Model Risk

M
o
d
e
l

R
i
s
k

E
d
i
t
e
d

B
y

D
a
n
i
e
l

R
s
c
h

a
n
d

H
a
r
a
l
d

S
c
h
e
u
l
e

EDITED BY DANIEL RSCH
AND HARALD SCHEULE
Model
PEFC Certified
This book has been
produced entirely from
sustainable papers that
are accredited as PEFC
compliant.
www.pefc.org
The rst years of the 21st Century saw the
nancial industry continue to pour vast
resources into building models to measure
nancial risks. Yet, as the nancial crisis that
began in 2007 has shown, these models
predictions were often being used without
recognition of the assumptions behind the
models themselves; acknowledgement of their
limitations; and understanding of the context in
which they were developed. The consequences
of model risk have been clear to see in the
nancial turmoil.
Drawing on experiences and data from
the nancial crisis, Model Risk: Identication,
Measurement and Management provides detailed
analysis of the shortcomings in the design
and application of modern risk models and
offers solutions for better understanding and
use in the post-crisis era. The book sets out
how to include model risk into existing risk
measurement frameworks solutions and how to
build better models as a result.
Part one of the book begins by setting out
frameworks for model risk. Four subsequent
sections tackle the models nancial institutions
use by risk type:
Macroeconomic and Capital Models
Credit Portfolio Risk Models
Liquidity, Market and Operational Risk
Models
Risk Transfer and Securitisation Models
Chapters address:
Sensitivity of regulatory and
economic capital to market stress
Systematic risk in a CDO
portfolio
Transmission of macro shocks
Improving estimations of
probably of default
Cashows from derivative
portfolios
Adequacy of market risk models
A new concept of potential
market risk
To date, model risk has lacked a clear
denition. This book explains the
different types of model risk; and
illustrates these with experiences
from the current nancial crisis.
Model Risk stands out as the guide to
better risk management in uncertain
times.
Risk
Identication,
Measurement
and Management
Model Risk
Model Risk
Identication, Measurement and Management
Edited by Harald Scheule and Daniel Rsch
Published by Risk Books, a Division of Incisive Financial Publishing Ltd
Haymarket House
2829 Haymarket
London SW1Y 4RX
Tel: + 44 (0)20 7484 9700
Fax: + 44 (0)20 7484 9797
E-mail: books@incisivemedia.com
Sites: www.riskbooks.com
www.incisivemedia.com
2010 Incisive Media
ISBN 978-1-906348-25-0
British Library Cataloguing in Publication Data
Acatalogue record for this book is available from the British Library
Publisher: Nick Carver
Commissioning Editor: Lucie Carter
Managing Editor: Jennifer Gibb
Designer: Lisa Ling
Copy-edited and typeset by T
&
T Productions Ltd, London
Printed and bound in the UK by PrintonDemand-Worldwide
Conditions of sale
All rights reserved. No part of this publicationmay be reproduced inany material formwhether
by photocopying or storing in any medium by electronic means whether or not transiently
or incidentally to some other use for this publication without the prior written consent of
the copyright owner except in accordance with the provisions of the Copyright, Designs and
Patents Act 1988 or under the terms of a licence issued by the Copyright Licensing Agency
Limited of 90, Tottenham Court Road, London W1P 0LP.
Warning: the doing of any unauthorised act in relation to this work may result in both civil
and criminal liability.
Every effort has been made to ensure the accuracy of the text at the time of publication, this
includes efforts to contact each author to ensure the accuracy of their details at publication
is correct. However, no responsibility for loss occasioned to any person acting or refraining
from acting as a result of the material contained in this publication will be accepted by the
copyright owner, the editor, the authors or Incisive Media.
Many of the product names contained in this publication are registered trade marks, and Risk
Books has made every effort to print them with the capitalisation and punctuation used by the
trademark owner. For reasons of textual clarity, it is not our house style to use symbols such
as TM, , etc. However, the absence of such symbols should not be taken to indicate absence
of trademark protection; anyone wishing to use product names in the public domain should
rst clear such use with the product owner.
While best efforts have been intended for the preparation of this book, neither the publisher,
the editor nor any of the potentially implicitly afliated organisations accept responsibility
for any errors, mistakes and or omissions it may provide or for any losses howsoever arising
from or in reliance upon its information, meanings and interpretations by any parties.
Contents
List of Figures ix
List of Tables xv
About the Editors xxi
About the Authors xxiii
Introduction xxxv
PART I CONCEPTS AND STOCHASTIC FRAMEWORKS FOR
MODEL RISK 1
1 Downturn Model Risk: Another View on the Global Financial
Crisis 3
Daniel Rsch; Harald Scheule
Leibniz Universitt Hannover; The University of Melbourne
2 Follow the Money from Boom to Bust 19
Jorge R. Sobehart
Citi Risk Architecture
3 Model Risk and Non-Gaussian Latent Risk Factors 45
Stef Hse and Stefan Huschens
Technische Universitt Dresden
4 Model Risk in Garch-Type Financial Time Series 75
Corinna Luedtke, Philipp Sibbertsen
Leibniz Universitt Hannover
PART II MACROECONOMIC AND CAPITAL MODELS 91
5 Monetary Policy, Asset Return Dynamics and the General
Equilibrium Effect 93
Kuang-Liang Chang; Nan-Kuang Chen; Charles Ka Yui Leung
National Chiayi University; National Taiwan University;
City University of Hong Kong
6 Capital Divergence: Sensitivity of Economic and Regulatory
Capital under Stress 137
Oleg Burd
KfW IPEX-Bank GmbH
v
MODEL RISK
PART III CREDIT PORTFOLIO RISK MODELS 153
7 Diversied Asset Portfolio Modelling: Sources and
Mitigants of Model Risk 155
Sean Keenan, Stefano Santilli, Sukyul Suh;
Andrew Barnes, Huaiyu Ma, Colin McCulloch
GE Capital; GE Global Research Center
8 Transmission of Macro Shocks to Loan Losses in a Deep Crisis:
The Case of Finland 183
Esa Jokivuolle; Matti Virn; Oskari Vhmaa
Bank of Finland; University of Turku and Bank of Finland;
University of Turku
9 Comparison of Credit-Risk Models for Portfolios of Retail
Loans Based on Behavioural Scores 209
Lyn C. Thomas; Madhur Malik
University of Southampton; Lloyds Banking Group
10 Validating Structural Credit Portfolio Models 233
Michael Kalkbrener, Akwum Onwunta
Deutsche Bank AG
11 Asymmetric Asset Correlation: Some Implications for the
Estimation of Probability of Default 263
Peter Miu; Bogie Ozdemir
McMaster University; BMO Financial Group
12 A Latent Variable Approach to Validate Credit Rating Systems 277
Kurt Hornik, Rainer Jankowitsch, Christoph Leitner, Stefan Pichler;
Manuel Lingo, Gerhard Winkler
Wirtschaftsuniversitt Wien; Oesterreichische Nationalbank
PART IV LIQUIDITY, MARKET AND OPERATIONAL RISK
MODELS 297
13 Modelling Derivatives Cashows in Liquidity Risk Models 299
Stefan Reitz
Hochschule fr Technik Stuttgart
14 Potential Future Market Risk 315
Manuela Spangler, Ralf Werner
Deutsche Pfandbriefbank
15 Market Risk Modelling: Approaches to Assessing
Model Adequacy 339
Carsten S. Wehn
DekaBank
vi
CONTENTS
16 Estimation of Operational Value-at-Risk in the Presence of
Minimum Collection Threshold: An Empirical Study 359
Anna Chernobai; Christian Menn; Svetlozar T. Rachev; Stefan Trck
Syracuse University; DZ Bank AG; Universitt Karlsruhe,
Finanalytica Inc, University of California at Santa Barbara;
Macquarie University
17 Operational Risk and Hedge Fund Performance: Evidence from
Australia 421
Robin Luo, Xiangkang Yin
La Trobe University
PART V RISKTRANSFER AND SECURITISATION MODELS 435
18 Identication and Classication of Model Risks in
Counterparty Credit Risk Measurement Systems 437
Marcus R. W. Martin
University of Applied Sciences, Darmstadt
19 Quantifying Systematic Risks in a Portfolio of Collateralised
Debt Obligations 457
Martin Donhauser, Alfred Hamerle, Kilian Plank
University of Regensburg
Epilogue 489
Index 493
vii
List of Figures
1.1 Seasonally adjusted delinquency rates for all commercial US
banks 4
1.2 Through-the-cycle (TTC) model and a point-in-time (PIT)
credit risk model 5
1.3 Credit-portfolio loss distributions 8
1.4 Volume of credit derivative transactions 9
1.5 Buyer of credit risk protection 10
1.6 Seller of credit risk protection 10
1.7 Average maturity of new credit derivatives 11
1.8 Interest and principal impairments of securitisations 12
1.9 Credit-portfolio loss distributions 12
1.10 Spread sensitivity of a senior tranche 13
1.11 Impairment rates by rating category 14
2.1 Distribution of normalised returns for the S&P 500 Index 32
2.2 Distribution of normalised returns for the DJI Index 32
2.3 Evolution of the marginal distribution of excess returns
p(, t) for = 1 and / = 0. 3 34
2.4 Evolution of the marginal distribution of log prices q(, t) for
= 1 and / = 0. 3 34
2.5 Comparison between the frequency of normalised returns
for the S&P 500 Index and Equation 2.20 for different time
horizons for the period 19502009 35
2.6 Comparison between the frequency of normalised returns for
the S&P 500 Index and Equation 2.20 as a function of the
reduced variable w for different time horizons using the
same data as in Figure 2.5 36
the DJI Index and Equation 2.20 for different time horizons in
the period 19282009 37
the DJI Index and Equation 2.20 as a function of the reduced
variable w for different time horizons using the same data as
in Figure 2.7 38
2.9 Comparison between the frequency of normalised returns
for the FTSE Index and Equation 2.20 as a function of the
reduced variable w for different time horizons in the period
19842009 38
the Nikkei Index and Equation 2.20 as a function of the
ix
MODEL RISK
reduced variable w for different time horizons in the period
19842009 39
4.1 Quantilequantile plot of S&P 500 77
4.2 Returns of S&P 500 78
4.3 Autocorrelation function for squared returns of S&P 500 78
5.1 (a) Federal funds rate (FFR), (b) interest rate spread (SPR),
(c) housing market returns (HRET) and (d) equity REIT
returns (REIT) 101
5.2 Smoothed probabilities for the SVAR(1) model of (FFR, SPR,
GDP, REIT) 106
GDP, HRET) 106
GDP, SRET) 107
5.5 Impulse responses of REIT to innovations in FFR when the
effect of SPR or GDP is shut off (FFR, SPR, GDP, REIT) 108
5.6 Impulse responses of HRET to innovations in FFR when the
effect of SPR or GDP is shut off (FFR, SPR, GDP, HRET) 109
5.7 Impulse responses of SRET to innovations in FFR when the
effect of SPR or GDP is shut off (FFR, SPR, GDP, SRET) 110
5.8 Simulation-based out-of-sample forecasts of stock returns
with 80% CI from 2006 Q1 to 2006 Q4 based on information
available at 2005 Q4 120
5.11 Simulation-based out-of-sample forecasts of housing returns
6.1 IRBAmaturity adjustment as function of pd and maturity 141
6.2 Maturity distribution of portfolio 145
6.3 Regional distribution of portfolio 146
7.1 Risk streams requiring aggregation 157
7.2 Accuracy of the mark-to-par algorithm 166
x
LIST OF FIGURES
7.3 Comparison of the CCM and IPM loss distributions 169
7.4 Conditional versus unconditional loss distribution 172
7.5 Capital rates: portfolio model versus meta model 175
7.6 Model implied capital rates versus PD by maturity band 175
7.7 US public company default rates 177
7.8 US public company default rates by sector 178
8.1 Industry-specic default rates 184
8.2 Relationship between loan losses and the aggregate default
rate 184
8.3 Comparison of OLS and SUR estimates of output gap for
different sectors 193
8.4 The estimated average output gap annual LGD against the
actual LGD 196
8.5 Distribution of loan losses (xed LGD) 199
8.6 Expected losses and the length of depression: feedback from
defaults to output 200
8.7 Comparison of effects of macro shocks 201
8.8 Fit of the constant LGD and the endogenous LGD loan-loss
models 203
9.1 Monte Carlo simulation run to calculate appropriate K value 217
9.2 ROC curve for model Aand model B of proportional hazards
example 222
12.1 Rating bias for bank/industry combinations
g,j
of the
13 Austrian banks 289
12.2 Standard deviations
g,j
of the rating errors for
bank/industry combinations of the 13 Austrian banks 289
12.3 Residual analysis for all 13 banks across the legal forms:
limited and unlimited companies 292
12.4 Residual analysis for two banks (bank 13 and bank 8) across
the relative exposure 292
13.1 Exercise probabilities for a one-year call option 306
13.2 Simulated paths of L(t, 5, 6) 307
13.3 Probabilities for a positive cashow at expiry 310
13.4 Simulated paths of ln(S
/K) and x() 312

14.1 Time series of risk factors from February 11, 2004, to July 17,
2009 319
14.2 Historical values of (a) volatilities and (b) correlations 320
14.3 Joint inuence of interest rate and credit-spread level on
bond value at issuance 322
14.4 Bond and portfolio values over time 322
14.5 Portfolio sensitivity over time for N
i
equal to 1 million 324
xi
MODEL RISK
14.6 Decomposition of total risk in interest rate risk and
credit-spread risk (at xing dates) 324
14.7 Impact of ageing, interest rate and credit-spread levels on
portfolio credit-spread sensitivity 325
14.8 Impact of sensitivity and covariance parameters on portfolio
VaR 326
14.9 Interest rate and credit-spread paths obtained by historical
bootstrapping 332
14.10 Quantiles of simulated sensitivities over one year 332
14.11 Quantiles of simulated volatilities over one year 333
14.12 Potential future credit-spread VaR and actual credit-spread
VaR evolution under the bootstrapping model 333
14.13 Potential future credit-spread VaR and actual credit-spread
VaR evolution under the bootstrapping model with stressed
scenarios 335
15.1 Embedding the results of backtesting in a regular validation
and backtesting process 353
16.1 Ratios of estimated parameters to the true (complete-data)
parameter values, for the lognormal example, u = 50 368
16.2 Ratios of estimated fraction of missing data (Q) to the true
(complete-data) fraction, for the lognormal example, u = 50 369
16.3 Ratios of estimated one-year EL, 95% VaR and 95% CVaR to
the true (complete-data) values, for the lognormal example,
u = 50, = 100 370
16.4 Annual accumulated number of external operational
losses, with tted cubic and Poisson models 372
16.5 Upper quantiles of tted truncated loss distributions to the
external-type losses, together with the empirical distribution 375
16.6 Fitted frequency functions to the operational losses 392
16.7 Upper quantiles of tted truncated loss distributions to
operational losses, together with the empirical distribution 397
17.1 Distribution of operational issues contributing to operational
risk in hedge funds 422
18.1 Sample paths for EuroStoxx50 generated by a GBM ESG 438
18.2 Sample paths for a call option on EuroStoxx50 439
18.3 Counterparty exposure prole for a single uncollateralised
call option on EuroStoxx50 440
18.4 Portfolio of a European put and call on the EuroStoxx50 440
18.5 Three steps for generating exposure proles and
counterparty measures 442
18.6 Stressing the at implied volatility assumption by a
deterministic time-dependence for illustrative purposes 447
xii
LIST OF FIGURES
19.1 Hitting-probability proles of the BBB mezzanine tranche
and a BBB bond 467
19.2 EL proles of the BBB mezzanine tranche and a BBB bond 468
19.3 Goodness of t of approximated conditional expected loss 474
xiii
List of Tables
4.1 S&P 500: descriptive statistics 76
4.2 Declaration of models used 83
4.3 Parameters of the data-generating processes 83
4.4 DGP Garch(1,1)-N: mean squared error and p-values of the
DieboldMariano test 84
4.5 DGP EGarch(1,1)-N: mean squared error and p-values of the
4.6 DGP GJR-Garch(1,1)-N: mean squared error and p-values of
the DieboldMariano test 85
4.7 DGP APArch(1,1)-N: mean squared error and p-values of the
4.8 DGP FIGarch(1,1)-N: mean squared error and p-values of the
4.9 DGP HYGarch(1,1)-N: mean squared error and p-values of
the DieboldMariano test 86
5.1 Statistical summary of federal funds rate, interest rate
spread, housing market returns, and equity REIT returns
(1975 Q22008 Q1) 100
5.2 Correlation coefcients (1975 Q22008 Q1) 102
5.3 AIC values for various three-variable VAR(p) models of the
REIT system 107
5.4 AIC values for various three-variable VAR(p) models of the
HRET system 107
5.5 Statistical summary of federal funds rate, term spread, gross
domestic production growth rate, external nance premium,
market liquidity, stock index return and housing market
return (1975 Q22008 Q3) 113
5.6 Correlation coefcients (1975 Q22008 Q3) 114
5.7 List of models 114
5.8 Asummary of goodness of t for all eight models 117
5.9 A summary of in-sample forecasting performance
(four-quarter-ahead forecasts) 118
5.10 A summary of out-of-sample forecasting performance
(four-quarter-ahead forecasts) 119
5.11 Is the forecasted stock return within the 80% condence
interval? 126
5.12 Is the forecasted housing return within the 80% condence
interval? 127
xv
MODEL RISK
5.13 Do models forecast stock return better in the presence of
housing return? 127
5.14 Do models forecast housing return better in the presence of
stock return? 128
5.15 A summary of in-sample forecasting performances
(four-quarter ahead forecasts) 129
5.16 A summary of out-of-sample forecasting performances
(four-quarter ahead forecasts) 130
6.1 Asset correlation in IRBAand multi-factor models 142
6.2 Rating distribution of portfolio 143
6.3 Single name concentration in portfolio 144
6.4 Regulatory and economic capital requirements as percentage
of total exposure 144
6.5 Regulatory and economic capital requirements in various
stress scenarios 147
6.6 Increase in regulatory capital requirements with constant
maturity and constant correlation 147
6.7 Sensitivity (%) of regulatory capital requirements with xed
maturity and stressed correlation of 28.4% 148
8.1 Estimation results of the basic default-rate model for the
various industries 188
8.2 Diagnostic tests 191
8.3 Summary of simulations 198
9.1 KolmogorovSmirnov (KS) results for alternative models 217
9.2 Coefcients in the case study of the proportional hazard model 220
9.3 Numbers of predicted and actual defaults in out of time
sample using proportional hazard models 223
9.4 First-order stationary transition matrix 226
9.5 Parameters for second-order Markov chain with age and
economic variables 227
9.6 Results of default predictions for the two transition matrix
models 230
10.1 Asset correlations of rating cohorts 246
10.2 intra-correlations (%) of industry cohorts 247
10.3 Variation (%) of intra-correlations over time 251
10.4 intra-correlations (%) using rating data 255
10.5 Variation of intra-correlation (%) and log-likelihood function
with degrees of freedom 256
11.1 True parameters of the asymmetric correlation model 270
11.2 Performance of LRPD
ave
and LRPD
cc
in estimating
unconditional probability of default 272
xvi
LIST OFTABLES
11.3 Summary statistics of the asset return correlation estimator 273
12.1 Descriptive statistics of the characteristics of the rating
information and the 13 Austrian banks in the data set 284
12.2 Distribution of the co-ratings of the 13 Austrian banks across
industries 285
12.3 Industry-specic means
g
and PD intervals measured in
basis points (10
4
) 286
12.4 Rating bias
g,j
for bank/industry combinations of the
13 Austrian banks 288
12.5 Standard deviations
g,j
bank/industry combinations of the 13 Austrian banks 290
16.1 Fraction of missing data, F
0
(u), for the lognormal(
0
,
0
)
example with nominal threshold of u = 50 369
16.2 Fitted frequency functions to the external"-type losses 372
16.3 Estimated and F
(u) values for the external-type

operational loss data 374
16.4 Results of in-sample GOF tests for external-type
operational losses 376
16.5 Estimates of expected aggregated loss, VaR and CVaR for
external-type losses 377
16.6 Average estimates of forecast errors for external-type
aggregated losses 380
16.7 LR statistic and p-values for external-type aggregated
losses in the seven-year forecast period 384

operational loss data, under the robust approach 386
external-type losses, under the robust approach 387
16.10 Average estimates of forecast errors for external-type
aggregated losses, under the robust approach 388
16.11 LR statistic and p-values for external-type aggregated
losses in the seven-year forecast period, under the robust
approach 390
16.12 Frequency functions tted to the operational losses 392
(u) values for the relationship,

human, processes and technology-type operational
loss data 393
16.14 Results of in-sample GOF tests for relationship-type
16.15 Results of in-sample GOF tests for human-type operational
losses 399
16.16 Results of in-sample GOF tests for process-type operational
losses 400
xvii
MODEL RISK
16.17 Results of in-sample GOF tests for technology-type
relationship-type losses 402
16.19 Estimates of expected aggregated loss, VaR, and CVaR for
human-type losses 403
process-type losses 404
technology-type losses 405
16.22 Average estimates of forecast errors for relationship-type
16.23 Average estimates of forecast errors for human-type
16.24 Average estimates of forecast errors for process-type
16.25 Average estimates of forecast errors for technology-type
16.26 LR statistic and p-values for relationship-type aggregated
16.27 LR statistic and p-values for human-type aggregated losses
in the seven-year forecast period 415
16.28 LR statistic and p-values for process-type aggregated
16.29 LR statistic and p-values for technology-type aggregated
17.1 Australian hedge funds: legal structure 425
17.2 Descriptive statistics of Australian hedge funds 430
17.3 Empirical results 431
19.1 Asset pool conguration 462
19.2 Structure of liabilities 462
19.3 Results: CDO risk measures 465
19.4 Approximation results for the bond representation 473
19.5 Risk measures for different portfolio sizes 476
19.6 ABS CDO collateral pool composition 479
19.7 Outer CDO structure based on expected tranche loss 479
19.8 Risk measures for the ABS CDO 480
19.9 Risk measures for thin mezzanine tranches 482
19.10 Risk measures for thin senior tranche and super senior tranche 483
19.11 Risk measures of investment alternatives 486
xviii
About the Editors
Harald Scheule teaches nance and banking in the Department of
Finance at the University of Melbourne. He has worked globally
as a consultant on credit risk, structured nance and securitisation
projects for banks, insurance and other nancial service companies.
He maintains strong research relationships with theAustralian, Ger-
man and Hong Kong regulators for nancial institutions. He has
published extensively and organised executive training courses in
his discipline.
Daniel Rsch is professor of nance and head of the Institute
of Banking and Finance at the Leibniz Universitt Hannover. He
received his PhD from the University of Regensburg. Daniels work
covers a broad range of subjects within asset pricing and empirical
nance. He has published numerous articles on risk management,
credit risk, banking andquantitative nance in leading international
journals. Daniel has also led numerous executive training courses
and is a consultant to nancial institutions on credit risk issues.
xix
About the Authors
Andrew Barnes is a researcher at the Risk and Value Manage-
ment Technologies Laboratory of the GE Global Research Center
in Niskayuna, New York. Since joining General Electric in 2004, he
has worked on quantitative nance problems with a focus on risk
measurement and analysis for large commercial loan and asset port-
folios. Before joining General Electric, Andrew spent several years
working on partial differential equations and electromagnetic scat-
tering problems at Duke University. He holds a BS in mathemat-
ics from Yale University, and a PhD in mathematics from Duke
University.
Joseph L. Breeden is president and chief operating ofcer of Strate-
gic Analytics Inc. Joseph has spent the past 12 years designing
and deploying risk management systems for retail loan portfolios.
At Strategic Analytics, which he co-founded in 1999, he leads the
design of advanced analytics and takes a leading role working with
client institutions. He has personally experienced and created mod-
els through the 1995 Mexican peso crisis, the 1997 Asian economic
crisis, the 2001 global recession, the 2003 HongKongSARSrecession,
and the 2007 US mortgage debacle. These crises have provided him
with a unique perspective on crisis management and the analytics
needs of executives for strategic decision-making. Joseph received
separate BS degrees in mathematics and physics in 1987 from Indi-
ana University. He earned a PhD in physics in 1991 from the Uni-
versity of Illinois. His thesis work involved real-world applications
of chaos theory and genetic algorithms. In the mid 1990s, he was a
member of the Santa Fe Institute. Since 1987, he has published more
than 40 articles in various journals on subjects including portfolio
forecasting, economic capital, evolutionary computation, non-linear
modelling, astrophysics and nuclear physics.
Oleg Burd is a vice president in the risk management department of
KfW IPEX-Bank GmbH and specialises in measurement and man-
agement of credit risk concentrations. His current responsibilities
include credit-portfolio modelling as well as supervision and imple-
mentation of active management of banks credit portfolio. Prior
xxi
MODEL RISK
to joining KfW IPEX-Bank in 2004, Oleg worked at the German
branch of Maple Financial Group, Maple Bank, where he devel-
oped, reviewed and implemented quantitative models for statistical
arbitrage trading. Oleg holds an MSc in economics and an MSc in
mathematics, both from the University of Gttingen.
Kuang-Liang Chang received his MAand PhD at the National Tai-
wan University in 1999 and 2004, respectively. He is an assistant
professor at the Department of Applied Economics, National Chi-
ayi University. Kuang-Liang has published in Applied Economics, The
Manchester School and Economic Modelling, among other journals.
Nan-Kuang Chen received his BA and MA at the National Taiwan
University in 1987 and 1989, respectively, and his PhD at the Uni-
versity of Minnesota in 1997. He is a professor in the Department of
Economics, National Taiwan University. He was a visiting scholar at
the London School of Economics in 2003 and has published articles
in numerous journals on economics and real estate.
Anna S. Chernobai is an assistant professor of nance at the M. J.
Whitman School of Management at Syracuse University, NewYork.
The focus of her research is operational risk, default risk, stochastic
processes, and applied statistics. She is an author of the book Opera-
tional Risk: AGuide to Basel II Capital Requirements, Models, andAnalysis
and is an FDICresearch fellowand JPMorgan Chase research fellow.
Anna earned her PhD in statistics and applied probability from the
University of California at Santa Barbara in 2006. She also holds a
Masters degree in nance from the Warwick Business School at the
University of Warwick, UK, and a Bachelors degree in economics
from Sophia University, Japan.
Martin Donhauser is a research assistant at the chair of statistics at
the University of Regensburg. He studiedeconomics andpreviously
workedas aconsultant at RiskResearchProf. HamerleGmbH, where
he was mainly involved with the development and implementation
of credit risk management techniques and solutions for medium-
sized German banks and international nancial institutions. Martin
is nishing his doctoral dissertation. His research focuses on the
valuation and risk analysis of structured nance products and the
dynamic modelling of credit risk.
Alfred Hamerle is a professor of statistics at the faculty of business,
economics andinformationsystems at the Universityof Regensburg.
xxii
ABOUTTHE AUTHORS
Prior to serving in his present position, he was professor of statistics
at the University of Konstanz and professor of statistics and econo-
metrics at the University of Tbingen. He is the founder and CEOof
Risk Research Prof. Hamerle GmbH. His primary areas of research
include statistical and econometric methods in nance, credit risk
modelling and Basel II as well as multivariate statistics. Alfred has
publishedeight books andmore than80 articles inscientic journals.
Kurt Hornik is the head of the Research Institute for Computational
Methods and the chair of the Department of Statistics and Mathe-
matics at the Vienna University of Economics andBusiness. He com-
pletedhis doctoral researchandhabilitationat the Vienna University
of Technology. His research interests include statistical computing,
statistical graphics, statistical and machine learning, data mining
and a variety of application domains for quantitative data analysis,
in particular quantitative risk management. Kurt has co-authored
around 200 publications in refereed journals and conference pro-
ceedings, is among the ISI 100 most highly cited researchers in the
engineering category and holds the Gold Merit Decoration of the
Republic of Austria for his scientic achievements.
Stef Hse is a postdoctoral fellow at the Technische Universitt
Dresden, Faculty of Business and Economics, and chair of quantita-
tive methods, especially statistics, where she works in quantitative
risk analysis. Her current research focuses on credit risk manage-
ment, in particular on the modelling of dependence structures by
means of risk factor and mixture models, on the simultaneous esti-
mation of dependence and default parameters and on the involved
model risk. Stef has been a trainer in the SRP/IRB qualication
programme for supervisors of the Deutsche Bundesbank and the
Federal Financial Supervisory Authority (Bundesanstalt fr Finanz-
dienstleistungsaufsicht) since 2004. She holds an academic degree
in business management and a doctoral degree from the Technische
Universitt Dresden.
Stefan Huschens holds the chair of quantitative methods, spe-
cialising in statistics at the Technische Universitt Dresden. He
holds a doctoral degree in economics and a habilitation degree in
statistics and economics from the Ruprecht-Karls-Universitt Hei-
delberg. Stefan has been a trainer in the SRP/IRB qualication
programme for supervisors of the Deutsche Bundesbank and the
xxiii
MODEL RISK
Federal Financial Supervisory Authority (Bundesanstalt fr Finanz-
dienstleistungsaufsicht) since 2004. His major research interests
are statistical and econometric methods of market and credit risk
management.
Rainer Jankowitsch is assistant professor of nance at the Vienna
University of Economics and Business. He completed his doctoral
research at the University of Vienna and recently nished his habil-
itation. His research is focused on credit and liquidity risk, banking,
risk management and nancial markets. In the past ve years he has
published in various nance journals such as the Journal of Banking
and Finance and The Journal of Risk. Rainer received the Best Paper
Award fromthe German Finance Association in 2008 for his work on
liquidity risk, which was produced in cooperation with New York
University. His current research is focused on the 20089 nancial
crisis.
Esa Jokivuolle is a research supervisor in the Bank of Finlands
Research Unit, specialising in nancial markets research. He is also
an adjunct professor of nance in the Helsinki School of Economics.
Previouslyhe workedinthe Bankof Finlands Financial Markets and
Statistics Department, and as a senior quantitative analyst in Leonia
plc in Helsinki. He has published several academic research articles.
Esa earned a PhD in nance in 1996 from University of Illinois.
Michael Kalkbrener is head of the portfolio-modelling teamwithin
the risk analytics and instruments department of Deutsche Bank
and he specialises in developing risk measurement and capital allo-
cation methodologies. His responsibilities include credit-portfolio
modelling and the development of a quantitative model for oper-
ational risk. Prior to joining Deutsche Bank in 1997, he worked at
Cornell University and the Swiss Federal Institute of Technology,
where he received the Venia Legendi for mathematics. Michael holds
a PhDin mathematics fromthe Johannes Kepler University Linz. He
has published numerous research articles on mathematical nance
and scientic computation.
Sean Keenan is the portfolio analytics leader at GE Capital, respon-
sible for credit risk systems and quantitative risk modelling. Prior to
joining GE he held quantitative research positions at Citigroup and
Moodys Investors Service. He holds a PhD in economics and a BA
xxiv
ABOUTTHE AUTHORS
in history, both from New York University. He has written a variety
of articles on quantitative credit risk topics and is regular speaker at
conferences.
Christoph Leitner is a research assistant at the Department of Statis-
tics and Mathematics, Vienna University of Economics and Busi-
ness. His research interests focus on the analysis of ratings, in both
nance and sports. He has recently contributed to several confer-
ences and workshops on credit risk, including the Annual Meeting
of the Southern Finance Association 2009. In the matter of sports
ratings, he has contributed articles to the International Journal of
Forecasting and to the proceedings of the 2nd International Confer-
ence on Mathematics in Sport (IMASport 2009 in Groningen, The
Netherlands).
Manuel Lingois ananalyst at Oesterreichische Nationalbank, where
he is responsible for the development of operations of the Inhouse
Credit Assessment System (ICAS) used for Eurosystem monetary
operations. Before joiningOesterreichische Nationalbankhe worked
as researchassistant at the Vienna Universityof Economics andBusi-
ness and as a freelance consultant for PricewaterhouseCoopers. He
publishes in journals related to credit risk (The Journal of Credit Risk
and The Journal of Risk Model Validation). His current research focuses
on rating system development and validation. Manuel holds a PhD
in nance from the Vienna University of Economics and Business.
Charles Ka Yui Leung received his BSc at the Chinese University of
Hong Kong in 1991 and his PhD at the University of Rochester in
1996. He taught at the Department of Economics, Chinese University
of Hong Kong and is an associate professor at the the Department of
Economics and Finance, City University of Hong Kong. He received
the Fulbright Scholarship (Research) in 20045 and has been a vis-
iting scholar at both the Fisher Center for Real Estate and Urban
Economics at the Haas School of Business, University of California,
Berkeley and the Hoover Institution, Stanford University. He has
published in the Journal of Monetary Economics, Journal of Urban Eco-
nomics, Journal of Regional Science, Journal of Real Estate Finance and
Economics, Journal of Real Estate Research and Journal of Housing Eco-
nomics, among other journals. He serves on the Editorial Board of
International Real Estate Review, the Board of Directors of the Asian
Real Estate Society (AsRES) and the Board of Directors of the Global
xxv
MODEL RISK
Chinese Real Estate Congress (GCREC). He also served as a guest
editor of the Journal of Housing Economics in 2007.
Corinna Luedtke is a PhD student at the Institute of Statistics at
the Leibniz Universitt Hannover. Her main research interests are
time series analysis and quantitative risk management. Corinna
graduated in economics and business administration at the Leibniz
Universitt Hannover in 2008.
Robin Luo is senior lecturer of nance at La Trobe University, Aus-
tralia. Prior to joining La Trobe University, he taught and researched
at Auckland University of Technology in New Zealand, Nanyang
Technological University in Singapore, anda couple of other tertiary
institutions inAsia. Dr Luo is a Financial Risk Manager (FRM), a Fel-
lowmember of the Global Associationof Risk Professionals (GARP),
co-director of GARP Regional Chapter in Melbourne and director of
GARP College Chapter at La Trobe University. His current research
interests focus on nancial risk management, asset pricing, market
efciency, international nance and Asia-Pacic nancial markets.
He has published in Economic Modelling, Applied Financial Economics,
the Global Economy Journal and Applied Economics Letters.
Huaiyu (Harry) Ma is a statistician in the Applied Statistics Lab
at the GE Global Research Center. He received his PhD in decision
sciences andengineering systems fromRensselaer Polytechnic Insti-
tute. His research interests include data analysis, simulation, time-
series analysis, statistical computing and their applications in risk
management, engineering and online social networks problems.
Madhur Malik is a senior analyst with the Lloyds Banking Group,
where he specialises in developing advanced nancial models for
portfolio credit risk, Basel II and macroeconomic time-series data.
Prior to joining Lloyds Banking Group, he was a research fellow
at the University of Southampton, where he applied a number of
innovative approaches such as survival analysis and Markov chains
to estimate portfolio level credit risk of retail loans. Madhur holds a
Masters degree in applied mathematics fromthe Indian Institute of
Technology in Roorkee and a PhD in mathematics from the Indian
Statistical Institute.
Marcus R. W. Martin is professor of nancial mathematics and
stochastics at the University of Applied Sciences in Darmstadt (Ger-
many). From 2002 to 2008, he was with Deutsche Bundesbank,
xxvi
ABOUTTHE AUTHORS
where he headedthe Quantitative RiskModel ExaminationGroupat
Hauptverwaltung Frankfurt of Deutsche Bundesbank fromNovem-
ber 2004. In this position he was responsible for conducting regula-
tory audits of IRBA, IMM, IAA, internal market risk and liquidity
risk models of Germanbanks. His current researchinterests focus on
modelling counterparty risk as well as asset liability, liquidity risk
and commodity risk modelling.
Colin C. McCulloch is a statistician in the Applied Statistics Lab-
oratory at the GE Global Research Center. He has worked in the
area of nancial risk modelling for seven years. In that time he
has developed models of capital adequacy and capital allocation for
GE Capitals credit and market risk exposures. Colin holds a PhD
in statistics from Duke University and has published 14 articles in
peer-reviewed journals.
Christian Menn works as senior equity derivatives trader at DZ
Banks structured product division. Before joining DZ Bank, he held
the position of equity derivatives trader at Sal. Oppenheim. After
gaining his PhD in economics at the University of Karlsruhe, Chris-
tian worked as visiting assistant professor at the School of Opera-
tions Research at Cornell University. He holds a degree in mathe-
matics from the University of Karlsruhe and the Universit Joseph
Fourier in Grenoble.
Peter Miu is an associate professor of nance at DeGroote School of
Business, McMaster University. He teaches nancial institutions as
well as international nancial management at both the undergrad-
uate and MBAlevels. His research has been conducted primarily in
such areas as credit risk modelling and forecasting, pricing and risk
management of credit portfolios, and Basel II implementation and
validation. He has consultedona number of Basel II implementation
projects and is a frequent speaker at both academic and professional
conferences on credit risk and Basel II. Peter obtained his PhD and
MBAin nance from the University of Toronto.
AkwumOnwunta is the Marie Curie Early Stage Research Fellowin
the COMISEF (Computational Optimization Methods in Statistics,
Econometrics and Finance) project at Deutsche Bank, Frankfurt,
Germany. He holds a BSc in mathematics, an MSc in physical and
mathematical analysis and a Diplme Universitaire in mathematical
xxvii
MODEL RISK
models in economics and nance. His research is focused on credit
risk modelling.
Bogie Ozdemir is a vice president of the BMO Financial Group
responsible for economic capital, stress testing, Basel analytics and
jointly responsible for ICAAP. Previously he was a vice president in
Standard & Poors Credit Risk Services group. In this role, he was
responsible for globally engineering new products and solutions,
business development andmanagement. He has co-authoredpapers
in The Journal of Credit Risk and published in the The Journal of Risk
Model Validation. His joint paper Discount Rate for Workout Recov-
eries: An Empirical Study with Brooks Brady, Peter Chang, Peter
MiuandDavidSchwartz wonthe Best Paper Awardat the FifthNTU
International Conference in 2007. Bogie has also co-authored a book
titled Basel II Implementation: A Guide to Developing and Validating a
Compliant, Internal Risk Rating System.
StefanPichler is a professor andthe chair of the Institute for Banking
and Finance at the Vienna University of Economics and Business.
He completed his doctoral studies at the University of Graz and
previously worked as an associate professor of nance at the Vienna
University of Technology. He has publishednumerous articles in the
Journal of Banking and Finance, Review of Finance, Quantitative Finance
and The Journal of Risk. His research focus is on risk management in
nancial and public institutions.
Kilian Plank is a research assistant and lecturer at the University
of Regensburg. In his doctoral dissertation he was concerned with
statistical modelling of growth processes in marketing. His research
focuses on statistical modelling and analysis of structured credit
products. Kilian has several years of work experience in the banking
industry and is engaged in consulting projects at Risk Research Prof.
Hamerle GmbH.
Svetlozar (Zari) Rachev holds the chair-professorship in statistics,
econometrics and mathematical nance at the University of Karl-
sruhe, and is the author of 12 books and over 300 published arti-
cles on nance, econometrics, statistics and actuarial science. At the
University of California at Santa Barbara, Zari foundedthe PhDpro-
gramme inmathematical andempirical nance. He holds PhD(1979)
and Doctor of Science (1986) degrees from Moscow University and
xxviii
ABOUTTHE AUTHORS
Russian Academy of Sciences. Zari was a co-founder and president
of BRAVO Risk Management Group, which has been acquired by
FinAnalytica, where he serves as chief scientist.
StefanReitzholds aPhDinmathematics andis professor of nancial
mathematics at the University of Applied Sciences in Stuttgart, Ger-
many. He also works as a consultant in the nancial industry in var-
ious projects (risk controlling, risk management, pricing of deriva-
tives). Prior tohis current positionhewas anauditor andaudit super-
visor within the banking examination department of the Deutsche
Bundesbanks regional ofce in Frankfurt. He conducted interna-
tional audits at major and regional banks in portfolio risk models,
pricingof derivatives, riskmanagement, minimumrequirements for
trading activities and Basel II implementation.
Stefano Santilli is vice president of portfolio analytics at GE Capi-
tal in Norwalk, CT, where he is responsible for portfolio modelling
in the Risk Management department. Prior to joining GE in 2003,
he worked as a credit risk controller with Dresdner Bank in Frank-
furt, Germany, and as an account manager with Ersel Sim in Milan,
Italy. Stefano holds an undergraduate degree in Business Adminis-
tration from Bocconi University, a Masters degree in nance from
the University of Alabama and is a CFAcharterholder.
PhilippSibbertsenis professor for statistics anddirector of the Insti-
tute of Statistics at the Leibniz Universitt Hannover. His research
interests are in nancial statistics and especially in statistical models
for measuring nancial risk and time series econometrics. Philipp
has numerous publications in these areas in highly ranked inter-
national journals and is a regular speaker at conferences on these
topics. He has also experience in applying statistical models to prac-
tical problems. Philipp holds a Diploma in mathematics from the
University of Hamburg and a PhD in statistics from the University
of Dortmund.
Jorge R. Sobehart is a managing director at Citigroup Risk Archi-
tecture. He is involved in credit risk capital measures and allocation,
stress testing, advanced portfolio loss models for wholesale expo-
sures, credit migration and default risk models. Previously, he was a
member of Moodys Standing Committee on Quantitative Tools and
VP senior analyst in Moodys Risk Management Services, where
xxix
MODEL RISK
he developed advanced default risk models, early warning tools
and model validation metrics and procedures. During his career, he
has worked and acted as a scientic consultant for several presti-
gious companies and institutions making contributions in differ-
ent elds. He has also acted as a referee for many professional
journals in nance, physics and mathematical modelling. Jorge has
advanced degrees in physics and has postdoctoral experience at the
Los Alamos National Laboratory.
Manuela Spangler works as a nancial engineer in the Risk Mod-
elling teamat Deutsche Pfandbriefbank. She studied nancial math-
ematics at the Technical University of Munich and at the National
University of Singapore. Her research interests include market risk
modelling and pricing of credit derivatives.
Sukyul Suh is vice president of portfolio modelling at GE Capital,
responsible for performing economic capital analyses and develop-
ing a risk modelling system for capital adequacy and capital allo-
cation. Prior to joining GE Capital in 2000, Sukyul was a process
engineer at SK energy, where he was responsible for improving
product quality and yield by applying statistical process control.
He holds an MBA degree from the University of Minnesota and a
BE degree in chemical engineering from Korea University. He is a
CFAcharterholder and a Certied Financial Risk Manager.
Lyn Thomas is professor of management science at the University
of Southampton. His interests are in applying operational research
andstatistical ideas inthe nancial area, particularlyincredit scoring
and risk modelling in consumer lending. He is a founder member
of the Credit Research Centre at the University of Edinburgh and
one of the principal investigators for the Quantitative Financial Risk
Management Centre based at Southampton. He has authored or co-
authored four books in the area, including Consumer Credit Models:
Pricing, Prot and Portfolios and Credit Scoring and its Applications. He
is a Fellowof the Royal Society of Edinburgh, a past president of the
Operational Research Society and was awarded the Beale Medal of
that Society in 2008.
Stefan Trck is an associate professor in the economics department
of Macquarie University, Sydney. He has held positions at Queens-
land University of Technology and at the University of Karlsruhe in
xxx
ABOUTTHE AUTHORS
Germany, where he received a PhD in statistics. His research inter-
ests focus on risk management and nancial econometrics includ-
ing the elds of credit risk, operational risk, power markets and
real estate nance. He has several years of consulting experience
for nancial institutions and has published in various international
journals including The Journal of Banking and Finance, the European
Journal of Finance, Energy Economics and The Journal of Credit Risk and
he is anauthor of the book Rating Based Modeling of Credit Risk: Theory
and Application of Migration Matrices.
Oskari Vhmaa is a PhD student in economics at the University
of Turku. He has previously worked as a research assistant at the
Research Unit of the Bank of Finland.
Matti Virn is a professor of economics at the University of Turku
anda scientic advisor to the Bank of Finland. Previously he worked
at the Bank of Finland as a research supervisor and in the Finnish
Government Institute for Economic Research as the research direc-
tor and as deputy director. He has published more than 100 arti-
cles in refereed journals and books. He studied at the University
of Chicago with a Fulbright scholarship, and gained his doctoral
degree (economics) from the University of Helsinki in 1980.
Carsten S. Wehn is head of market risk control at DekaBank, Frank-
furt. Market risk control is responsible for the measurement of mar-
ket and liquidity risk of the bank and the development of risk meth-
ods and models as well as the validation of the adequacy of the
respective risk models. Before joining DekaBank, he was responsible
for supervising and conducting regulatory examinations for inter-
nal market risk models with Deutsche Bundesbank. Carsten studied
in Siegen, Germany, as well as in Nantes, France. He holds a PhD
in mathematics and regularly gives lectures at universities. He has
published more than 30 articles and other publications including
three books.
Ralf Werner heads the global Risk Methods & Valuation Depart-
ment at Deutsche Pfandbriefbank and is mainly in charge of risk
methodology, nancial engineering and economic capital mod-
elling. Before joining Deutsche Pfandbriefbank, Ralf was responsi-
ble for market risk methodology at Allianz Group Risk Controlling.
In the past he has held positions as nancial engineer for credit
xxxi
MODEL RISK
risk topics and as consultant for investment strategies and asset
liability management at Risklab Germany, as well as prop trader
(Xetra, Eurex) for SchmidtBank Nrnberg. Ralf publishes regularly
in nance- and optimisation-related journals and speaks at inter-
national conferences. Since 2002, he has continuously supported
the HVB Institute for Mathematical Finance at TU Mnchen as lec-
turer for nancial optimisation andsimulation. Ralf holds a diploma
and a PhDin mathematics fromthe FriedrichAlexander Universitt
Erlangen-Nrnberg.
GerhardWinkler is deputyheadof Oesterreichische Nationalbanks
credit division. His research interests focus on credit risk measure-
ment, risk model validation and bank efciency. Before joining the
central bank he worked as an assistant professor at the Institute
of Banking and Finance at the Vienna University of Economics
and Business, where he recently completed his habilitation. He is
author of several academic publications in the eld of nancial risk
management and credit risk measurement.
Xiangkang Yin is a professor of economics and nance at La Trobe
University, Australia. His research interests cover a wide range of
topics in economics and nance, including capital asset pricing, cor-
porate nance and governance, industrial organisation and applied
microeconomic theory. Prior to jointing La Trobe University, he held
various positions at Shanghai Jiaotong University, Universite Louis
Pasteur and Monash University. Xiangkang has published articles
in top-tier economics and nance journals, including The Journal of
Finance, Journal of Development Economics, Journal of Economic Behavior
and Organization and the Australian Journal of Management.
xxxii
Introduction
The 1970s witnessed the start of a new era in nance. Starting with
the BlackScholesMerton option pricing formula, sophisticated
mathematical models for pricing risky securities found their way
into capital markets. Banks, insurance companies and hedge funds,
among others, soon migrated to using these models for pricing,
hedging, arbitrage or speculation.
At the same time the breakdown of the Bretton Woods system
rendered the nancial world riskier and the increased use of risk
measurement and management methods led to globalised, inter-
dependent markets andstronglyincreasingtradingvolumes inmore
and more complex nancial instruments. Consequently, the risk of
large failures due to mis-pricing and mismanagement increased and
many of these realised failures are still important objects in learn-
ing lessons about the malfunctioning of risk models. Among others,
these include the cases of Metallgesellschaft in 1993, the Bank of
Tokyo and Mitsubishi in 1997 and NatWest Capital Markets in 1997
and most recently the global nancial crisis.
After introducing a global regulation framework for strength-
ening the equity positions of nancial institutions (the so-called
Basel Accord or Basel I), banks were allowed to calculate capital
charges by using internal models for market risk, thereby honour-
ing the industrys advances in risk measurement approaches. Sim-
ilarly, Basel II acknowledges efforts made in recent years by basing
regulatory capital on bank-internal credit-rating models.
Financial risk models have become increasingly important for
nancial institutions, markets and instruments. These models are
individually crafted and then generally assembled to generate port-
folio measures. The occurrence of the 20089 global nancial crisis
suggests that many existing nancial risk models were unable to
predict the increase in loss rates prior to the crisis. This was partic-
ularly true for new markets such as asset securitisations and credit
derivatives. The consequence was a general loss in credibility, which
has resulted in changes of economic and regulatory requirements.
The global nancial crisis has resulted in changes for regulatory
requirements. The Basel Enhancement to the Basel II framework
xxxiii
MODEL RISK
(Basel Committee on Banking Supervision 2009) issued, in July, new
riskweights for ratedresecuritisations andis stresses the importance
of bank internal due diligence processes:
A bank should conduct analyses of the underlying risks when
investinginthe structuredproducts andmust not solelyrelyonthe
external credit ratings assigned to securitization exposures by the
credit ratingagencies. Abankshouldbe aware that external ratings
are a useful starting point for credit analysis, but are no substitute
for full and proper understanding of the underlying risk, espe-
cially where ratings for certain asset classes have a short history or
have been shown to be volatile. Moreover, a bank also should con-
duct credit analyses of the securitization exposure at acquisition
and on an ongoing basis. It should also have in place the necessary
quantitative tools, valuation models and stress tests of sufcient
sophistication to reliably assess all relevant risks.
Inadditionto this, The Turner Review: ARegulatory Response to
the Global Banking Crisis (Financial Services Authority 2009) has
also stressed the importance of increased capital ratios for market
riskexposures toreect the interactionbetweenmarket andliquidity
risk.
One lesson learned is that risk models are substantial parts of
a sound risk management process and important ingredients for
nancial decision making. As important as risk models themselves
is knowledge about the limitations and shortcomings of the models,
ie, the acknowledgement that risk models and their outcomes may
be wrong.
In the spirit of Socrates (we should be aware of our own igno-
rance), this book is designed to illuminate shortcomings and to
show ways overcoming the limitations within sound risk manage-
ment processes. The book examines the failings of existing nancial
risk models, and shows ways to address this model risk in existing
risk measurement and management frameworks. Aportfolio of case
studies, lessons learned and implications of the nancial crisis are
presented. Twenty groups of authors from around the world have
writtencontributions about their workexperiences andresults; these
are arranged into ve parts, organised by various risk categories.
Part I shows concepts and stochastic frameworks for model risk.
In Chapter 1 Daniel Rsch and Harald Scheule address the interac-
tion of the economy and credit-portfolio model risk. In Chapter 2
Jorge Sobehart investigates the role of imperfect information and
xxxiv
INTRODUCTION
investors behaviour. In Chapter 3 Stef Hse and Stefan Huschens
measure model risk in relation to non-Gaussian latent risk factors.
After dening model risk in general, they show the impact of a
potential misspecication of the factor distributions on credit risk
measures and derive upper and lower bounds for the value-at-risk.
In Chapter 4 Corinna Luedtke and Philipp Sibbertsen analyse time-
series properties of value-at-risk. They compare Garch-type models
with respect to their in-sample robustness and their out-of-sample
performance when the value-at-risk is forecasted using alternative
model specications. Theyshowthat various stylisedfacts mayhave
a serious impact on forecasting errors.
Part II looks at model riskingeneral economic andcapital models.
In Chapter 5 Kuang-Liang Chang, Nan-Kuang Chen and Charles Ka
Yui Leung analyse asset return dynamics and monetary policy. Oleg
Burd (Chapter 6) shows ways to manage economic and regulatory
capital through the business cycle. He nds that economic capital
is much more sensitive to stress scenarios than regulatory capital,
mainly due to maturity adjustment and asset correlation specica-
tion, and that this fact must be taken into account in the capital
management process.
Part III focuses on credit risk models. Chapter 7 Andrew Barnes,
Sean Keenan, Harry Ma, Colin McColloch, Stefano Santilli and
Sukyul Suh present their experiences on credit risk models dur-
ing the nancial crisis. Esa Jokivuolle, Oskari Vhmaa and Matti
Virn (Chapter 8) show the transmission of macro shocks to loan
losses. Lyn Thomas and Madhur Malik (Chapter 9) compare credit
risk models for portfolios of retail loans based on behavioural
scores. Michael Kalkbrener andAkwumOnwunta (Chapter 10) val-
idate structural credit-portfolio models. They review moment and
maximum-likelihood estimators for intra- and inter-sector asset cor-
relations under different distributional assumptions and analyse
their ability to capture the dependence structures. Peter Miu and
Bogie Ozdemir (Chapter 11) show the implications on estimating
and validating the probability of default if asset correlations are
stochastic. Finally, Kurt Hornik, Rainer Jankowitsch, Christoph Leit-
ner, Manuel Lingo, Stefan Pichler and Gerhard Winkler (Chapter 12)
focus on rating validation in terms of tests of the accuracy of prob-
ability of default estimates and present a latent variable approach
to validate credit rating systems. Using a large sample of Austrian
xxxv
MODEL RISK
banks and obligors, the authors conduct an extensive benchmarking
exercise.
Part IVcombines liquidity, market andoperational riskmodels. In
Chapter 13 Stefan Reitz addresses liquidity in derivatives contracts.
He shows how pricing models can be used to derive the expected
cashow for non-path-dependent and path-dependent derivatives.
Manuela Spangler and Ralf Werner (Chapter 14) are concerned with
the quantication of market risk over longer horizons. They derive
the concept of potential future market risk, a promising approach
similar to the concept of potential future exposure in counterparty
credit risk. In Chapter 15 Carsten Wehn focuses on market risk mod-
els. He systematically addresses the most common model errors in
market riskandprovides anoverviewof the most recent back-testing
approaches. AnnaChernobai, ChristianMenn, Svetlozar Rachevand
StefanTrck (Chapter 16) developoperational risk models for value-
at-risk in the presence of data biases. Robin Luo and Xiangkang Yin
(Chapter 17) analyse the operational risk for hedge funds.
Part V looks at risk transfer and securitisation models. In Chap-
ter 18 Marcus Martin models counterparty risk for over-the-counter
derivatives and develops a framework for addressing model risk
issues therein. Martin Donhauser, Alfred Hamerle and Kilian Plank
(Chapter 19) quantify the systematic risk of securitisations by con-
sidering various risk measures. The authors introduce the concept of
a bondrepresentation andexamine typical pooling andstructuring
approaches with respect to their systematic risk exposure.
ACKNOWLEDGEMENTS
We thank Joe Breeden for writing the epilogue to this book. We are
very grateful for the support from Lucie Carter and Jennifer Gibb
from Risk Books and Journals for their tremendous help in man-
aging the editing process. We hope that the book will provide new
insights for practitioners and regulators, as well as researchers on
applications, regulations and techniques presented in this book and
we encourage the reader to share any thoughts andexperiences with
our community.
Daniel Rsch and Harald Scheule
Melbourne and Hannover, November 2009
xxxvi
Part I
Concepts and Stochastic
Frameworks for Model
Risk
1
Downturn Model Risk: Another
View on the Global Financial Crisis
Daniel Rsch; Harald Scheule
Leibniz Universitt Hannover; The University of Melbourne
Researchers and practitioners have spent ample resources mod-
elling credit, explaining correlations between risk models as well
as inputs and outputs. One popular example is asset correlation,
which describes the co-movement between the asset value returns
of corporate borrowers or issuers. Other examples are default cor-
relations, correlations between default and recovery processes and
correlations between risk categories such as credit, interest, liquidity
or market risk.
Instatistical terms, correlations are oftenplaceholders for relation-
ships which cannot be explained and are also known as seeming
correlations. The 20089 global nancial crisis caught us by sur-
prise and showed that, starting with US subprime mortgage mar-
kets, other markets such as equity, credit and commodity markets
have declinedglobally. These links have not beenincludedintoexist-
ing risk models, and this chapter identies these links and shows
how to address these relationships in risk models.
We show that the insufcient incorporation of economic infor-
mation into valuation models for nancial instruments may partly
explain why the nancial industry was unable to predict, mitigate
andcover the recent losses. Economic downturns are generally well-
known. Unfortunately, todate the nancial industryhas struggledto
incorporate econometric properties into forecasting models. These
models were often propagated by the industry and supported by
a number of academic studies on the information content of credit
ratings.
We donot claim, nor intend, toaddress the nancial crisis compre-
hensivelyinthis chapter, andother experts have put complementary
3
MODEL RISK
Figure 1.1 Seasonally adjusted delinquency rates for all commercial
US banks
0
2
4
6
8
1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007
D
e
l
i
n
q
u
e
n
c
y

r
a
t
e

(
%
)
Business loans
Loans secured by real estate
Source: Board of Governors of the US Federal Reserve System and authors cal-
culations. Delinquency rates are the ratios of the US dollar amount of a banks
delinquent loans to the USdollar amount of total loans outstanding in that category.
proposals forward (Hull 2009; Crouhy et al 2008; Franke and Krah-
nen 2008). Their explanations mainly focus on misaligned incentive
structures and a lack of transparency as a consequence thereof. This
chapter provides another perspective on the lack of transparency:
the ignorance of risk models with regard to econometric proper-
ties of risk, as well as the assessment of model risk. Credit and credit
derivative markets maynot be able torecover unless these important
issues have been resolved.
CREDIT RISK AND BUSINESS CYCLES
Figure 1.1 shows a proxy for credit-portfolio risk, the delinquency
rate. It is apparent that the delinquency rate, and thus credit risk,
changes over time and follows cyclical patterns. For instance, the
years 1991 (rst Gulf War) and 20012 (terrorist attacks on the US)
were periods of high delinquency rates for business loans. Delin-
quency rates for business loans have changedsurprisingly little dur-
ing the current (20089) nancial crisis, while loans secured by real
estate have dramatically increased.
Generally speaking, the risk may be measured by two funda-
mentally different approaches (Rsch and Scheule 2005). Firstly, we
can take the average over the business cycle; this is known as the
through-the-cycle (TTC) approach. Secondly, we can try to measure
4
DOWNTURN MODEL RISK: ANOTHERVIEW ONTHE GLOBAL FINANCIAL CRISIS
Figure 1.2 Through-the-cycle (TTC) model and a point-in-time (PIT)
credit risk model
0
2
4
6
8
1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007
D
e
l
i
n
q
u
e
n
c
y

r
a
t
e

(
%
)
GAP
GAP
GAP
Delinquency rate
TTC
PIT
Source: Board of Governors of the USFederal Reserve System. Delinquency rates
are the ratios of the US dollar amount of a banks delinquent loans to the US dollar
amount of total loans outstanding in that category. A TTC model assumes the
average default rate for every period. A PIT model forecasts the default rate based
on an empirical model.
the credit risk for a given point in time; this is known as the point-
in-time (PIT) approach. PIT models are generally based on forecast
models, which explain the credit risk for a future point in time, by
information which is available at the time when the forecast is made.
Figure 1.2 shows the real delinquency rate for loans secured by
real estate, as well as the estimateddelinquency rate by a TTCmodel
and a PIT model.
It can be seen that neither model estimates the default rate accu-
rately. However, PIT models generally approximate the default rate
better for most points in time. Thus, a PIT model reects the reality
muchbetter andshouldbe the aimof every soundrisk measurement
framework.
Unfortunately, the majority of the nancial industry focuses on
TTC approaches. Various reasons for this deserve to be mentioned.
Simplicity. TTC approaches have gained acceptance because
they offer simplicity. PIT models have been propagated but
estimated based on modest recent loss experiences due to lim-
ited data availability. In other words, building a risk model
based on the experience of multiple boom years may be
inadequate to provision for credit losses during downturns.
5
MODEL RISK
Regulatory requirements. Regulators have accepted both TTC
and PIT methods but have often preferred TTC methods
(Financial Services Authority 2009). Inaddition, withthe intro-
duction of Basel II, the concern was raised that the capital
of nancial institutions may uctuate with the credit risk
assessment during the business cycle. This pro-cyclicality may
require the issue of newcapital during an economic downturn
when capital is expensive or restricted in availability. In other
words, regulators tried to avoid pro-cyclicality by accepting
risk models which do not take the current state of the economy
into account. The current crisis demonstrated that accounting
practice implies a pro-cyclical capital requirement for market
risk exposures, as the accounting values of marketable assets
based on current market values, delinquent credit exposures
are provisioned for and defaulted credit exposures are written
off.
Guidance by rating agencies. Rating agencies provide cate-
gorical ratings; common rating categories are AAA(Aaa), AA
(Aa), A(A), BBB (Baa), BA(Ba), B (B), CCC (Caa), CC (Ca) and
C (C) for Standard & Poors rating agency and Fitch rating
agency (Moodys rating agency). These agencies have histor-
ically propagated TTC models and have explicitly excluded
the economy and focused on idiosyncratic risk drivers which
were considered to be fundamental. For these efforts, credit
ratings reect an opaque mix of a TTC and PIT model out-
put, as some idiosyncratic information naturally reects the
business cycle. As a result, the degree of cyclicality which is
embedded in public credit ratings is difcult to assess and
investors are uncertain as to whether they should associate
time-constant, time-varying (or a mix of both) default rates to
these ratings categories. Rating agencies may have no incen-
tive to change this opaque practice, as the crucial calibration
step (ie, the conversion from categorical ratings to numeric
default rates needed by modern risk models) lies within the
responsibility of investors.
The result of using a through-the-cycle approach is obvious: the
model positively surprises in an economic boom, as the loss out-
come is less than predicted by the model and disappoints in an eco-
nomic downturn (eg, the 20089 nancial crisis) as the loss outcome
6
is higher than predicted by the model. As a result, parties that relied
on these models were disappointed in the 20089 crisis. This may
be conrmed by the public criticism of rating agencies during the
current and previous nancial crises.
CORRELATIONS
Incredit-portfolio risk measurement, correlations play a central role.
Credit defaults happen very rarely and a corporate borrower often
ceases to exist after a default event. Thus, for any given pair of
borrowers, it may be difcult to estimate correlations, as multiple
default observations are not available. Some nancial institutions
apply expert values between zero and one. For instance, the default
correlation for two borrowers which are part of the same holding
company may be set to one.
This apparent co-movement is generally drivenby systematic risk
factors and may be quantied by specifying an econometric factor
model or correlations if a credit-portfolio risk model is unable or
unwilling to include parts of the systematic (economic) informa-
tion. In other words, correlations may be derived which capture the
gap between a portfolios actual and a models predicted credit loss.
Figure 1.2 shows that the gap is dependent on the period as well as
the applied model. In other words, the chosen model methodology
(TTC or PIT) has a major impact on the correlations. The correlation
is high for a TTC model and low for a PIT model. This is empiri-
cally conrmedby a number of studies including Rsch andScheule
(2005).
1
Generally speaking, the econometric estimation of correlations
requires the availability of long data histories, ie, multiples of whole
business cycles. Note that the estimationof correlations is more com-
plicated for securitisations such as asset-backed securities, collater-
alised debt obligations and mortgage-backed securities. Many of
these products have been issued recently and long default histories
are unavailable. We will refer to this point below.
CREDIT PORTFOLIO RISKS
Credit-portfolio risk models aggregate the risk characteristics for
individual borrowers on a portfolio level. The nancial industry has
identied a set of key risk drivers which are probabilities of default,
7
MODEL RISK
Figure 1.3 Credit-portfolio loss distributions
0 10 20 30 40 50 60 70 80 90 100
PIT (boom) PD = 0.5%
PIT (recess.) PD = 10%
TTC PD = 1%
Portfolio loss (%)
D
e
n
s
i
t
y
A TTC model assumes the average default rate for every period. A PIT model
forecasts the default rate based on an empirical model.
loss rates given default, exposures at default and default correla-
tions. Alarge literature exists for everyone of these parameters. Note
that default events are oftengeneratedbasedonso-calledasset value
models, which rely on asset rather than default correlations. Proba-
bilities of default, loss rates given default and exposures at default
describe expectations for random variables. For example, a proba-
bility of default generally describes the expectation of occurrence, ie,
the likelihood of a default event during an observation period. Con-
trarily, the default or non-default observation describes the outcome
of this random variable.
In an uncertain world, the credit loss of a portfolio is random,
and modern credit-portfolio risk measurement focuses on the quan-
tication of the loss distribution. A loss distribution describes the
frequency for various levels of future losses. Figure 1.3 shows that
the loss distribution depends on the risk of the individual borrow-
ers (eg, the probability of default, PD) as well as the correlations (eg,
the asset correlation). Figure 1.3 shows three loss distributions: one
is based on a PIT model during an economic boom, one is based
on a TTC model and the other is based on a PIT model during an
economic downturn.
In this example, the 99.9th percentile which is often used as a
proxy for credit-portfolio risk is 7%, 15% and 46%. This may lead to
the interpretation that during an economic downturn the PIT model
may provide the highest risk assessment and thus the highest level
of capital, and thus the highest required level of protection for credit
8
Figure 1.4 Volume of credit derivative transactions
0
5,000
10,000
15,000
20,000
25,000
30,000
35,000
2002 2001 2000 2003 2006 2004 2008
G
l
o
b
a
l

c
r
e
d
i
t

d
e
r
i
v
t
i
v
e
s

v
o
l
u
m
e
(
U
S
$

b
i
l
l
i
o
n
)
1996 1998 1999
Source: British Bankers Association.
losses. Conversely, nancial institutions which rely on TTC models
dramatically underestimate credit-portfolio risks during economic
downturns (15%versus the reasonable number of 46%)! Similar con-
clusions may be drawn for other risk parameters such as loss rates
given default and the correlation between the default and recovery
processes (Rsch and Scheule 2009a, 2010).
SECURITISATIONS
Securitisations involve a real or synthetic sale of an asset portfolio to
investors. Common assets are of a nancial nature such as insurance
policies, leases or loans. Many securitisations involve large mone-
tary values and investors are pooled. Important investors are gen-
erally large nancial institutions and hedge funds. Some countries
like Australia and New Zealand enable retail investors (also known
as mums and dads) to trade a small selection of such securities at
public exchanges. The following gures describe the credit deriva-
tive market prior to the nancial crisis. Figure 1.4 shows that the
global monetary volume of credit derivatives has increased expo-
nentially during the past decade. Credit derivative and securitisa-
tion markets shared similar trends in the past. The numbers for 2008
are projections made before the nancial crisis. Figure 1.5 identies
banks andhedge funds as the mainbuyers of credit protection, while
Figure 1.6 identies banks, insurance companies and hedge funds
as the main sellers of credit protection. Figure 1.7 shows that the
9
MODEL RISK
Figure 1.5 Buyer of credit risk protection
80
70
90
50
40
60
20
10
0
30
F
r
a
c
t
i
o
n

o
f

g
l
o
b
a
l

m
a
r
k
e
t

(
%
)
Banks Hedge
funds
Corporate Insurance Others
2000 2002 2004 2006 2008
Figure 1.6 Seller of credit risk protection
70
50
40
60
20
10
0
30
F
r
a
c
t
i
o
n

o
f

g
l
o
b
a
l

m
a
r
k
e
t

(
%
)
Banks Hedge
funds
Corporate Insurance Others
2000 2002 2004 2006 2008
maturities of newsecuritisations are generally between one and ve
years.
Different investors have different requirements in relation to the
riskreturn prole of their investments. Thus, large asset securi-
tisations generally involve the separation (tranching) of investors
into distinct riskreturn categories where the proceeds from the
asset portfolio are forwarded according to an agreed set of rules.
Most securitisations have a unique structure which may or may not
be different from the examples presented here. The International
10
Figure 1.7 Average maturity of new credit derivatives
50
40
60
20
10
0
30
F
r
a
c
t
i
o
n

o
f

g
l
o
b
a
l

m
a
r
k
e
t

(
%
)
<1 year 15 years 510 years >10 years
2000 2002 2004 2006 2008
Swaps and Derivatives Association (ISDA)
2
has published guide-
lines to improve the standardisation of securitisations and credit
derivatives.
In the example presented, investors invest in a junior tranche,
a mezzanine tranche and a senior tranche. The yield to maturity
decreases from the junior to the senior tranche. Typical yields may
include a credit spread (above a reference rate) of 30% for junior
investors, 3% for mezzanine investors and 50 basis points for senior
investors. Generally, 10% of the total investment amount may be
raised with junior investors, 20%with mezzanine investors and 70%
with senior investors.
The proceeds are forwarded to the senior, then to the mezzanine
and lastly to the junior investors according to predetermined rules if
the asset portfolio cashows are sufcient. The junior tranche is also
knownas the equitytranche due tothis conditional andthus residual
payout policy. If a tranche does not receive an agreed payment, it is
impaired. The concept of impairment is comparable to the concept
of default for loans. Impairment is most likely to happen upon the
payment of the largest contractual amount: the principal. Figure 1.8
shows that principal impairments are far more commonthaninterest
impairments.
In addition, Figure 1.8 shows that the number of impairments
increased dramatically during the GFC. There may be several
reasons for this. Firstly, the economy is currently
3
experiencing a
11
MODEL RISK
Figure 1.8 Interest and principal impairments of securitisations
Interest payment year
Principal payment year
I
m
p
a
i
r
m
e
n
t

r
a
t
e
0.2
0
0.4
0.6
0.8
1.0
1997 1999 2001 2003 2005 2007
Source: Rsch and Scheule (2009b), Moodys credit rating agency. Impairments
comprise principal and interest impairments. Principal impairments include secu-
rities that have suffered principal write-downs or principal losses at maturity and
securities that have been downgraded to Ca/C, even if they have not yet experi-
enced an interest shortfall or principal write-down. Interest impairments, or interest-
impaired securities, include securities that are not principal impaired and have
experienced only interest shortfalls.
Figure 1.9 Credit-portfolio loss distributions
Junior Mezz. Senior
D
e
n
s
i
t
y
Portfolio loss (%)
0 10 20 30 40 50 60 70 80 90 100
PIT (boom), PD = 0.5%
TTC, PD = 1%
PIT (recess.), PD = 10%
A TTC model assumes the average default rate for every period. A PIT model
forecasts the default rate based on an empirical model.
downturn. In particular the asset class US mortgage loans are expe-
riencing a major stress. Secondly, a comparison of Figures 1.1 and1.4
reveals that the growth in the securitisation market occurred dur-
ing an economic boom, ie, good years. This has the implication
that no market participant has a loss experience, which comprises
a whole business cycle to calibrate econometric models. Thirdly, a
12
Figure 1.10 Spread sensitivity of a senior tranche
0
20
40
60
80
100
0
40
80
120
160
200
S
p
r
e
a
d

(
b
p
)
S
p
r
e
a
d
(
m
i
l
l
i
o
n
s
,
%

o
f

5
%

b
e
n
c
h
m
a
r
k
)
Spread (bp)
Spread (% of 5% benchmark)
0 0.1 0.2 0.3 0.4 0.5
Correlation
comparison of Figures 1.8, 1.4 and 1.7 provides evidence that prin-
cipal impairments are the most common, that most securitisations
were originated in recent years (ie, the past 15 years) and that most
securitisations pay back the principal after 15 years. These factors
may explain the low impairment rates in recent years which are
no longer true. Fourthly, market participants may have relied on
TTCmodels. Figure 1.9 shows the implications for the example from
Figure 1.3 (junior tranche equals 10%, mezzanine tranche 20% and
senior tranche 70% of total assets).
It is obvious that the attachment probability is higher for the mez-
zanine and senior tranches during an economic downturn and the
PIT model. As the PIT model is more accurate than the TTC model,
the use of the latter results in a dramatic underestimation of credit
riskfor mezzanine andsenior tranches duringeconomic downturns.
The attachment probabilities for losses for the mezzanine tranche are
0.4% in the TTC model and 39.8% in the PIT (recession) model. The
attachment probabilities for losses for the senior tranche are 0% in
the TTC model and 2% in the PIT (recession) model. In summary, a
TTCrating model may underestimate the risk to the mezzanine and
the senior tranches dramatically.
Generally speaking, tranche credit risk as measured by a risk
model is crucially inuenced by the rating methodology and the
correlations used in the model. Chernih et al (2006) show that the
correlationmay vary from0.5%to 45%, depending onthe estimation
methodology and the data used.
13
MODEL RISK
Figure 1.11 Impairment rates by rating category
1997 1999 2001 2003 2005 2007
0.8
0.6
0.4
0.2
0
I
m
p
a
i
r
m
e
n
t

r
a
t
e
AaaA Baa Ba B
Source: Rsch and Scheule (2009b), Moodys credit rating agency. Impairment
rates per Moodys rating category. A TTC model results in increased default or
impairment rates during an economic downturn such as the nancial crisis.
As an example for the correlation sensitivity of tranches, consider
a tranching which employs a simple credit risk model to calculate
the fair spread of the senior tranche. Figure 1.10 shows on the left
axis that the impliedspreadof the tranche is close to zero for lowcor-
relations andincreases monotonically with the correlation. The right
axis exhibits the spread measured relatively to the spread which is
based on a correlation benchmark of 5%. For higher correlations the
spreadmay be upto a multiple of many hundreds of thousandtimes
of the benchmark.
Rating agencies are one of the most prominent proponents of TTC
models as well as levels of asset correlations. Figure 1.11 shows the
impairment rates of securitisations rated by Moodys rating agency.
The impairment rates per rating category uctuate over time as
the ratings are TTC. See Rsch and Scheule (2009b) for a discussion
of the performance of public credit ratings of securitisations.
CATALYST I: RESECURITISATIONS
It was shown by Hull (2008) that demand for medium-rated mezza-
nine tranches was limited before the 20089 nancial crisis. Hence,
the originating institutions pooled unsold mezzanine tranches and
resecuritised these assets, which led to a new set of junior, mezza-
nine and senior investment tranches. Resecuritisations are found to
14
magnify the problempresented in the previous section. The implied
correlations may be higher for resecuritisations thanfor plainvanilla
asset securitisations.
CATALYST II: CONCENTRATION OFTHE MODEL PROVIDER,
FINANCIAL INTERMEDIATION AND MODEL AUDITING
INDUSTRY
The model provider, nancial intermediation and model auditing
industry are highly concentrated. The concentration leads to sys-
temic risk. Several examples sufce: the small number of credit rat-
ing agencies for bond and structured nance issues; the growing
market share of too big to fail nancial institutions; joint ven-
tures in model construction designedto reduce costs. The problemis
compounded by the use of similar quantitative frameworks and/or
frameworks which are calibrated based on similar loss experiences.
An anecdote illustrates this point: some years ago the chief risk of-
cer of a major US bank presented the asset correlation matrix used
by that institution. Another major nancial institution present at
the event conrmed its use of the same matrix. While the institu-
tions were fundamentally different in nature, they shared the same
reputable consulting rm. To date, the rms model has not been
formally validated. The oligopolistic structure was nurtured and
promoted by the data available to a limited few as well as by the
propensity of nancial institutions to outsource risk modelling.
WHAT LESSONS CAN BE LEARNED GOING FORWARD?
Franke and Krahnen (2008) show that structured nance transac-
tions offer many benets suchas improvedrisk allocationanddiver-
sication. Therefore, a strict ban of securitisations may contradict
market efciency. However, the lack of transparency leads to illiq-
uidity of the associated assets and liabilities and is thus a major
driver of the current nancial crisis. Appropriate risk measurement
and management may have to complement government nancial
stabilityprogrammes toprovide equity, debt or nancial guarantees.
Our main objective in this chapter was to show the impact of
the economy on credit-portfolio risk and derivative products. The
inclusionof the impact of the economyis paramount toa soundmea-
surement and management of credit-portfolio risks. Historically, the
15
MODEL RISK
nancial industry did not exploit all the available econometric infor-
mation. Transparencymayincrease withthe econometric forecasting
of credit-portfolio risks. Credibility in internal and external models
may not be restored otherwise.
The following complementary suggestions may contribute addi-
tional elements for a new framework of global nancial markets.
Homogenisation and renement of regulation
This chapter recommends changes in regulations for the following
areas.
Bank models should be point-in-time and be able to forecast
the credit risk for future periods with a reasonable degree of
accuracy.
Regulation should address pro-cyclicality. Firstly, it has to be
determined whether nancial institution capital should be
pro-cyclical, neutral or countercyclical. While regulators in
the past have tried to avoid pro-cyclicality, the 20089 crisis
supports dynamic capital adequacy regulations which antici-
pate economic downturns. Secondly, requirements may have
to be developed which take the individual risk drivers of cap-
ital (ie, assets, liabilities and off-balance-sheet activities) into
account. One example for an inconsistent process is that assets
on the trading book are marked-to-market, while assets on
the banking book are marked-to-delinquency or marked-to-
default. Capital andcredit markets maybe correlatedbut share
quite different econometric patterns.
Existing regulations should be homogeneous across countries
and industries. Past securitisations were often inspired by dif-
ferences inregulations betweencountries, industries andother
categories. The homogenisation may
improve the transparency,
regulate previously unregulatedindustries closely linked
to regulated industries and avoid transactions which
aim to exploit regulatory differences (also known as
regulatory arbitrage transactions).
In addition, stricter rules for information disclosure beyond
Basel II may have to be implemented and potentially high risk
strategies limited. One example is the possible limitation of
securitisations to older and thus better known loan vintages.
16
These changes have to be carefully assessed for their ability to pre-
vent similar crises in the future. This also includes other recent
proposals such as compulsory retention levels for securitisations,
renement of capital adequacy rules or the introduction of alterna-
tive incentive mechanisms. These changes, while benecial intheory,
maycontradict constitutional rights, the needfor transparencyor the
function of nancial institutions as asset transformer and nancial
intermediary, or they may propagate nancial crises.
Pooling and publication of information
One major challenge to pricing and research is the availability of
data for various stages of the business cycles. As a result of this lim-
itation, past research has often focused on a few data sets which
are publicly available. It may be important to make data, in partic-
ular credit loss histories which are pooled but not aggregated over
lenders, available to researchers, nancial institutions and their reg-
ulators. A global and compulsory data warehouse for credits may
be a rewarding approach. The data may be led anonymously and
be validated by the national regulators. In addition, models of var-
ious participants may have to be transparent and available in an
open source environment. This includes, in particular, internal mod-
els of rating agencies. With regard to gaming (eg, the provision of
biased information of an issuer to obtain a favourable rating), our
experience indicates that gaming is concentrated in areas where no
objective models exist.
Evaluation of model risk and stress-testing of risk models
For pricing and risk measurement of complex nancial products,
models have to be used. Each model relies on a number of assump-
tions andtherefore provides onlya simplicationof reality. As recent
research has shown, the risk measures and implied prices of struc-
tured products are highly sensitive to the underlying pricing model
andthe assumptions. Model riskcannot easilybe mitigatedbydiver-
sication and may therefore lead to substantial systemic risk. There
are only a limited number of providers of risk models using similar
assumptions in their models. Sharing the same advisors may imply
that many banks use the same models, and thus, model errors may
be consistent across the entire industry. In other words, these mod-
els are called market standard, but standard may not imply
quality.
17
MODEL RISK
Facilitation of knowledge transfer
Financial markets are large and complex. Their scope may be mea-
sured in number of products, monetary volume or, more impor-
tantly, the number of employees. Tocope withthe involvedcomplex-
ities, a sound education in disciplines such as econometrics, nance,
mathematics or statistics is paramount. Unfortunately, education in
these markets has not matched the growth. Training has been kept
within nancial institutions, with the aim of prot maximisation
rather than industry stability. This problem may be addressed by
sponsoring existing and new credit risk research centres organised
by the government, supported by the industry and academia. The
initiative may be organised under the national rescue packages and
comprise a marginal cost in the budget.
1 See also Chernih et al (2006) for a comparison of existing empirical studies.
2 See http://www.isda.org.
3 At the time of writing in October 2009.
REFERENCES
Chernih, A., S. Vanduffel and L. Henrard, 2006, Asset Correlations: ALiterature Review
and Analysis of the Impact of Dependent Loss Given Defaults, Working Paper, X-act
Consulting, Vrije Universiteit Brussels, Fortis.
Crouhy, M., R. Jarrow and S. Turnbull, 2008, The Subprime Credit Crisis, Journal of
Derivatives, 130.
Financial Services Authority, 2009, The Turner Review: A Regulatory Response to the
Global Banking Crisis, Report, UK Financial Services Authority, March.
Franke, G., andJ. P. Krahnen, 2008, The Future of Securitization, Working Paper, Center
for Financial Studies, Goethe-Universitt Frankfurt.
Hull, J., 2009, The Credit Crunch of 2007: What Went Wrong? Why? What Lessons Can
Be Learned?, The Journal of Credit Risk 5(2), pp. 318.
Rsch, D., and H. Scheule, 2005, A Multi-Factor Approach for Systematic Default and
Recovery Risk, Journal of Fixed Income 15(2), pp. 6375.
Rsch, D., and H. Scheule, 2009a, Credit Portfolio Loss Forecasts for Economic
Downturns, 2009, Financial Markets, Institutions and Instruments 18(1), pp. 126.
Rsch, D., and H. Scheule, 2009b, Rating Performance and Agency Incentives of
Structured Finance Transactions, Working Paper.
Rsch, D., and H. Scheule, 2010, Downturn Credit Portfolio Risk, Regulatory Capital
and Prudential Incentives, forthcoming, International Review of Finance.
18
2
Follow the Money from
Boom to Bust
Jorge R. Sobehart
Citi Risk Architecture
THE ROLE OF IMPERFECT INFORMATION AND
INVESTORS BEHAVIOUR
There is hardly a more difcult and timely subject in economics
today than the prediction, or even the analysis, of nancial crises,
market reaction and its impact on securities and asset prices.
Changes in securities prices and market reaction are rarely the ide-
alisedrandomevents denedinprobability textbooks andacademic
literature. Market reactionandprice changes always occur as a result
of a particular sequence of events, which might be unknown to
individual investors, or might be too complex to model or forecast.
Insightful discussions on market dynamics and their historical con-
text canbe foundinMalkiel (1981), Jones (1991), Kindleberger (2000),
Mahar (2003) and Bookstaber (2007).
Most technical discussions on securities prices are based on the
assumption that prices are driven by the activity of rational mar-
ket participants, whose aggregateddecisions result in instantaneous
relative price changes that can be represented by Brownian motion
driven by noise with a (normal) Gaussian bell-shaped distribution.
In general, Gaussian noise provides a reasonably good approxima-
tion of many types of real noises. Note, however, that the Gaus-
sian bell-shaped distribution does not hold a monopoly on possible
limit distributions despite the general misconception that adding up
many independent random effects immediately requires the use of
normally distributedrandomchanges. Evenif relative price changes
converge toGaussiannoise, slowconvergence canleadtoobservable
non-normal effects over extended periods of time (Gopikrishnan et
al 1999; Sobehart and Farengo 2002, 2003).
19
MODEL RISK
The origin of the standard random-walk assumption for price
movements can be traced back at least as far as the pioneering con-
tribution of Bachelier (1900). Early in the development of pricing
models, Kendall (1953) highlighted the limitations of the standard
random-walk assumption thus:
It may be that the motion [of stocks] is genuinely random and
that what looks like a purposive movement over a long period is
merely a kindof economic Brownian motion. But economists and
I cannot help sympathizing with them will doubtless resist any
such conclusion very strongly.
Empirical work on the distribution of price movements and the
observation that most investors and equity portfolio managers do
not consistently outperform passive benchmarks over long periods
of time, combined with the application of the theory of stochas-
tic processes to nancial markets, led to the formulation of an ele-
gant and technically sound theory of efcient markets based on the
standard random-walk assumption. The modern literature on price
movement andrandomwalks usually begins withSamuelson(1965)
and his application of the lawof iterated expectations to anticipated
prices, although it was the articulation of what efcient markets
mean in the 1970s (Fama 1970) that nally solidied the efcient
market hypothesis framework andthe idealisedassumptionof price
movements as normal random walks.
Debates about the relevance of investor behaviour were usually
dismissed by arguments such as non-rational investors would tend
to lose money, giving these investors a short enough life to make
themirrelevant, or under- and overreactions to price changes would
produce on average no observable net effect. But, as attractive as the
standard random-walk assumption is, it is not perfectly consistent
with the properties of historical returns. Evidence accumulated over
the years has shown consistently that investors are far less rational
in their decision making than the theory assumes (DeBondt and
Thaler 1985; DeLong et al 1990). Note, however, that investors irra-
tionality is neither random nor senseless and sometimes can mani-
fest itself as predictable and systematic. Analysis of the distribution
of changes in securities prices indicates that returns systematically
have more mass inthe tail areas thanwouldbe predictedbya normal
distribution (excess kurtosis).
20
FOLLOWTHE MONEY FROM BOOMTO BUST
Over the years, three basic approaches have been proposed in
order to extend the standard Brownian model of asset returns to
include the observed effects. The rst approach (known as the local
volatility model) makes price volatility a function of the underly-
ing securitys price, which assumes, counterfactually, that price and
volatility are perfectly correlated. The second approach increases
the likelihood and severity of price changes by introducing random
variations of volatility, interest rates and other state variables. The
third approach includes models that replace the standard random-
walk process with jump processes, Lvy ights, power laws, frac-
tal random walks or similarly complex stochastic processes (Cont
and Bouchard 1998; Bibby and Sorensen 2001; Barndorff-Nielsen
and Sheppard 2001; Andersen and Sornette 2002; Dragulescu and
Yakovenko2002; AntonuccioandProebsting2003; Sornette et al 2003;
Metzler and Klafter 2004; Gabaix 2008). These non-standard mod-
els reect a wide range of behavioural effects based on prospect
theory, regret and cognitive dissonance, anchoring, overcondence,
over- and underreaction, the representativeness heuristic, irrational
speculation, herding, contagion effects and a variety of documented
cognitive biases that may affect risk perception and decision mak-
ing. A detailed review of the literature on the subject can be found
in Hirshleifer (2001).
Although there is no shortage of alternative stochastic models,
none of these models has been rmly established to date because
of their mathematical complexity, lack of economic interpretation or
acceptance among academics andpractitioners. More recently, alter-
native models based on investors behaviour have been proposed
with different degrees of success. Here we discuss a phenomenolog-
ical model of price dynamics that includes investor behaviour pat-
terns based on price momentum and trends, and produces realistic
fat-tailed distributions of price returns in agreement with historical
data over a wide range of time horizons.
Before performing any formal analysis, we step back and identify
some obvious problems with the standard assumption of normally
distributed random changes commonly associated with models of
rational markets and perfectly informed market participants.
One of the fundamental tenets of most of these model extensions
is the assumption that market participants are perfectly rational in
the sense that not only do they act in their own self-interest but they
21
MODEL RISK
do not follow price trends because price changes are unpredictable
in nature and have no memory effects. The argument is that, at each
instant, prices correctly reect assets true values, based not only on
current economic conditions but also on the best estimates of how
those conditions will evolve inthe future. Inthis idealisedviewof the
markets, changes in prices are only the result of market participants
responding to a constant ow of new information on fundamen-
tals or exogenous shocks (the efcient market hypothesis). In this
frameworkthere is no roomfor bubbles, busts or non-randomexcess
returns. Notice, however, that investors and arbitrageurs are limited
in their ability to restore price changes instantaneously, which may
create temporary price distortions. Adiscussion on price distortions
for assets and derivatives due to market uncertainty and trading
limitations is given in OHara (1998), Sobehart (2005), Sobehart and
Keenan (2005) and references therein.
Even if security prices are set in the aggregate under quasi-
equilibrium conditions that may eliminate sustainable arbitrage
opportunities and price distortions, the equilibrium may not nec-
essarily be rational in the restricted sense described above (Var-
ian 1985; Fama 1970, 1998; OHara 1998; Thaler 1999; Shefrin 1999;
Shleifer 2000 and references therein). An illustrative example of
these market dynamics is the 2000 irrational exuberance bubble,
when investors kept buying Internet and telecoms stocks driven
by price momentum despite the lack of sustainable earnings and
the poor business models of those rms. Similar bubbles occurred
in the tronic boom and the Nifty Fifty craze of the 1960s and
1970s (Malkiel 1981). Apathological repetition of this phenomenon
was observed in the period 20089 when the rapid deterioration
of credit conditions combined with high levels of uncertainty about
the viability of many nancial institutions ledto dramatic changes in
investors condence on a daily basis. This was reected in an over-
all price decline in markets around the world, with large daily price
swings. Inaddition, evenfullyinformedsecurities prices are actually
marginal prices and, therefore, primarily reect marginal realloca-
tion due to supply and demand conditions rather than the funda-
mental value of the aggregate equityor debt of arm. The small daily
volume of trades relative to the aggregated number of shares held
by investors and price swings that accompany infrequently large
spikes in trade volume or takeover bids support this point.
22
Purely rational behaviour, in the strict sense used in economics
and nance, is more a convenient assumption in order to make the
pricing problem analytically tractable than a realistic assumption
about investors behaviour. Classical rational models simply do not
address the impact of investor psychology on investment goals and
strategies. In practice, emotions and feelings are deeply enmeshed
in the human reasoning process and can override rational decisions
under stressful or uncertain conditions. Furthermore, recent physio-
logical and psychological studies have revealed that certain aspects
of emotions and feelings are indispensable in rational decision mak-
ing(Damasio1994; Ariely2008). Emotions andfeelings assist us with
the daunting task of predicting in an uncertain environment and
planning our actions accordingly without having to exercise logical
deduction or induction in each action, including making investment
decisions based on perceived price momentum and trends regard-
less of the technical soundness of their estimation. After all, trading
strategies based on price trend analysis (or technical analysis) are
commonly usedby professional andcasual investors, andhave been
popularised through countless books and trade articles.
Investors behaviour, irrationality and their impact on markets
are not new topics. Economists have been debating the effect that
investors behaviour, imperfect information and trading noise have
on securities prices for a while (Campbell et al 1996; Thaler 1999;
Shiller 2000; Hirshleifer 2001; Baker and Stein 2002; Sobehart 2005).
Here we extend the debate by revisiting a recent model of securities
prices based on the under/overreaction of market participants to
price momentum(Sobehart and Farengo 2002, 2003; Sobehart 2003).
The addition of a small amount of behavioural noise due to under-
and overreaction to price trends allows us to produce realistic fat-
tailed distribution of price returns for different time horizons that
would be absent in a perfectly rational world.
More importantly, under the standard assumption of random
walks with normally distributed returns, extremely severe price
changes are absent, as their probability is not only embarrassingly
small but also results in serious underestimation of risk. In practice,
the frequency of extreme events (including bubbles and crashes) is
sometimes orders of magnitude larger than what the extrapolation
based on small and moderate events would suggest under a normal
distribution of returns based on the behaviour of rational investors
23
MODEL RISK
(Bouchaud et al 1999; Andersen and Sornette 2002; Gabaix 2008).
As market crises, rallies and bubbles reoccur, it is becoming harder
for supporters of the idealised random-walk models to water-down
these events and rationalise them as being perfectly explainable
mathematical ukes.
INVESTOR BEHAVIOUR AND PRICE DYNAMIC
We begin our analysis by describing a standard model for securities
prices and then introduce a model that includes a behaviourally
motivated form of price momentum and trends. Let = log(S/S
0
)
be the securities logprice, where Sis the instantaneous price andS
0
is
a reference price. The standard assumption is that a securities return
follows a Brownian motion with normally distributed independent
increments (ie, a Wiener process). That is
d = (
1
2
2
0
) dt +
0
dZ (2.1)
Here the random variable Z follows a Wiener process, is the secu-
ritys growth rate and
0
is its volatility. (Equation 2.1 is usually
written in the more familiar form dS/S = dt +
0
dZ.)
Next we incorporate a specic form of cognitive biases of market
participants basedonprice momentumconsistent withsome trading
patterns of investors described in the literature (Thaler 1999; Shiller
2000; Hirshleifer 2001; Baker and Stein 2002). The resulting price
dynamic yields realistic fat-tailed distributions of returns that allow
extreme price changes and match observed frequencies for a wide
range of time horizons.
In principle, an ideal model of the dynamics of security prices
must include a market-microstructure description of supplyde-
mand dynamics and investors preferences that lead to changes in
returns. Here we simplify the problem by including a phenomeno-
logical description of the basic features of the dynamics of security
prices when market participants under/overreact to price trends,
forcing the securitys return to randomly deviate from its funda-
mental value. Our model reduces to Equation 2.1 when all market
participants are rational (in the sense that they do not follow price
trends) and information is incorporated into prices instantaneously.
The economic interpretation of our model is as follows. Investors
have some prior views about the company in question and some
idiosyncratic responsiveness to new information. We describe this
24
situation with a (dissipative) mean-reversion process for the secu-
ritys excess return above (below) the mean return. During peri-
ods of uncertainty, investors may tend to react to past price trends
as opposed to actual news or fundamental analysis. For exam-
ple, bullish investors may attach themselves to an over-optimistic
view of the company during upward price trends. Similarly, bear-
ish investors may attach themselves to an over-pessimistic view
during a downward price trend. In doing so, they discount the
possibility that the bullish (bearish) price trends are the result of
aggregated behavioural dynamics rather than a change in the future
prospects of the company. This gives rise to the market overreac-
tion to upward and downward price moves that can create posi-
tive feedback and, therefore, potentially unstable dynamics for both
stock prices and excess returns (Shleifer 2000; Sobehart and Farengo
2002). We describe this situation, assuming that the securitys return
candeviate temporarilyfromits fundamental value throughthe ran-
domfeedbackprocess causedbythe over- andunderreactiontoprice
trends.
Technically, we use two random variables to model the effect
described above: the log price and the excess return of the
security above/below the expected mean return . The stochas-
tic processes driving variables and have embedded in them
assumptions reecting bullishbearish directional changes and ran-
domfeedback. More precisely, the evolution of log prices is assumed
to be driven by an equation similar to Equation 2.1 except that it
includes anextra termthat reects randomchanges ininstantaneous
excess returns
log-price adjustment: d = ( +
1
2
2
) dt +
0
dZ
0
(2.2)
The parameter
0
in Equation 2.2 was introduced for complete-
ness andcanbe generalisedtoinclude other randomprocesses. Note,
however, that the fat-tail effects described in this chapter are driven
mainly by the dynamics of the excess return .
Followingthe economic interpretationintroducedearlier, the evo-
lution of the excess return is driven by a mean-reverting Ornstein
Uhlenbeck process that includes random adjustments to the speed
of mean reversion (multiplicative noise) as well as random market
corrections of returns (additive noise)
excess return adjustment: d = ( dt + dZ
1
) +dZ
2
(2.3)
25
MODEL RISK
Here the parameter is the average sensitivity of market partic-
ipants to price return discrepancies, which leads to mean reversion
of returns, is the strength of the multiplicative price momentum
corrections and is the magnitude of random additive changes in
excess return resulting from investors under- and overvaluation of
returns.
When investors evaluate price changes they must do so against a
backdrop of economic uncertainty with volatile and cyclical mar-
kets. Market booms and crashes cannot be dissociated from the
level of economic activity and business cycles. However, the analy-
sis of behaviour patterns through average market conditions can
help to identify basic features of the price change distribution. These
ideas can then be generalised to include the effects of economic and
business cycles. For simplicity, parameters ,
0
, , and are
assumed constant, and 0 for a stable mean-reverting process.
For analytical convenience we also assume that the random vari-
ations of the noise and mean-reverting processes are independent
and normally distributed. That is, the random variables Z
0
, Z
1
and
Z
2
followindependent Wiener processes. Inpractice, high-frequency
excess returns sometimes exhibit a small linear autocorrelation that
is less signicant for longer time horizons. Model extensions that
include time-dependent parameters and correlations require more
demanding calculations, but are conceptually straightforward.
The effective volatility in Equation 2.2 is a function of the
key parameters of the model, as shown later, in the section on dis-
tribution of returns over time (see page 33 onwards). For nearly
rational markets, the effective volatility (also derived in that section)
is approximately
( = 0)
_
2
+
2
0
where
0
is the volatility of the standard Brownian motion in
Equation 2.2.
Because we assume that 0, the dissipative nature of the model
does not generate deviations of returns that last for extendedperiods
of time (ie, there are no long-term bubbles or busts). However, the
additional term dZ
1
in Equation 2.3 introduces a concept absent
from the standard rational investor framework: the possibility of
extreme price changes caused by positive random feedback result-
ing from investors under- and overreaction to price changes. This
26
feedback mechanism leads to short-lived instabilities (micro bub-
bles and busts) that can exacerbate price changes during crises and
periods of irrational exuberance.
In the limit of perfectly rational market participants ( = 0) and
ideal market efciency ( ), prices follow a random walk
in which price changes have a completely unpredictable compo-
nent. This is the result of rational market participants incorporat-
ing all possible information into prices and quickly eliminating any
prot opportunity (Farmer and Lo 1999). That is, given a small time
interval t > 1, Equation 2.3 yields
t

Z
2
.
This limit corresponds to the adiabatic elimination of fast relaxation
processes, where Equations 2.2 and 2.3 play a role similar to the
stochastic differential equation
dS
S
= dt + dZ
in standard securities pricing models. In simple terms, in the limit
= 0, any errors investors make in forecasting the future are ran-
dom errors rather than the result of stubborn bias towards either
optimism or pessimism (Sobehart and Farengo 2002).
When }= 0, market participants tend to over- and underreact
to price trends, leading to unstable situations that may result in
severe price swings. To illustrate this, let us consider the price adjust-
ment process in Equations 2.2 and 2.3 during a small time interval
t. Although for 0 the stabilising mean-reverting mechanism
for price returns is still present, when t < t + Z
1
mar-
ket participants overreact to price return discrepancies, and when
0 t +Z
1
< t they underreact with respect to the base case
= 0. In contrast, when t +Z
1
< 0, the price adjustment pro-
cess is such that if the actual return is less than the expected return ,
the adjustment may consist of some sellers actually increasing their
offers of the security in anticipation of further price declines (the
bearish view). Also, if the actual return is higher than the expected
return , the adjustment may consist of some buyers increasing their
bids for the security hoping for future price increases (the bullish
view). The result of this randommixture is a market that consistently
fails to perform in the way perfectly rational models predict.
27
MODEL RISK
Due to the natural tendency for investors to weight losses more
heavily than gains (loss aversion), we cannot expect the dynam-
ics of market reaction to price momentum described here with a
single parameter to be valid for arbitrarily large security prices
and excess returns. Also, when prices and returns are bigger than
some reference values (price anchoring), market participants may
show different sensitivity to price trends, which affects the param-
eters required in Equations 2.2 and 2.3 (Ariely 2008). To illustrate
this point, investors over- and underreaction to excess returns and
price trends could be very different when a security, initially priced
at US$20, reaches a US$200 value or drops belowUS$2. First, expec-
tations on the rms further revenue growth or decline could be
limited by market size, competitiveness and other effects. Second,
the investors population, their risk appetite and their investment
goals could differ for different levels of securities prices. For sim-
plicity, here we assume that our model is valid for prices and returns
within sensible value ranges. Above the upper bound price
U
and
return
U
(or belowthe lower price
L
andreturn
L
), the model may
no longer reect the actual price dynamics due to non-linearity in
over- and underreaction to price trends, transaction costs and other
effects. In this extreme regime, the term in Equation 2.3 may in
fact saturate to some critical value (
c
,
c
)
c
. Therefore, the ran-
dom process in Equation 2.3 may become a simple mean-reverting
process with only additive noise for extreme values. This may result
in asymptotically thin (normal) tails for extremely large (or small)
returns and prices that deviate from the fat-tail effects observed for
moderately long price changes described here. This description is
consistent with the sudden decrease in the observed frequency of
extreme returns relative to the frequency of moderate and large
returns reported in Goripkrishnan et al (1999), and also with theoret-
ical models of limited herd behaviour (Eguiluz and Zimmermann
2000).
In the general case }= 0 the stochastic process followedby the log
price andits instantaneous excess return in Equations 2.2 and2.3
can be described in terms of the probability distribution P(, , t) in
the phase space of a random log price
t
in the interval [, +d]
and random excess return
t
in the interval [, + d] at time t.
From Equations 2.2 and 2.3 the differential equation describing the
evolution of the probability distribution of log prices and excess
28
returns is given as follows (Sobehart and Farengo 2002)
P
t
+( +
1
2
2
)
P
[P] +
1
2
2
[(
2
+
2
2
)P] +

2
0
2
2
P
2
(2.4)
DISTRIBUTION OF EXCESS RETURNS
Here we explore a variety of solutions to Equation 2.4 for different
regimes. Notice that, because we assume that the model parameters
,
0
, , and are constant within a practical (moderate) range of
prices and returns, Equation 2.3 does not depend explicitly on the
log price . Therefore, the marginal distribution of excess returns
p(, t) =
_
P(, , t) d
is determined by the solution to the following equation
p
t
=

_
p +
1
2
((
2
+
2
2
)p)
_
(2.5)
The analysis of the stationary distribution of returns implied by
Equation 2.5 provides a rst look at the mechanics of the model
and the origin of the fat-tailed distribution of returns. The stationary
andhomogeneous marginal distribution of the instantaneous excess
return is obtained by setting the time and log price derivatives in
Equation 2.5 to zero
p
0
() =
1
(1 +(/)
2
)
+1
_
A+B
_
0
_
1 +
_
s
_
2
_
ds
_
,
=

, =

2
_
_
(2.6)
Here A and B are normalisation constants and denes the
characteristic scale of returns.
The solution of Equation 2.6 needs to satisfy specic boundary
conditions. Assuming that the excess returns are unbounded (but
still small relative to the threshold values
U
and
L
for observing
non-linear effects), the probability density in Equation 2.6 should
vanish for arbitrarily large returns. That is, p
0
() 0 as .
Imposing this condition together with the restriction that the prob-
ability density must lead to a meaningful cumulative loss distribu-
tion, we obtain B = 0. In this situation the stationary probability
29
MODEL RISK
density p
0
() is a modied t-distribution with = 2+1 degrees of
freedom. Notice that the fat-tailedt-distributiondescribedhere is the
natural consequence of the positive feedbackmechanismintroduced
in Equations 2.2 and 2.3 to model investors over- and underreaction
to price trends.
Furthermore, from Equations 2.5 and 2.6 we can obtain the fol-
lowing approximate time-dependent solution p(, t) for small and
moderate values of normalised excess returns
p(, t)
( +1)
( +
1
2
)
1
_
2
(1 +(/)
2
)
+1
(2.7)
and
2
(t) =
2
+(
2
0
2
)(1 e
2t/
) (2.8)
Here = 1/(+
2
) is the characteristic time for the mean reversion
of returns and
0
is the initial width of the distribution of excess
returns. Equation 2.8 indicates that the approximate distribution of
excess returns rapidly approaches the stationary distribution p
0
()
for t > . Early empirical studies of price changes suggest 5
10 minutes (Goripkrishnan et al 1999). In the following we assume
that t > andneglect the transient effects describedinEquations 2.7
and 2.8. FromEquations 2.7 and 2.8 the standard deviation of excess
returns is simply
_
E(( E())
2
) =
(t)
_
2 1

_
2 1
Notice that as market participants become more rational ( 0) the
distribution is asymptotically normal ( , ) and security
prices behave like ideal randomwalks with Gaussian randomincre-
ments. More precisely, in the limit >
2
, we obtain the following
density of excess returns
p
0
() =
( +1)
( +
1
2
)
1
2
(1 +(/)
2
)
+1

>
2
1
_
2
2
e
2
/2
2
2
=

2
2
_
_
(2.9)
In this limit, the probability of observing an excess return lower than
the value is
F() =
_
p
0
(s) ds
1
_
2
2
_
e
s
2
/2
2
ds =
_

_
(2.10)
30
Here (/) is the cumulative normal probability for observing a
random excess return lower than . Equations 2.9 and 2.10 can be
recognisedas the normal assumptionusedinstandardrational mod-
els based on the efcient market hypothesis, which can be derived
here inthe limit whenmarkets have nofeedbackmechanisms caused
by over- and underreaction to price trends.
In contrast, when
2
the additional variability resulting from
investors following price trends causes asset returns to exhibit a fat-
tailed t-distribution that is asymptotically similar to a power lawfor
an excess return
p
0
() =
( +1)
( +
1
2
)
1
2
(1 +(/)
2
)
+1

>
C
_

_
2(+1)
(2.11)
Here C is the normalisation constant for the truncated power
law. Equation 2.11 indicates that the return distribution exhibits a
decreasing linear pattern in a loglog (power-law) plot of severity
frequency against observed return . More precisely, the probability
of observing an excess return lower than the value is
log[F()] = log(C) +(2 +1)[log() log()] (2.12)
The fat-tail distribution in Equation 2.12, which contrasts with the
normal distribution in Equation 2.10, is consistent with descriptions
of price returns in terms of truncated power-lawdistributions found
in the literature (Goripkrishnan et al 1999; Plerou et al 2001; Gabaix
et al 2003; Novak 2007; Gabaix 2008). Here, however, we present a
more fundamental analysis on the origin of the fat tails based on a
behavioural model of market under- andoverreactiontoprice trends
instead of imposing exogenously a power-law distribution with no
economic interpretation.
The approximate time-dependent solution p(, t) in Equations 2.7
and 2.8 can be normalised as follows
p(z)
( +1)
( +
1
2
)
1
_
(2 1)(1 +z
2
/(2 1))
+1
(2.13)
z(, t) =
_
2 1

(t)
(2.14)
The distribution of normalised excess returns z in Equation 2.13 has
zero mean and unit standard deviation. The reason for introducing
the distributionof normalisedreturns will be clear whenwe describe
the asymptotic marginal distribution of log prices, , later.
31
MODEL RISK
Figure 2.1 Distribution of normalised returns for the S&P 500 Index
10
1
10
0
10
2
10
3
10
4
10
5
10
6
F
r
e
q
u
e
n
c
y
20 15 10 5 0 5 10 15 20
Normalised return Z
Circles denote minute returns for the S&P 500 Index for the period April 2008
to August 2009; the solid line denotes Equation 2.14 with = 1 ( = 3) and
= 0. 1% and the dotted line denotes normal (Gaussian) distribution.
Figure 2.2 Distribution of normalised returns for the DJI Index
10
1
10
0
10
2
10
3
10
4
10
5
10
6
F
r
e
q
u
e
n
c
y
20 15 10 5 0 5 10 15 20
Normalised return Z
Circles denote minute returns for the DJI Index for the period April 2008 to August
2009; the solid line denotes Equation 2.14 with = 1 ( = 3) and = 0. 1% and
the dotted line denotes normal (Gaussian) distribution.
For illustration purposes the stationary and homogeneous distri-
butions of instantaneous excess returns in Equations 2.13 and 2.14
can be compared with the average distribution of high-frequency
(minute-by-minute) returns observed over an extended period of
time (t > and(t) ). The average high-frequencydistribution
is not equivalent to the distribution of instantaneous excess returns,
32
but it provides a rst-order approximation that shows very similar
patterns for extreme returns. Figure 2.1 shows the empirical aver-
age distributions of normalised minute-by-minute returns for the
S&P500 Index for the periodApril 2008 toAugust 2009 (over 127,000
observations). For consistency, only minute returns calculated dur-
ing trading hours after the rst minute of trading are used in this
analysis. Price jumps at the market opening are excluded from the
sample. Figure 2.1 also shows the stationary distribution of instant-
aneous excess returns in Equation 2.13 with 1 ( 3), and a
normal (Gaussian) distribution (standard Brownian motion) tted
to the central peak of the distribution. Notice the reasonable agree-
ment between Equation 2.13 and the empirical distribution over a
wide range of returns. The value of the parameter is also in good
agreement with the values reported in the literature (Goripkrish-
nan et al 1999; Plerou et al 2001; Novak 2007; Gabaix 2008). Other
equity indexes show similar patterns, as shown in Figure 2.2 for the
DowJones Industrial (DJI) Index. Figures 2.1 and 2.2 highlight the
failure of the normal distribution of price returns to capture the cor-
rect frequency of extreme events, which can have consequences for
portfolio risk calculations.
DISTRIBUTION OF RETURNS OVERTIME
Although the distribution of instantaneous excess returns p(z)
exhibits characteristics similar to the observed average empirical
distributions of high-frequency returns, it does not represent the
actual probability distribution for an arbitrary log price , excess
return andtime t. The distributions of log prices andexcess returns
can be derived from the analytical or numerical solution to Equa-
tion 2.4 or, more easily, from stochastic simulations of Equations 2.2
and 2.3. Figures 2.3 and 2.4 illustrate the evolution of the normalised
marginal distributions of returns q(, t) and p(, t) obtained from
500,000 Monte Carlo simulations of Equations 2.2 and 2.3 for = 1,
/ = 0. 3. Notice that the distribution of excess returns in Figure 2.3
is nearly stationary and exhibits a fat tail similar to Equation 2.13. In
contrast, the distribution of log prices in Figure 2.4 diffuses for long
time horizons, while preserving some basic features of the fat-tail
distributionof excess returns. Thedrift-diffusionprocess resembles a
standard Brownian motion, except that the shape of the distribution
reects long-term fat-tail effects.
33
MODEL RISK
Figure 2.3 Evolution of the marginal distribution of excess returns
p(, t) for = 1 and / = 0. 3
10
1
10
0
10
2
10
3
10
4
F
r
e
q
u
e
n
c
y
10 5 0 5 10
Excess return
t = 0.02
t = 0.05
t = 0.07
t = 0.10
t = 0.15
t = 0.20
Notice the convergence to the stationary distribution.
Figure 2.4 Evolution of the marginal distribution of log prices q(, t) for
= 1 and / = 0. 3
10
1
10
1
10
0
10
2
10
3
10
4
F
r
e
q
u
e
n
c
y
3 2 1 0 1 2 3
Log price
t = 0.02
t = 0.05
t = 0.07
t = 0.10
t = 0.15
t = 0.20
Distributions are not normalised to highlight the diffusion process.
Because of the quasi-normal diffusive nature of the log-price
changes for long time horizons, in the following we introduce the
normalised log price x
x(, t) =

0
(
1
2
2
)(t t
0
)
t t
0
(2.15)
34
Figure 2.5 Comparison between the frequency of normalised returns
for the S&P 500 Index and Equation 2.20 for different time horizons for
the period 19502009
10
1
10
0
10
2
10
3
10
4
10
5
10
6
F
r
e
q
u
e
n
c
y
20 15 10 5 0 5 10 15 20
Normalised return Y
Crosses, 19502009 monthly returns; triangles, 19502009 weekly returns;
squares, 19502009 daily returns; circles, 20089 minute returns; solid line, fat-tail
model; dashed line, normal distribution.
Following Sobehart and Farengo (2002, 2003), we can derive a
rst-order approximate solution to the marginal distribution
q(, t) =
_
P(, , t) d
of a log price at time t obtainedfromEquation2.5 using the Fourier
transformation. The approximate marginal distribution is given by
the following expression
q(, t)
e
x
2
/2
_
2
2
(t t
0
)
_
n=0
a
n
(t)H
n
(x)
n!(
t t
0
)
n
(2.16)
Here x is dened in Equation 2.15 and the functions H
n
(x) are the
Hermite polynomials of degree n, which describe corrections to the
normal tails of the distribution. The coefcients a
n
are determined
by additional conditions imposedon the initial distribution of prices
and excess returns and are dened by the following expressions
35
MODEL RISK
for the S&P 500 Index and Equation 2.20 as a function of the reduced
variable w for different time horizons using the same data as in
Figure 2.5
2
0
4
6
8
10
12
L
o
g

f
r
e
q
u
e
n
c
y
0 2 4 6 8 10 12 14
w
model.
(Sobehart and Farengo 2002)
a
n
(t) =
i
n
A(0, t)
n+1
n
A
k
n
(0, t) (2.17)
A(k, t) =
_
m=0
_
+
a
m
H
m
(g(, k))e
g
2
(,k)/2
(1 +(/)
2
)
(+1)/2
d
exp
_
1
2
2
__
(+1)
m
_
(t t
0
)
_
(2.18)
Here we introducedthe variable g(, k) =
c(u+b/c
2
), where / =
sinh(u), b = ik, c =
_
(+1), = +
1
2
and
m
= 2m+1.
Equations 2.152.18 show the corrections to the normal distribu-
tionrequiredfor describingthe distributionof logprices over time.
Also note the decay over time for the different contributions to the
coefcient A(k, t). From Equation 2.18, the characteristic timescale
for the exponential decay of different contributions for m 1 and
>1 is approximately
m

2
2
(
m
1)

1
m
36
for the DJI Index and Equation 2.20 for different time horizons in the
period 19282009
10
1
10
0
10
2
10
3
10
4
10
5
10
6
F
r
e
q
u
e
n
c
y
20 15 10 5 0 5 10 15 20
Normalised return Y
model; dashed line, normal distribution.
The slow time decay for m = 0 in Equation 2.18 (representing the
long-term normal distribution) is caused by the approximations
required to calculate Equation 2.16 and can therefore be neglected.
Equation 2.16 indicates that the bulk of the marginal distribu-
tion of log prices resembles a normal distribution with an effective
volatility determined by the sensitivity of market participants to
price trends (bullishbearish views) and fat-tail contributions
4
(+1)
+
2
0
_
2
__
1 +
1
2
__
1 +
3
2
__
1
+
2
0
(2.19)
Notice that, in the limit 0 ( ), market participants
become rational and the effective volatility approaches the value

_
(/)
2
+
2
0
as discussed in the second section (see page 24).
We should expect the leading normal distribution term in Equa-
tion 2.16 to be valid for a moderate range of securities price changes
conned to the central peak of the distribution, with fat-tail com-
ponents for moderately large returns. For long-term horizons the
37
MODEL RISK
for the DJI Index and Equation 2.20 as a function of the reduced variable
w for different time horizons using the same data as in Figure 2.7
1
3
5
7
9
11
L
o
g

f
r
e
q
u
e
n
c
y
0 2 4 6 8 10 12 14
w
model.
for the FTSE Index and Equation 2.20 as a function of the reduced
variable w for different time horizons in the period 19842009
1
3
5
7
9
11
L
o
g

f
r
e
q
u
e
n
c
y
0 2 4 6 8 10 12 14
w
squares, 19842009 daily returns; solid line, fat-tail model.
38
for the Nikkei Index and Equation 2.20 as a function of the reduced
variable w for different time horizons in the period 19842009
1
3
5
7
9
11
L
o
g

f
r
e
q
u
e
n
c
y
0
2 4 6 8 10 12 14
w
squares, 19842009 daily returns; solid line, fat-tail model.
impact of the fat-tail contributions inEquation2.16 diminishes, lead-
ing to a normal distribution of log prices. Note, however, that, due
to the diffusive nature of the process, the fat-tail characteristics can
be preserved for relatively long periods of time.
Although Equation 2.16 describes the approximate impact of fat-
tail effects due to market under- and overreaction to price trends,
its mathematical complexity limits its usefulness for practical appli-
cations. Even the computation of plain vanilla European options
requires demanding calculations, as shown in Sobehart andFarengo
(2002). Here we introduce simpliedasymptotic expressions that can
be used for some practical applications.
The asymptotic analysis of Equations 2.15 and 2.16 for moderate
values of and t suggests that the distribution of normalised log
prices x can be approximated roughly with a t-distribution similar
to Equations 2.132.14, whose degrees of freedom change slowly
over time approaching a normal distribution
q(y)
( +1)
( +
1
2
)
_
_
(2 1)
_
1 +
y
2
2 1
_
+1
_
1
(2.20)
y(, t) =
_
2 1
x
(2.21)
39
MODEL RISK
Here (t) and (t) 1 are slowly varying functions of time
that approximate asymptotically the shape of the distribution of log
prices for different time horizons. Equations 2.20 and 2.21 provide a
simple asymptotic expressionfor moderate values of the normalised
variable x that preserves the main characteristics of the fat-tailed
distribution of excess returns over long periods of time.
Figure 2.5 shows the distributionof returns for different time hori-
zons for the S&P 500 Index against Equation 2.20 with = = 1
( = 3) and = 1. The same data is presented in Figure 2.6, high-
lighting the linear relationship between the logarithm of the fre-
quency of returns, log(q), and the reduced variable w = log(1 +
y
2
/(21)). Notice the reasonable agreement betweenthe observed
returns and Equation 2.20 from minute-by-minute observations for
the period from April 2008 to August 2009 to monthly observa-
tions for the period 19502009. For each time horizon, the difference
between the empirical frequency and the asymptotic approxima-
tion (Equation 2.20) is most noticeable for large returns, when data
is sparse and sample frequencies can be driven by a small num-
ber of volatile observations. Figures 2.7 and 2.8 show similar results
for the DJI Index for the period 19282009 using the same parame-
ters as for the fat-tail model used in Figure 2.5. Figures 2.9 and 2.10
show similar results for the FTSE 100 (UK) and Nikkei (Japan)
Indexes, reecting similar characteristics across market segments
and geographies.
Of course, the asymptotic approximations 2.13 and 2.20 (or the
more complex Equation 2.16) are valid only within reasonable
bounds of excess returns and log price and time t as we can-
not expect the dynamics of market participants described here with
constant parameters and simple functional forms to be valid for
arbitrarily large security prices and excess returns. This is rein-
forced by the fact that frequent changes in the behaviour of mar-
ket participants and the natural tendency for investors to weight
losses more heavily than gains may result in both time-dependent
and non-linear processes. Prices can also include sudden jumps in
response to changes in information or risk perception beyond the
effects described in our model. However, the results are encourag-
ingandsuggest that models basedonbehavioural patterns drivenby
price momentumandtrends canhelpus to gaina better understand-
ing of the dynamics of securities prices during crises beyond the
40
limitations of traditional models of perfectly rational investors that
require large changes in their parameters to capture the observed
price moves.
REFERENCES
Andersen, J. V., andD. Sornette, 2002, ANonlinear Super-Exponential Rational Model of
Speculative Financial Bubbles International, Journal of Modern Physics C 13(2), pp. 17188.
Antonuccio, F., and M. Proebsting, 2003, A Risk-Neutral Approach to Option Pricing
with Jumps and Diffusion, The Journal of Risk 5(2), pp. 73105.
Ariely, D., 2008, Predictably Irrational (New York: Harper Collins).
Bachelier, L., 1900, Theory of Speculation. English translation in P. Cootner, 1964, The
Random Character of Stock Market Prices, pp. 1778 (Cambridge, MA: MIT Press).
Baker, M., and J. C. Stein, 2002, Market Liquidity as a Sentiment Indicator, Working
Paper, Harvard Business School.
Barndorff-Nielsen, O. E., and N. Sheppard, 2001, Non Gaussian OrsteinUhlenbech-
Based Models and Some of Their Uses in Financial Economics, Journal of the Royal
Statistical Society B 63(2), pp. 167241.
Bibby, B. M., and M. Soresen, 2001, Hyperbolic Processes in Finance, Working Paper
MathPhySto MPS-RR 2001-21, University of Aarhus.
Bookstaber, R., 2007, A Demon of Our Own (New York: John Wiley & Sons).
Bouchaud, J. P., P. Cizeau, L. Leloux and M. Potters, 1999, Mutual Attractions: Physics
and Finance, Physics World, January, pp. 259.
Campbell, J., A. W. Lo and A. C. MackKinlay, 1996, The Econometrics of Financial Markets,
pp. 943, 87147 (Princeton University Press).
Cont, R., andJ. P. Bouchard, 1998, HerdBehavior andAggregate Fluctuations inFinancial
Markets, Working Paper, Centre dEtudes de Saclay.
Damasio, A. R., 1994, Descartes Error: Emotion, Reason and the Human Brain (New York:
Putnam Books).
DeBondt, W. F. M., and R. Thaler, 1985, Does the Stock Market Overreact?, Journal of
Finance, July, pp. 793805.
DeLong, J. B., A. Shleifer, L. Summers and R. Waldeman, 1990, Noise Trader Risk in
Financial Markets, Journal of Political Economy 98, pp. 70338.
Dragulescu, A. A., and V. M. Yakovenko, 2002, Probability Distribution of Returns in
the Heston Model with Stochastic Volatility, Working Paper, Department of Physics,
University of Maryland.
Eguiluz, V. M., and M. G. Zimmermann, 2000, Transmission of Information and Herd
Behavior: An Application to Financial Markets Physics Review Letters 85, pp. 565962.
Fama, E. F., 1970, Efcient Capital Markets: A Review of Theory and Empirical Work,
Journal of Financial Economics, May, pp. 383417.
Fama, E. F., 1998, Market Efciency, Long-TermReturns and Behavioral Finance, Journal
of Financial Economics 49, pp. 283306.
Farmer, J. D., and A. W. Lo, 1999, Frontiers of Finance: Evolution and Efcient Markets,
Report SFI-99-06-039, Santa Fe Institute.
41
MODEL RISK
Gabaix, X., 2008, Power Laws inEconomics andFinance National, WorkingPaper 14299,
National Bureau of Economic Research.
Gabaix, X., P. Gopikrishnan, V. Plerou and H. E. Stanley, 2003, A Theory of Large
Fluctuations in Stock Market Activity, Working Paper, MIT Economics Department.
Goripkrishnan, P., V. Plerou, L. A. Nunez-Amaral, M. Meyer and E. Stanley, 1999, Scal-
ing of the Distributionof Fluctuations of Financial Market Indices, Physical ReviewE60(5),
pp. 530516.
Hirshleifer, D., 2001, Investor Psychology and Asset Pricing, Working Paper, Fisher
College of Business, Ohio State University.
Jones, D., 1991, The Politics of Money: The Fed under Alan Greenspan (New York Institute of
Finance).
Kendall, M., 1953, The Economic Analysis of Time Series. I. Prices, Journal of the Royal
Statistical Society, 13.
Kindleberger, C., 2000, Manias, Panics and Crashes: AHistory Of Financial Crises (NewYork:
John Wiley & Sons).
Mahar, M., 2003, Bull: A History of the Boom 19821999, pp. 331 (New York: Harper
Business).
Malkiel, B. G., 1981, A Random Walk Down Wall Street (New York: Norton).
Metzler, R., and J. Klafter, 2004, The Restaurant at the End of the Random Walk: Recent
Developments in the Description of Anomalous Transport by Fractional Dynamics,
Journal of Physics A37, pp. R161208.
Novak, S. Y., 2007, Measures of Financial Risk and Market Crashes, Theory of Stochastic
Processes 13(29), pp. 18293.
OHara, M., 1998, Market Microstructure Theory (Oxford: Blackwell).
Plerou, V., P. Gopikrishnan, X. Gabaix, L. Nunes Amaral and H. E. Stanley, 2001, Price
Fluctuations, Market Activity and Trading Volume, Quantitative Finance 1(2), pp. 2629.
Samuelson, P. A., 1965, Proof That Properly Anticipated Prices Fluctuate Randomly,
Industrial Management Review 6, pp. 419.
Shefrin, H., 1999, Irrational Exuberance and Option Smiles, Financial Analyst Journal,
November/December, pp. 91103.
Shiller, R. J., 2000, Irrational Exuberance, pp. 13568 (Princeton University Press).
Shleifer, A., 2000, Inefcient Markets: An Introduction to Behavioral Finance, pp. 152, 11274
(Oxford University Press).
Sobehart, J. R., 2003, A Mathematical Model of Irrational Exuberance and Market
Gloom, GARP Risk Review, July/August, pp. 226.
Sobehart, J. R., 2005, A Forward Looking, Singular Perturbation Approach to Pricing
Options under Market Uncertainty and Trading Noise, International Journal of Theoretical
and Applied Finance 8(8), pp. 63558.
Sobehart, J. R., and R. Farengo, 2002, Fat Tailed Bulls and Bears, Risk, December, S204.
Sobehart, J. R., and R. Farengo, 2003, ADynamical Model of Market Under- and Over-
Reaction, The Journal of Risk, 6(4), pp. 91116.
Sobehart, J. R., and S. C. Keenan, 2005, Capital Structure Arbitrage and Market Timing
under Uncertainty and Trading Noise, The Journal of Credit Risk 1(4), pp. 129.
42
Sornette, D., Y. Malevergne andJ. F. Muzy, 2003, What Causes Crashes?, Risk, February,
pp. 6771.
Thaler, R. H., 1999, The Endof Behavioral Finance, Financial Analyst Journal 55(6), pp. 12
17.
Varian, H. R., 1985, Differences of Opinion in Financial Markets, Working Paper,
Department of Economics, University of Michigan.
43
3
Model Risk and Non-Gaussian
Latent Risk Factors
Stef Hse and Stefan Huschens
Technische Universitt Dresden
Risk-factor models withlatent risk factors forma widely usedmodel
class in market and credit-risk analysis. Since latent risk factors are
not directly observable, it is usually assumed that their stochas-
tic behaviour can be modelled by Gaussian distributions. The role
of this assumption, which seems to be arbitrary, and the model
risk arising from it will be discussed in this chapter using credit-
portfolio models based on risk factors, which are standard in the
banking industry. In these models, the systematic risk factor and
all idiosyncratic risk factors are generally taken to be Gaussian dis-
tributed (Basel Committee on Banking Supervision 2006, pp. 634).
The purpose of this assumption is a parsimonious modelling of the
dependence structure of default events so that only one parameter
or a small number of parameters determine the loss distribution and
especially its right tail. The crucial question is What are the effects
of this assumption on the resulting credit-portfolio loss distribution
and on the corresponding risk measures quantifying the credit risk
of the portfolio? or, more generally, What is the model risk arising
from this assumption?
The remainder of the chapter is structuredas follows. Since deni-
tions and measures of model risk differ among authors, a review of
theexistingliteratureis giveninthenext section. Inthefollowingsec-
tion, a basic credit-risk model with Gaussian distributed latent risk
factors, the well-known single-risk-factor model and a special case
of it with homogeneous potential losses, default probabilities and
correlation parameters are introduced. The distributional assump-
tions for the latent risk factors are then reduced, and generalisations
of both models are presented. Next, the model risk arising from
45
MODEL RISK
Gaussian latent risk factors is quantied by means of risk measures
such as the value-at-risk and the average value-at-risk. The chapter
concludes with a summary.
DEFINING AND MEASURING MODEL RISK
A basic requirement for the quantication of model risk is its def-
inition. Since the denition of the term model risk differs among
authors, a review of the existing literature is given as a rst step.
Measures of model risk are introduced in a second step.
Denitions of model risk
Derman (1996, pp. 356) was one of the rst to analyse and dene
the term model risk in the eld of nancial engineering. He dis-
tinguished between seven types of model risk: inapplicability of
modelling; incorrect modelling due to incorrect model assump-
tions; correct model but incorrect analytical solution; correct model
but inappropriate use; badly approximated solution due to numer-
ical problems; software and hardware bugs and unstable data
(parameter uncertainty). Crouhy et al (1999, pp. 2739) present a
similar typology and incorporate everything in their denition of
model risk that might be related to the statistical model used.
This includes model misspecication, data contamination, incorrect
implementation, calibration and application.
In contrast to this application-oriented point of view, a rather
research-oriented approach is presented by the following authors.
Meyer zu Selhausen (2004, pp. 2746) used the fact that credit-
portfolio models are prediction models and proposed the concepts
of rst- and second-order model risk. The rst-order model risk is
dened as the prediction error, ie, the difference between the model-
based prediction of the portfolio loss and its real value. The risk that
the distribution of the prediction error is not estimable is called the
second-order model risk. Credit-portfolio models suffer from both
types of model risk due to the limited data availability and the fact
that model parameters such as default probabilities and correlations
are not directly observable. Kerkhof et al (2010, pp. 268, 2723) inter-
preted model risk as the hazard of working with a potentially incor-
rect model. More precisely, a decomposition of model risk into three
components is proposed: the estimation risk which represents the
part of the model risk that is caused by estimation errors in a given
46
MODEL RISK AND NON-GAUSSIAN LATENT RISK FACTORS
class of parametric models; the misspecicationrisk, whichis caused
by the possibility of an incorrect choice of model; the identication
risk, which arises when observationally indistinguishable models
lead to different results, eg, to different values of the risk measure of
interest. In the approach of Sibbertsen et al (2008, pp. 6871), model
risk can occur at three levels of the statistical modelling procedure.
At the rst level, the model riskis inducedbythe probabilityspace of
the model. At the second level, it is caused by the uncertainty about
the specication of the statistical model (misspecication risk). At
the third level, model risk is caused by the uncertainty about the
model parameters (estimation risk).
In this chapter, only one aspect of model risk is addressed: the
misspecication risk associated with the assumption of Gaussian
distributed latent risk factors in a credit-portfolio model based on
risk factors. The following measures can be used to quantify this
kind of model risk.
Measures of model risk
The credit risk of a portfolio is often measured by a single number
such as the value-at-risk or the average value-at-risk. For these risk
measures, the concept of a model risk set is introduced as the region
of uncertainty resulting from the misspecication risk associated
with the distributional assumptions of the credit-portfolio model
used.
In order to dene such sets, the following notation will be used
throughout this chapter: F denotes the set of all distribution func-
tions and X F means that the random variable X has distribution
function F F, ie, X F Pr(X x) = F(x) for all x R.
Denition 3.1. Let Mbe a non-empty subset of the set F of all distri-
bution functions and let m[] be a real-valued risk measure dened
for all random variables X with a distribution function in M.
(i) The set
?:= m[X] X F, F M (3.1)
is called model risk set for the measure m[] with respect to
the set of distribution functions M.
(ii) The supremum of the model risk set is called the worst-case
risk. The inmum of the model risk set is called the best-case
risk.
47
MODEL RISK
(iii) A distribution with distribution function F M is called a
worst-case distribution or a best-case distribution, if X F
and if
m[X] = max ? or m[X] = min ? (3.2)
respectively.
Latent systematic and idiosyncratic risk factors in credit-portfolio
models are oftenassumedtofollowthe standardnormal distribution
with distribution function . This strong distributional assumption
canbe weakenedto the assumptionof anarbitrary distributionfunc-
tion F belonging to a set M of distribution functions with M,
eg, the set which contains all distributions with mean 0, variance 1
and continuous distribution functions. Then the model risk set for
a measure m[] quanties the misspecication risk associated with
the assumption of Gaussian distributed latent risk factors. The aim
of this chapter is to dene model risk sets for the value-at-risk and
the average value-at-risk in a generalised single risk factor model
and to nd the related worst-case risks and worst-case distributions
with respect to the set of standardiseddistributions with mean 0 and
variance 1.
GAUSSIAN LATENT RISK FACTORS
Before model risks can be quantied, the underlying model must
be dened. For this purpose, credit-risk models with Gaussian dis-
tributed latent risk factors are presented in this section. The well-
knownsingle-risk-factor model anda special case of it withhomoge-
neous potential losses, default probabilities and correlation param-
eters are introduced. Additionally, their properties are discussed
and the corresponding risk measures value-at-risk and average
value-at-risk are dened.
A basic single-risk-factor model
The default behaviour of each obligor i is modelled by a Bernoulli
distributed default variable D
i
, where D
i
= 1 indicates the default
of obligor i. The potential loss caused by this obligor in case of its
default is denoted by v
i
. Without loss of generality, it is assumed that
the potential losses are positive and normalised. Thus, the loss of a
48
portfolio with n obligors can be modelled by the random variable
V
n
:=
n
_
i=1
v
i
D
i
with v
i
> 0, i = 1, . . . , n and
n
_
i=1
v
i
= 1 (3.3)
The dependence structure of the default variables is modelled by
D
i
:= 1
_
i
Z +
_
1
i
Y
i

1
(p
i
)
0 < p
i
< 1, 0 <
i
< 1, i = 1, . . . , n (3.4)
where Z is the latent systematic risk factor that simultaneously
affects the creditworthiness of all obligors, and where Y
1
, . . . , Y
n
are
the latent idiosyncratic risk factors (Bluhmet al 2003, p. 84; Huschens
and Vogl 2002, p. 287). Here, 1 denotes the indicator function and
1
() denotes the inverse of the distribution function of the stan-
dard normal distribution. In order to specify the loss distribution for
this model, an assumption about the joint distribution of all latent
risk factors has to be made.
Assumption 3.2. The random variables Z, Y
1
, . . . , Y
n
are indepen-
dent and follow the standard normal distribution.
Denition 3.3. The single-risk-factor (SRF) model is dened by the
two model equations 3.3 and 3.4 and by Assumption 3.2.
The random variables
A
i
:=
_
i
Z +
_
1
i
Y
i
, i = 1, . . . , n (3.5)
can be seen as the returns of the stochastic processes of the credit-
worthiness of the different obligors. Assumption 3.2 implies that the
random variables A
i
are identically distributed, follow the standard
normal distribution and are correlated with (Huschens and Vogl
2002, p. 287)
Corr[A
i
, A
j
] =
_
j
, i j, i, j = 1, . . . , n
In the next lemma, known facts about the SRF model are sum-
marised.
Lemma 3.4. Suppose the conditions of the SRF model given in
Denition 3.3 hold.
(i) The default variables D
i
are Bernoulli distributed with default
probabilities p
i
D
i
Ber(p
i
), i = 1, . . . , n
49
MODEL RISK
(ii) Given a realisation z of the systematic risk factor Z, the
default variables are stochastically independent and Bernoulli
distributed
D
i
Z = z Ber(p
i
(z))
with conditional default probabilities
Pr(D
i
= 1 Z = z) = p
i
(z)
:=
_
1
(p
i
)
i
z
_
1
i
_
, i = 1, . . . , n
(iii) The unconditional distributionof the default variables is given
by
Pr(D
1
= d
1
, . . . , D
n
= d
n
) =
_
i=1
Pr(D
i
= d
i
Z = z) d(z)
with d
i
0, 1 for i = 1, . . . , n.
The proof of this lemma and the proofs of the following lem-
mas and theorems can be found in the appendix. As can be seen
in Equation 3.3, the distribution function F
V
n
() := Pr(V
n
) of
the loss variable V
n
depends on the unconditional n-dimensional
distribution of the default variables given in the lemma above. In
order to quantify the credit risk arising from this distribution, the
risk measures value-at-risk and average value-at-risk will be used.
For a given probability level 0 < < 1, the value-at-risk of the loss
distribution is dened by
VaR
[V
n
] := minx F
V
n
(x) (3.6)
and the average value-at-risk (also called expected shortfall or
tail value-at-risk) is dened by (McNeil et al 2005, pp. 38, 44)
AVaR
[V
n
] :=
1
1
_
1
VaR
u
[V
n
] du (3.7)
Homogeneous single-risk-factor model
In order to derive a model for which the asymptotic loss distribution
canbe givenexplicitly, the SRFmodel is combinedwiththe following
assumption of homogeneity.
Assumption 3.5.
(i) Homogeneous potential losses: v
i
= 1/n for i = 1, . . . , n.
50
(ii) Homogeneous default probabilities: p
i
= for i = 1, . . . , n
with 0 < < 1.
(iii) Homogeneous correlation parameters:
i
= for i = 1, . . . , n
with 0 < < 1.
Denition 3.6. The homogeneous single-risk-factor (HSRF) model
is dened by the two model equations (Equations 3.3 and 3.4) and
by Assumptions 3.2 and 3.5.
The assumed homogeneity of the potential losses implies that the
portfolio loss V
n
coincides with the stochastic default rate, ie
V
n
=
1
n
n
_
i=1
D
i
(3.8)
Furthermore, the assumed homogeneity of the default probabili-
ties and correlation parameters implies that the default variables
D
1
, . . . , D
n
are identically Bernoulli distributed with default proba-
bility , and that the random variables A
1
, . . . , A
n
are identically
Gaussian distributed with equicorrelation (Schnbucher 2003,
pp. 305ff). In this case, the portfolio loss distribution depends only
on three parameters: the number of obligors n; the default probabil-
ity 0 < < 1; and the correlation 0 < < 1. In the next lemma,
known facts about the HSRF model are summarised.
Lemma 3.7. Suppose the conditions of the HSRF model given in
Denition 3.6 hold.
(i) The conditional default probabilities are given by
Pr(D
i
= 1 Z = z) = p(z)
:=
_
1
()
z
_
1
_
, i = 1, . . . , n
(ii) The portfolio loss V
n
given in Equation 3.8 has the discrete
probability distribution
Pr
_
V
n
=
k
n
_
=
_
_
n
k
_
p(z)
k
(1 p(z))
nk
d(z), k = 0, 1, . . . , n
51
MODEL RISK
(iii) The portfolio loss V
n
converges for n in quadratic mean
and with probability 1 (almost surely) to the random variable
p(Z) = p() Z with Z (3.9)
The distribution of p(Z) is called Vasicek distribution with
parameters and .
(iv) The -quantile of p(Z) is given by
p(
1
(1 )) = p(
1
()), 0 < < 1
Since the convergence in quadratic mean implies the convergence
in distribution, the portfolio loss V
n
converges for n in distribu-
tion to the Vasicek distributed random variable p(Z). For portfolios
with a large number of obligors, this fact justies the approximation
of the loss distributionbya Vasicekdistribution. Inaddition, the con-
vergence in distribution is equivalent to the convergence in quantile
(Shorack 2000, p. 112). For a given probability level 0 < < 1,
this allows the approximation of the value-at-risk VaR
[V
n
] and
the average value-at-risk AVaR
[V
n
] for a portfolio with a nite but
large number of obligors by
VaR
[p(Z)] = p(
1
()) =
_
1
() +
1
()
_
1
_
(3.10)
and by
AVaR
[p(Z)] =
1
1
_
1
VaR
u
[p(Z)] du (3.11)
respectively.
NON-GAUSSIAN LATENT RISK FACTORS
Before the misspecication risk associated with the assumption of
Gaussian distributed latent risk factors in the SRF and the HSRF
model can be quantied, generalisations of these models will be
presented. More precisely, the distributional assumptions for the
latent risk factors in the SRF and the HSRF model will be weakened
as follows.
52
A generalised single-risk-factor model
In this section, an SRF model with non-Gaussian latent risk factors
is introduced. For this purpose, Assumption 3.2 is replaced by the
following assumption, where the set of distribution functions
F
0,1
:=
_
F
F is distribution function,
_
R
x dF(x) = 0,
_
R
x
2
dF(x) = 1
_
characterises all standardised distributions. If a random variable X
has distribution function F
X
F
0,1
, then E[X] = 0 and V[X] = 1, ie,
the random variable X is standardised.
Assumption 3.8.
(i) The randomvariables Z, Y
1
, . . . , Y
n
are stochastically indepen-
dent.
(ii) The random variable Z is distributed with distribution func-
tion F
Z
F
0,1
.
(iii) The random variables Y
1
, . . . , Y
n
are identically distributed
with distribution function F
Y
F
0,1
.
It should be noted that Assumption 3.8 is a necessary but not suf-
cient condition for Assumption 3.2. The implications of Assump-
tion 3.8 for the distribution of the random variables A
1
, . . . , A
n
are
presented in the next lemma.
Lemma 3.9. Let Assumption 3.8 hold and let A
i
be dened by
Equation 3.5.
(i) The distributionfunctions of the randomvariables A
i
are given
by
F
A
i
() = F
Z
_

i
_
F
Y
_

_
1
i
_
, i = 1, . . . , n (3.12)
where denotes the convolution of distribution functions.
(ii) The random variables A
i
are standardised, ie, F
A
i
F
0,1
for
i = 1, . . . , n.
(iii) If
i
=
j
, then the random variables A
i
and A
j
are identically
distributed.
53
MODEL RISK
Since the random variables A
i
are not necessarily Gaussian dis-
tributed, the denition of the default variables given in Equation 3.4
has to be replaced by
D
i
:= 1
_
_
i
Z +
_
1
i
Y
i
F
A
i
(p
i
)
_
0 < p
i
< 1, 0 <
i
< 1, i = 1, . . . , n (3.13)
where
F
X
(u) := minx F
X
(x) u, 0 < u < 1 (3.14)
denotes the lower quantile function of a random variable X. In gen-
eral, the default variables D
i
are Bernoulli distributed with default
probability F
A
i
(F
A
i
(p
i
)) p
i
. Since the random variables A
i
do not
necessarily have a continuous and strictly increasing distribution
function under Assumption 3.8, it is additionally assumed that
F
A
i
(F
A
i
(p
i
)) = p
i
, i = 1, . . . , n (3.15)
to ensure that obligor i has the default probability p
i
. This means
that p
i
is in the range of function F
A
i
. For example, if F
A
i
= , then
the generalised inverse F
A
i
=
dened by Equation 3.14 coincides

with the common inverse
1
so that (
1
(p)) = p for all 0 < p < 1
and therefore condition 3.15 is fullled.
Denition 3.10. The generalised single-risk-factor (GSRF) model is
dened by the three model equations (Equations 3.3, 3.13 and 3.15)
and by Assumption 3.8.
The implications of the GSRF model are summarised in the next
theorem.
Theorem 3.11. Suppose the conditions of the GSRF model given in
Denition 3.10 hold.
(i) The default variables are Bernoulli distributed, D
i
Ber(p
i
)
for i = 1, . . . , n.
(ii) The conditional default probabilities are given by
Pr(D
i
= 1 Z = z) = p
i
(z; F
Z
, F
Y
) := F
Y
_
F
A
i
(p
i
)
i
z
_
1
i
_
for all elements z in a set
Z
R with Pr(Z
Z
) = 1.
54
(iii) The unconditional distributionof the default variables is given
by
Pr(D
1
= d
1
, . . . , D
n
= d
n
) =
_
Z
n
i=1
Pr(D
i
= d
i
Z = z) dF
Z
(z)
with d
i
0, 1 for i = 1, . . . , n.
Generalised homogeneous single-risk-factor model
If the GSRF model with xed distribution functions F
Z
and F
Y
for
the systematic and the specic risk factors is combined with the
homogeneity assumption, a model results for which the asymptotic
loss distribution is derived.
Denition 3.12. The generalised homogeneous single-risk-factor
(GHSRF) model is dened by the three model equations 3.3, 3.13
and 3.15 and by Assumptions 3.5 and 3.8.
The standard HSRF model of the previous section (see page 50)
is a special case of the GHSRF model for F
Z
= F
Y
= . In both
of these models, the portfolio loss V
n
coincides with the stochastic
default rate, as can be seen in Equation 3.8. Further implications of
the GHSRF model are given in the next theorem.
Theorem 3.13. Suppose the conditions of the GHSRF model given
in Denition 3.12 hold.
(i) The random variables A
1
, . . . , A
n
are identically distributed
with distribution function
F
A
() = F
Z
_

_
F
Y
_

_
1
_
F
0,1
(ii) The default variables are identically Bernoulli distributed,
D
i
Ber() for i = 1, . . . , n.
(iii) The conditional default probabilities are given by
Pr(D
i
= 1 Z = z) = p(z; F
Z
, F
Y
) := F
Y
_
F
A
()
z
_
1
_
for all elements z in a set
Z
R with Pr(Z
Z
) = 1.
55
MODEL RISK
(iv) The portfolio loss V
n
given in Equation 3.8 has the probability
distribution
Pr
_
V
n
=
k
n
_
=
_
Z
_
n
k
_
p(z; F
Z
, F
Y
)
k
(1 p(z; F
Z
, F
Y
))
nk
dF
Z
(z) (3.16)
for k = 0, 1, . . . , n with 0
0
:= 1.
(v) The portfolio loss V
n
converges for n in quadratic mean
and with probability 1 (almost surely) to the random variable
V
:= p(Z; F
Z
, F
Y
) = p(; F
Z
, F
Y
) Z with Z F
Z
(3.17)
Denition3.14. The distributionof the randomvariable V
dened
in Equation 3.17 is called generalised Vasicek distribution with
parameters and and risk factor distributions F
Z
and F
Y
.
Theconvergenceinquadratic meaninTheorem3.13(v) implies the
convergenceindistributionsothat theportfolioloss V
n
converges for
n in distribution to the random variable V
with distribution
function F
V
. For portfolios with a large number of obligors, this fact

justies the approximation of the loss distribution by a generalised
Vasicekdistributionandtherefore the approximationof the value-at-
risk VaR
[V
n
] and the average value-at-risk AVaR
[V
n
] (compare
Equations 3.6 and 3.7) by
VaR
[V
] = minx F
V
(x) (3.18)
and
AVaR
[V
] =
1
1
_
1
VaR
u
[V
] du (3.19)
respectively.
In the special case F
Z
= F
Y
= , it holds that p(Z; F
Z
, F
Y
) =
p(Z; , ) = p(Z) with p(Z) given in Equation 3.9. Then the ran-
dom variable V
given in Equation 3.17 follows a common Vasicek

distribution with parameters and . The value-at-risk VaR
[V
]
in Equation 3.18 specialises to the value-at-risk in Equation 3.10 and
the average value-at-risk AVaR
[V
] in Equation 3.19 specialises to

the average value-at-risk in Equation 3.11.
The following proposals for distribution functions F
Z
F
0,1
and
F
Y
F
0,1
are made in the literature as alternative choices to F
0,1
.
56
Standardised t-distribution. If random variable X
follows
a t-distribution with (degrees of freedom) parameter
3, 4, . . . , thenE[X
] = 0 andV[X
] = /(2). The random

variable
:=
, 3, 4, . . .
has a standardised t-distribution with parameter and dis-
tribution function F
(x) := Pr(
x), where F
F
0,1
(cf
Wang et al (2009, p. 272) and the double t copula in Burtschell
et al (2009, p. 14)).
Standardised generalised t-distribution with non-integer
parameter. If the stochastically independent random vari-
ables U and V follow a standard normal distribution and a
chi-squared distribution with non-integer parameter > 0,
respectively, then the random variable
X
:=
U
_
V/
, > 0
follows a generalised t-distribution with non-integer parame-
ter (also called fractional degrees of freedom). The random
variable
:=
, > 2
has a standardised generalised t-distribution with non-integer
parameter and distribution function F
(x) := Pr(X
x),
where F
F
0,1
(Wang et al 2009, pp. 2801).
Mixture of two Gaussian distributions. Let X be a random
variable with distribution function
F(x) = c
_
x
1
_
+(1 c)
_
x
2
_
, 0 < c < 1
Then the distribution of X is a mixture of the two Gaussian
distributions N(0,
2
1
) and N(0,
2
2
). Aparameterisation with
c
2
1
+ (1 c)
2
2
= 1 ensures that F F
0,1
(Wang et al 2009,
pp. 27980).
Mixture of more than two Gaussian distributions. If X is a
random variable with distribution function
F(x) =
_
_
x
_
dH()
57
MODEL RISK
where the mixing distribution function H has the two proper-
ties
H(0) = 0 and
_
dH() = 1
then X has the distribution of a mixture of Gaussian distribu-
tions N(0, ) and F F
0,1
. Apossible choice of H is the inverse
Gaussian distribution suitably parameterised(Wang et al 2009,
pp. 2723).
Mixture of a standardised generalised t-distribution with
non-integer parameter andthe standardnormal distribution.
Let F
denote the distribution function of a standardised gen-

eralisedt-distributionwithnon-integer parameter > 2. Then
the distribution function
F(x) = cF
(x) +(1 c)(x), 0 < c < 1

characterises a mixture of a standardised generalised t-dis-
tribution with non-integer parameter > 2 and a standard
normal distribution. It holds that F F
0,1
(Wang et al 2009,
p. 281).
Standardised smoothly truncated stable distribution. The
smoothly truncatedstable distribution

S
(, , ) is derived
froma commonstable distributionS
(, , ) sothat the dis-

tributionfunctionof

S
(, , ) coincides withthe distribution

function of S
(, , ) in an interval [a, b]. It further coincides

with the Gaussian distribution function ((x
1
)/
1
) for
x < a and with ((x
2
)/
2
) for x > b. The four parameters
, , and determine uniquely the cut-off points a and b
and the four parameters
1
,
1
,
2
and
2
. Whereas stable
distributions have nite variances only if = 2, the smoothly
truncated stable distributions

S
(, , ) have nite means

andvariances for all parameters , , andandcantherefore
be standardised (Wang et al 2009, pp. 2823).
The concept of t-distributed risk factors combined with a t copula
(Bluhm et al 2003, p. 111; Burtschell et al 2009, p. 14; Hamerle and
Rsch 2005, p. 43) does not t in this context since the systematic
and idiosyncratic risk factors are not independent as required by
Assumption 3.8, even if the correlation parameter is zero.
If the risk factors are observable, a choice between the competing
distributions discussed above is possible with statistical methods
58
based on observed risk factor values. However, in the context of the
latent risk factors considered here, it is more reasonable to specify
subsets of permitted distributions than to replace a single distri-
bution by another. Using these subsets of distributions, model risk
sets in the sense of Equation 3.1 are determined and worst-case and
best-case distributions in the sense of Equation 3.2 may be derived.
RISK MEASURES AND MODEL RISK
In this section we analyse the implications for the distribution of
the portfolio loss and for the associated risk measures, ie, the value-
at-risk and the average value-at-risk, if the SRF model is replaced
by the GSRF model with non-Gaussian latent risk factors Z F
Z
and Y F
Y
. For this purpose, the distribution functions F
Z
and F
Y
are only restricted by sets F
Z
F
0,1
and F
Y
F
0,1
, where generally
F
Z
F
Y
is possible. In this chapter, the special case F
Z
= F
Y
= F
0,1
is analysed, whereas further restrictions of the sets are also possible
(eg, continuous distributionfunctions, symmetric distributions, uni-
modal distributions or distributions with nite moments of a certain
order).
VaR, AVaR and model risk sets
The model risk, or, more precisely, the misspecication risk associ-
ated with the assumption of Gaussian distributed latent risk factors,
can be quantied by means of model risk sets as dened by Equa-
tion 3.1. For this reason, model risk sets for the value-at-risk and
the average value-at-risk of a credit portfolio with respect to the
risk-factor distribution functions F
Z
and F
Y
, each restricted only by
the set F
0,1
, are given in the next three problems. These problems
seem not to have been considered in the literature so far, either for
homogeneous portfolios with a small number of obligors or for the
asymptotic portfolio loss V
in the GHSRF model.

Problem 3.1 (GSRF model). Let 0 < < 1 be a given probability
level. Suppose the conditions of the GSRF model given in Deni-
tion 3.10 hold. Then the distribution of the loss variable V
n
depends
onthevectors v := (v
1
, . . . , v
n
), p := (p
1
, . . . , p
n
) and := (
1
, . . . ,
n
)
of potential losses, default probabilities and correlations.
Which properties do the model risk sets
`
,v,p,
:= VaR
[V
n
] F
Z
F
0,1
, F
Y
F
0,1
59
MODEL RISK
and
,v,p,
:= AVaR
[V
n
] F
Z
F
0,1
, F
Y
F
0,1
have for the value-at-risk and the average value-at-risk with respect
to F
Z
F
0,1
and F
Y
F
0,1
? More particularly, what are the worst-
case value-at-risk sup`
,v,p,
, the best-case value-at-risk inf `
,v,p,
,
the worst-case average value-at-risk sup
,v,p,
and the best-case
average value-at-risk inf
,v,p,
?
Problem 3.2 (GHSRF model: nite case). Let 0 < < 1 be a given
probability level. Suppose the conditions of the GHSRF model given
in Denition 3.12 hold. Then the distribution of the loss variable V
n
depends on the number of obligors n, on the default probability
and on the correlation .
`
,n,,
:= VaR
[V
n
] F
Z
F
0,1
, F
Y
F
0,1
and
,n,,
:= AVaR
[V
n
] F
Z
F
0,1
, F
Y
F
0,1
have? More particularly, what are the worst-case value-at-risk

sup`
,n,,
,n,,
, the worst-case
average value-at-risk sup
,n,,
and the best-case average value-
at-risk inf
,n,,
?
Problem 3.3 (GHSRF model: asymptotic loss distribution). Let
0 < < 1 be a given probability level. Suppose the conditions of
the GHSRF model given in Denition 3.12 hold and let the random
variable V
be dened by Equation 3.17. Then the distribution of

V
depends on the default probability and on the correlation .

`
,,,
:= VaR
[V
] F
Z
F
0,1
, F
Y
F
0,1
and
,,,
:= AVaR
[V
] F
Z
F
0,1
, F
Y
F
0,1
have? More particularly, what are the worst-case value-at-risk

sup`
,,,
,,,
, the worst-
case average value-at-risk sup
,,,
and the best-case average
value-at-risk inf
,,,
?
60
The special case F
Z
= F
Y
= F
0,1
shows that the six sets
dened in Problems 3.13.3 are not empty. We conjecture, but are
not able to prove in full generality, that the six sets are convex, ie, are
intervals. The derivation of the worst-case and best-case risks men-
tioned above can be based on the extremal risk-factor distributions
presented in the following subsection.
Extremal risk-factor distributions
Aprerequisite for nding extremal risk-factor distributions is given
in the next lemma, where bounds for the lower and upper quantiles
for all standardised random variables are stated.
Lemma 3.15. Let
0 < < 1, a
:=
and b
:=

1
Then the inequalities
a
() F
+
() b
(3.20)
hold for each distribution function F F
0,1
with lower quantile
function F
(cf Equation 3.14) and upper quantile function

F
+
(u) := infx F(x) > u, 0 < u < 1
These lower and upper bounds for the lower and upper quantiles
can be linked to the class of dichotomous distributions with mean 0
and variance 1. Dichotomous distributions concentrate their proba-
bility mass on exactly two different points so that the corresponding
distribution functions are step functions with two steps.
Lemma 3.16. Let a
and b
be dened as in Lemma 3.15.

(i) All dichotomous distributions with distribution functions in
F
0,1
are given by
Pr(X = a
) = and Pr(X = b
) = 1 (3.21)
with 0 < < 1.
(ii) For a xed 0 < < 1, the distribution function F
X
(x) :=
Pr(X x) of the random variable X dened in Equation 3.21
is
F
X
(x) = G
(x) :=
_
_
0 if x < a
if a
x < b
1 if b
x
(3.22)
61
MODEL RISK
The lower quantile function of X is
F
X
() =
_
_
_
a
if 0 <
b
if < < 1
(3.23)
and the upper quantile function of X is
F
+
X
() =
_
_
_
a
if 0 < <
b
if < 1
(3.24)
In particular, the lower and upper -quantiles of X are given
by
F
X
() = a
and F
+
X
() = b
(3.25)
The class of dichotomous distributions withmean0andvariance 1
can be characterised by G := G
0 < < 1 F
0,1
with distri-
bution function G
dened in Equation 3.22. The subset G species

discrete distributions which are extremal distributions in the set F
0,1
in the following sense.
Theorem3.17. Let 0 < < 1 be a xed probability level and let X be
a random variable with distribution function G
dened in Equa-
tion 3.22 and with lower -quantile F
X
() and upper -quantile
F
+
X
() given in Equation 3.25. Then the distribution of X minimises
the lower -quantile in the class F
0,1
, ie
minF
() F F
0,1
= F
X
()
and maximises the upper -quantile in the class F
0,1
, ie
maxF
+
() F F
0,1
= F
+
X
()
These properties qualify the class of dichotomous distributions
with mean 0 and variance 1 (characterised by the set G) in the search
for worst-case and best-case risks, eg, the worst-case value-at-risk
and the worst-case average value-at-risk. It should be noted that
these distributions are used only to derive bounds for extreme sce-
narios and are not used as substitutes for the true but unknown
distributions of the risk factors or as benchmark scenarios.
Worst-case VaR and AVaR
In this subsection, Problems 3.2 and3.3 are addressed. In the GHSRF
model, the value-at-risk VaR
[V
n
] and the average value-at-risk
62
AVaR
[V
n
] denedinEquations 3.6 and3.7 canbe approximatedby
VaR
[V
] andAVaR
[V
] giveninEquations 3.18 and3.19, respec-

tively. Since the aim is to nd the corresponding worst-case scenar-
ios of these risk measures, the quantile maximising and minimis-
ing properties given in Theorem 3.17 for the class of dichotomous
distributions with mean 0 and variance 1 are used.
Lemma 3.18. Suppose the conditions of the GHSRF model given in
Denition 3.12 hold. Let = 1
1 and F
Z
= F
Y
= G
, where
G
is dened by Equation 3.22.

(i) If
1 (1 (1
1 )
n
) < < 1, then
VaR
[V
n
] = AVaR
[V
n
] = 1 (3.26)
(ii) If
1 < < 1, then

VaR
[V
] = AVaR
[V
] = 1 (3.27)
where the random variable V
is dened in Equation 3.17.

Lemma 3.18 holds for relevant parameter values of and , eg,
for probability level = 99. 5%, default probability = 1%, any
positive correlation less than 1 and risk-factor distribution func-
tions F
Z
= F
Y
= G
with = 1
1 . Since the random vari-

ables V
n
and V
are bounded by the unit interval, the risk measures

VaR
[V
n
], AVaR
[V
n
], VaR
[V
] and AVaR
[V
] have the upper

bound 1. Lemma 3.18 shows that this upper bound is taken for some
parameter constellations. Animmediate consequence of Lemma 3.18
and of G F
0,1
is the next theorem, which gives a partial answer to
the questions of Problems 3.2 and 3.3.
Theorem 3.19. Suppose the conditions of the GHSRF model given
in Denition 3.12 hold.
(i) If

1 (1 (1
1 )
n
) < < 1, then the worst-case
value-at-risk and the worst-case average value-at-risk of the
portfolio loss V
n
reach the maximum value 1, ie
sup`
,n,,
= max `
,n,,
= 1
sup
,n,,
= max
,n,,
= 1
_
_
_
(3.28)
63
MODEL RISK
The maximums in Equation 3.28 are reached by the worst-case
portfolio loss distribution
Pr
_
V
n
=
k
n
_
=
_
_
_
n
k
_
k
(1 )
nk
(1 ) if k = 0, 1, . . . , n 1
+
n
(1 ) if k = n
(3.29)
with = 1
1 .
(ii) If
1 < < 1, then the worst-case value-at-risk and the

worst-case average value-at-risk of the asymptotic portfolio
loss V
reach the maximum value 1, ie

sup`
,,,
= max `
,,,
= 1
sup
,,,
= max
,,,
= 1
_
_
_
(3.30)
The maximums in Equation 3.30 are reached by the worst-case
asymptotic portfolio loss distribution
Pr(V
= 1) = 1
1
Pr(V
= 1
1 ) =
1
_
_
_
(3.31)
This theoremhas the followingimplications for the GHSRFmodel
with latent systematic and idiosyncratic risk factors. If the only
restriction for the distributions of the latent risk factors is the stan-
dardisation, then the risk measures value-at-risk and average value-
at-risk are only bounded by the trivial upper bound 1 for relevant
parameter constellations. To get sharper bounds, further restrictions
of the allowedrisk-factor distributions are necessary, eg, continuous,
symmetric or unimodal distributions with mean 0 and variance 1.
These restrictions appear to be arbitrary if the risk factors are latent
and not observable. The upper bounds in Equations 3.28 and 3.30
show impressively that the values for the value-at-risk and for the
average value-at-risk derived in the context of the common HSRF
model depend substantially on the assumption of normality for the
latent risk factors.
SUMMARY
This chapter focuses on the quantication of the model risk arising
from a potential misspecication of the distributions of latent risk
64
factors in a special class of credit-portfolio models. For this purpose,
new measures of model risk such as the model risk sets are intro-
duced. Since model risks can only be quantied if the underlying
model is dened, the SRF model with Gaussian distributed latent
risk factors and a special case, the HSRF model, with homogeneous
potential losses, default probabilities and correlation parameters,
are discussed. In addition, generalisations of both models are devel-
oped, where the distributional assumptions for the latent risk fac-
tors are reduced. More precisely, in the generalisedsingle-risk-factor
model andinthe generalisedhomogeneous single-risk-factor model
the latent risk factors are assumed only to be standardised (mean 0
and variance 1) instead of having a standard normal distribution.
Lower and upper bounds for the lower and upper quantiles for
all standardised random variables are given, and the correspond-
ing class of distributions which reach these bounds is presented.
Using these extremal distributions, it is shown that in the GHSRF
model with standardised systematic and idiosyncratic risk-factor
distributions for relevant parameter constellations no upper bound
less than 1 exists for the value-at-risk and the average value-at-risk.
The results highlight the fact that the values for the value-at-risk and
for the average value-at-risk usually calculated with the common
HSRF model depend substantially on the assumption of normality
for the latent risk factors.
MATHEMATICAL APPENDIX
Proof of Lemma 3.4 For the proof of the rst two statements see
Huschens and Vogl (2002, p. 287). The proof of the third state-
ment follows from McNeil et al (2005, pp. 3523) and the fact that
P(D
i
= 1 Z = z) = p
i
(z) and P(D
i
= 0 Z = z) = 1 p
i
(z).
Proof of Lemma 3.7
(i) For the proof of this statement see Schnbucher (2003, pp. 3078),
where a general default barrier is used instead of
1
().
(ii) The proof of this statement follows from Schnbucher (2003,
p. 308) and the fact that
Pr(nV
n
= k) = Pr
_
V
n
=
k
n
_
for k = 0, 1, . . . , n
65
MODEL RISK
(iii) Given that Z = z, the default variables D
i
are independent and
identically Bernoulli distributed with parameter E[D
i
Z = z] =
p(z) [0, 1].
It holds that
E[V
n
Z = z] = p(z) and V[V
n
Z = z] =
p(z)(1 p(z))
n
From
0 E[(V
n
p(Z))
2
] = E[(V
n
E[V
n
Z])
2
]
= E[E[(V
n
E[V
n
Z])
2
Z]]
= E[V[V
n
Z]]
= E
_
p(Z)(1 p(Z))
n
_
1
4n
it follows that V
n
converges in quadratic mean to p(Z), ie
lim
n
E[(V
n
p(Z))
2
] = 0
For eachxedvaluez R, thestronglawof largenumbers implies
that V
n
converges to p(z) with probability 1. As a result of this,
convergence with probability 1 to p(Z) also holds unconditionally
(Bluhm et al 2003, p. 89).
(iv) This statement follows fromBluhmet al (2003, Proposition 2.5.8)
and the fact that
1
() =
1
(1 ) for 0 < < 1.
Proof of Lemma 3.9
(i) The stochastic independence of Z and Y
i
implies that the random
variables

i
Z and
_
1
i
Y
i
are also stochastically independent.
Since the distribution functions of these random variables are
F
Z
_

i
_
and F
Y
_

_
1
i
_
the distribution function of A
i
can be expressed as the convolution
given in Equation 3.12.
(ii) Assumption 3.8 implies that
E[A
i
] = E[
_
i
Z +
_
1
i
Y
i
] =
_
i
E[Z] +
_
1
i
E[Y
i
] = 0
and
V[A
i
] = V[
_
i
Z +
_
1
i
Y
i
] =
i
V[Z] +(1
i
)V[Y
i
] = 1
(iii) This statement follows immediately from Equation 3.12.
66
Proof of Theorem 3.11
(i) FromEquation3.13, it follows that D
i
is Bernoulli distributedwith
default probability
Pr(D
i
= 1) = Pr(A
i
F
A
i
(p
i
)) = F
A
i
(F
A
i
(p
i
)) = p
i
for i = 1, . . . , n, where the last equality follows from Equation 3.15.
(ii) The probabilities Pr(D
i
= 1 Z = z) are dened uniquely for all
z
Z
R with Pr(Z
Z
) = 1. Then
Pr(D
i
= 1 Z = z) = Pr(A
i
F
A
i
(p
i
) Z = z)
= Pr(
_
i
Z +
_
1
i
Y
i
F
A
i
(p
i
) Z = z)
= Pr
_
Y
i

F
A
i
(p
i
)
i
z
_
1
i
_
= F
Y
_
F
A
i
(p
i
)
i
z
_
1
i
_
holds for i = 1, . . . , n, where the last two equalities follow from
Assumptions 3.8(i) and (iii).
(iii) Let d
i
0, 1 for i = 1, . . . , n. Then
Pr(D
1
= d
1
, . . . , D
n
= d
n
)
=
_
Z
Pr(D
1
= d
1
, . . . , D
n
= d
n
Z = z) dF
Z
(z)
=
_
Z
n
i=1
Pr(D
i
= d
i
Z = z) dF
Z
(z)
holds, where the last equality follows fromAssumption 3.8(i).
(i) This statement follows from Lemma 3.9 and the homogeneous
correlation parameters of Assumption 3.5.
(ii) This statement follows from Theorem 3.11(i) and the homoge-
neous default probabilities of Assumption 3.5.
(iii) This statement follows from Theorem 3.11(ii) and the homoge-
neous default probabilities and correlation parameters of Assump-
tion 3.5.
(iv) Equation 3.16 follows from
Pr
_
V
n
=
k
n
_
= Pr(nV
n
= k) =
_
Z
Pr(nV
n
= k Z = z) dF
Z
(z)
67
MODEL RISK
and the fact that for given Z = z the default variables are independ-
ent and identically Bernoulli distributed with parameter p(z; F
Z
, F
Y
)
so that nV
n
Z = z Bin(n, p(z; F
Z
, F
Y
)).
(v) Given Z = z, the default variables D
i
are independent and
identically distributed with E[D
i
Z = z] = p(z; F
Z
, F
Y
) [0, 1].
Analogously to the proof of Lemma 3.7(iii), it follows from
E[V
n
Z = z] = p(z; F
Z
, F
Y
),
V[V
n
Z = z] =
p(z; F
Z
, F
Y
)(1 p(z; F
Z
, F
Y
))
n
and
0 E[(V
n
p(Z; F
Z
, F
Y
))
2
]
= E
_
p(Z; F
Z
, F
Y
)(1 p(Z; F
Z
, F
Y
))
n
_
1
4n
that
lim
n
E[(V
n
p(Z; F
Z
, F
Y
))
2
] = 0
ie, V
n
converges in quadratic mean to p(Z; F
Z
, F
Y
).
For each xed value z in a set
Z
R with Pr(Z
Z
) = 1, the
strong lawof large numbers implies that V
n
converges to p(z; F
Z
, F
Y
)
with probability 1. Consequently, convergence with probability 1
to p(Z; F
Z
, F
Y
) also holds unconditionally (Bluhm et al 2003, p. 89).
This follows from the existence of a regular conditional probability,
whichis guaranteedfor real-valuedrandomvariables (Shorack2000,
pp. 168ff).
Proof of Lemma 3.15 The one-sided inequality of Cantelli for each
randomvariable XwithdistributionfunctionF
X
F
0,1
is (DasGupta
2008, p. 634)
F
X
(t)
1
1 +t
2
, t < 0
Let W be a random variable with distribution function
F
W
(t) =
_
_
1
1 +t
2
for t < 0
1 otherwise
68
Then F
X
(t) F
W
(t) for all t R and therefore F
W
() F
X
() for
all 0 < < 1 and especially F
W
() F
X
() with
F
W
() := mint F
W
(t) = min
_
t < 0
1
1 +t
2

_
= a
This proves the rst inequality in Equation 3.20. The secondinequal-

ity in Equation 3.20 follows from the denition of F
() and F
+
().
The third inequality in Equation 3.20 can be derived from the rst
inequality in Equation 3.20 as follows. If F
X
F
0,1
, then the distribu-
tion function F
X
of X is also contained in F
0,1
. The rst inequality
in Equation 3.20 gives

1
F
X
(1 )
and therefore
F
X
(1 ) b
Using F
+
X
() = F
X
(1 ) (Fllmer and Schied 2004, p. 177), the
third inequality in Equation 3.20 follows.
Proof of Lemma 3.16
(i) If X has a dichotomous distribution, then
Pr(X = a) = = 1 Pr(X = b)
for a parameter 0 < < 1 and two real numbers a < b. F
X
F
0,1
implies the two restrictions
E[X] = a +b(1 ) = 0 and E[X
2
] = a
2
+b
2
(1 ) = 1
for a and b. For xed , these equations hold if and only if a = a
and b = b
.
(ii) The distribution function in Equation 3.22 and the two quantile
functions in Equations 3.23 and 3.24 follow immediately from the
dichotomous distribution dened by Equation 3.21.
Proof of Theorem 3.17 The theorem follows immediately from Equa-
tions 3.20 and 3.25.
Proof of Lemma 3.18
(i) Let Z and Y be stochastically independent randomvariables with
F
Z
= F
Y
= G
and 0 < < 1. The discrete random variable

69
MODEL RISK
A :=
Z +
_
1 Y with 0 < < 1 has the support s
1
, s
2
, s
3
, s
4
with
s
1
=
_
a
+
_
1 a
s
2
=
_
a
+
_
1 b
s
3
=
_
b
+
_
1 a
s
4
=
_
b
+
_
1 b
From a
< 0 < b
, it follows that s
1
< 0 < s
4
with probabilities
Pr(A = s
1
) = Pr(Z = a
) Pr(Y = a
) =
2
and
Pr(A = s
4
) = Pr(Z = b
) Pr(Y = b
) = (1 )
2
In general, the support may consist of four different points
s
1
< mins
2
, s
3
maxs
2
, s
3
< s
4
If s
2
and s
3
coincide, then
Pr(A = s
2
= s
3
) = 2(1 )
else
Pr(A = s
2
) = Pr(A = s
3
) = (1 )
From
Pr(A maxs
2
, s
3
) = Pr(A < s
4
) = 1 Pr(A = s
4
) = 1 (1 )
2
and the assumption = 1
1 , which is equivalent to
1 (1 )
2
=
it follows that Pr(A maxs
2
, s
3
) = and therefore F
A
() =
maxs
2
, s
3
.
From Equation 3.16 and F
Z
= F
Y
= G
, it follows that
Pr
_
V
n
=
k
n
_
=
_
n
k
_
p(a
; G
, G
)
k
(1 p(a
; G
, G
))
nk
Pr(Z = a
)
+
_
n
k
_
p(b
; G
, G
)
k
(1 p(b
; G
, G
))
nk
Pr(Z = b
)
(3.32)
70
for k = 0, 1, . . . , n with Pr(Z = a
) = and Pr(Z = b
) = 1. From
Theorem 3.13(iii), it follows for the two Bernoulli parameters in the
mixture distribution of V
n
that
p(a
; G
, G
) = Pr(D
1
= 1 Z = a
)
= Pr(A F
A
() Z = a
)
= Pr(A maxs
2
, s
3
Z = a
)
= Pr(A < s
4
Z = a
)
= Pr(
_
Z +
_
1 Y < s
4
Z = a
)
= Pr(
_
a
+
_
1 Y < s
4
) (3.33)
= 1
and
p(b
; G
, G
) = Pr(A F
A
() Z = b
)
= Pr(
_
b
+
_
1 Y < s
4
) (3.34)
= Pr(Y < b
)
=
where Equations 3.33 and 3.34 follow from the independence of Z
and Y. Therefore
Pr(V
n
= 1) = p(a
; G
, G
)
n
Pr(Z = a
) +p(b
; G
, G
)
n
Pr(Z = b
)
= +
n
(1 )
which implies Equation 3.26, if
> 1 ( +
n
(1 )) =
1 (1 (1
1 )
n
)
(ii) Theorem 3.13 implies
V
= Pr(D
1
= 1 Z) = Pr(A F
A
() Z)
From the derivation above, it follows that
Pr(A F
A
() Z = a
) = 1 and Pr(A F
A
() Z = b
) =
Therefore, the distribution of V
is given by
Pr(V
= 1) = Pr(Z = a
) =
Pr(V
= ) = Pr(Z = b
) = 1
_
_
_
(3.35)
which implies Equation 3.27, if > 1 =
1 .
71
MODEL RISK
(i) Pr(V
n
1) = 1 implies VaR
[V
n
] 1 and AVaR
[V
n
] 1.
If >
1 (1 (1
1 )
n
) and if F
Z
= F
Y
= G
with
= 1
1 , then VaR
[V
n
] = AVaR
[V
n
] = 1, ie, the upper
bounds are reached. Using G
G, the equations
maxVaR
[V
n
] F
Z
, F
Y
G = maxAVaR
[V
n
] F
Z
, F
Y
G = 1
result. Since the upper bound 1 is reached as maximum in the set G
and since G F
0,1
, Equations 3.28 follow.
Using Equation 3.2 with value-at-risk and average value-at-risk
as risk measures and with V
n
as the random variable, the worst-
case loss distribution in Equation 3.29 follows from Equation 3.32
with Pr(Z = a
) = , Pr(Z = b
) = 1 , p(a
; G
, G
) = 1 and
p(b
; G
, G
) = , which are derived in the proof of Lemma 3.18.

(ii) The analogous reasoning for V
with >
1 leads to
maxVaR
[V
] F
Z
, F
Y
G = maxAVaR
[V
] F
Z
, F
Y
G
= 1
Since the upper bound 1 is reached as maximum in the set G and
since G F
0,1
, Equations 3.30 follow.
For the random variable V
the worst-case loss distribution

follows from Equation 3.35 with = 1
1 .
REFERENCES
Basel Committee on Banking Supervision, 2006, International Convergence of Capital
Measurement and Capital Standards: A Revised Framework, Comprehensive Version,
URL: http://www.bis.org/publ/bcbs128.htm.
Bluhm, C., L. Overbeck and C. Wagner, 2003, An Introduction to Credit Risk Modeling (Boca
Raton, FL: Chapman & Hall/CRC).
Burtschell, X., J. GregoryandJ.-P. Laurent, 2009, AComparativeAnalysis of CDOPricing
Models under the Factor Copula Framework, Journal of Derivatives 16(4), pp. 937.
Crouhy, M., D. Galai and R. Mark, 1999, Model Risk, The Journal of Financial Engineering
7, pp. 26788.
DasGupta, A., 2008, Asymptotic Theory of Statistics and Probability (New York: Springer).
Derman, E., 1996, Model Risk, Risk 9(5), pp. 347.
Fllmer, H., and A. Schied, 2004, Stochastic Finance: An Introduction in Discrete Time,
Volume 2 (Berlin: Walter de Gruyter).
Hamerle, A., and D. Rsch, 2005, Misspecied Copulas in Credit Risk Models: How
Good Is Gaussian?, The Journal of Risk 8(1), pp. 4158.
72
Huschens, S., and K. Vogl, 2002, Kreditrisikomodellierung imIRB-Ansatz von Basel II,
inA. Oehler (ed.) Kreditrisikomanagement: Kernbereiche, Aufsicht und Entwicklungstendenzen,
Second Edition, pp. 27995 (Stuttgart: Schffer-Poeschel).
Kerkhof, J., B. Melenberg and H. Schumacher, 2010, Model Risk and Capital Reserves,
Journal of Banking & Finance 34(1), pp. 26779.
McNeil, A. J., R. Frey and P. Embrechts, 2005, Quantitative Risk Management: Concepts,
Techniques and Tools (Princeton University Press).
Meyer zu Selhausen, H., 2004, Das Modellrisiko der Kreditportfoliomodelle: Konzep-
tionalisierung und Ursachen, in J. K. Thomas Burkhardt and U. Walther (eds) Banken,
Finanzierung und Unternehmensfhrung: Festschrift fr Karl Lohmann zum 65. Geburtstag,
pp. 27396 (Berlin: Duncker & Humblot).
Schnbucher, P. J., 2003, Credit Derivatives Pricing Models: Models, Pricing and Implementa-
tion (Chichester: John Wiley & Sons).
Shorack, G. R., 2000, Probability for Statisticians (New York: Springer).
Sibbertsen, P., G. Stahl and C. Luedtke, 2008, Measuring Model Risk, The Journal of
Risk Model Validation 2(4), pp. 6581.
Wang, D., S. T. Rachev and F. J. Fabozzi, 2009, Pricing Tranches of a CDO and a CDS
Index: Recent Advances and Future Research, in G. Bol, S. T. Rachev and R. Wrth (eds),
Risk Assessment: Decisions in Banking and Finance, pp. 26386 (Heidelberg: Physica).
73
4
Model Risk in Garch-Type
Financial Time Series
Corinna Luedtke, Philipp Sibbertsen
Leibniz Universitt Hannover
The market risk of a nancial institution is often measured by the
value-at-risk(VaR). The VaRis basicallya quantile of the forecast dis-
tribution of an underlying asset. The risk of obtaining an incorrect
VaRdue to the use of a misspecied statistical model is therefore the
risk of incorrect forecasting. Model uncertainty that stems frommis-
specication and incorrect estimation is a matter of high economic
relevance (Sibbertsen et al 2009).
It is well known that asset returns exhibit several empirical
stylised facts, such as volatility clustering, long memory in the
volatilities and a heavy-tailed return distribution. Many different
statistical models are around to reect these stylised facts. Models of
autoregressive conditional heteroscedasticity (Arch) introduced by
Engle (1982), and many varieties of these, model the volatility clus-
ters. These models may or may not exhibit long memory in order
to model the dependence structures of the volatilities. Furthermore,
models with different return distributions are considered to model
the heavy tails of the asset return distribution. However, it is not
yet clear which of these stylised facts and therefore different statis-
tical models have a signicant impact on the forecast of the return
distribution and thus on the VaR. Therefore, this chapter studies
the impact of various stylised facts and their statistical models on
measuring the market risk by the VaR.
The aim of this chapter is to discover which stylised facts have
an impact on the VaR. There exists a vast literature investigating
whichArch model is the best for out-of-sample VaRforecasting (see,
for example, Hansen and Lunde 2005). However, these papers only
use empirical data as input. We simulate 12 processes and estimate
these processes with different models of the Arch family. This leads
75
MODEL RISK
Table 4.1 S&P 500: descriptive statistics
Mean 0.0333
Standard deviation 1.0226
Skewness 0.0904
Excess kurtosis 4.0668
JB 2731.4830
p-value 0.0000
to universally valid evidence showing the model which performs
best.
The chapter is organised as follows. In the following section we
introduce the characteristics of nancial market data. These facts
can be captured by applying models of the Arch family, which are
describedinthe thirdsection. We next compute the VaRforecast with
Arch models to nd out whether the aforementioned stylised facts
do have an impact on the risk measure. The nal section contains
concluding remarks.
STYLISED FACTS
Stylised facts of nancial data are empirically observed character-
istics of the levels and returns of nancial data. For an overview
of stylised facts see Ghysels et al (1996). In this chapter we concen-
trate on those facts which are most important for modelling the VaR.
These are basically the fat tails of the distribution of stock returns,
volatility clustering, asymmetry in the volatilities and the high per-
sistence of the volatilities. These facts are described in more detail
in this section.
The stylised facts are empirically underpinned by the descriptive
statistics of the Standard & Poors 500 (S&P 500) stock index. The
daily data ranges fromAugust 31, 1990 to October 31, 2005. Table 4.1
displays the descriptive statistics of the time series.
Mandelbrot (1963) observed that returns are not normally dis-
tributed. Rather, they exhibit fat tails, meaning that extreme returns
are far more likely than under a normal distribution. On the other
hand, he observed that returns are more concentrated around their
mean, as expected under a normal distribution. Thus, stock returns
have a high probability mass at the tails and near the mean. They
therefore follow a leptokurtic distribution. Figure 4.1 displays a
76
MODEL RISK IN GARCH-TYPE FINANCIALTIME SERIES
Figure 4.1 Quantilequantile plot of S&P 500
2
2
4
6
0
0
4
6
2
2
Theoretical quantiles
S
i
m
p
l
e

q
u
a
n
t
i
l
e
s
quantilequantile plot of the S&P 500 stock index, where the empir-
ical quantiles are compared with the theoretical quantiles of the nor-
mal distribution. It can easily be seen that the tails deviate widely
from those of the normal distribution.
Furthermore, the stock returns exhibit excess kurtosis and the
JarqueBera test hypothesis of normal distributionis rejected. Exam-
ples of leptokurtic distributions are stable distributions with non-
existing rst or second moments. Distributions without an existing
variance are, however, problematic to apply in practice because the
variance is a good indicator for the risk of a nancial asset. Particu-
larly in the case of risk measurement, the underlying probability of
thedistributionshouldhaveexistingsecondmoments. Furthermore,
using leptokurtic distributions is a very inexible approach. Amore
exible approach is to assume time-varying conditional volatilities.
This approach also accounts for another stylised fact: the volatility
clustering.
Volatility clustering means that stock returns show patterns with
high volatility followed by patterns with lowvolatility. In Figure 4.2
the returns of the stock index tend to cluster. Volatility is quite low
at the beginning of the time series; in the middle, a cluster of high
volatility can be observed.
77
MODEL RISK
Figure 4.2 Returns of S&P 500
2
4
6
0
4
6
2
Figure 4.3 Autocorrelation function for squared returns of S&P 500
5
0.2
0
0.4
10 15 20 25 35 30
0.8
1.0
0.6
0
Lag
A
C
F
This pattern is an explanation for the leptokurtic distribution of
the stock returns. It also indicates that the conditional volatilities
of stock returns are time-varying. Since the 1980s an extensive lit-
erature has been published about modelling stochastic conditional
volatility. The most popular class of models are the Arch models
introduced by Engle (1982). Excepting the time-varying behaviour
of conditional volatilities, the volatility clustering also indicates a
strong persistence or long memory in the conditional volatilities.
Long memory in the conditional volatilities means that they have
a strong dependence structure with slowly hyperbolically decaying
78
autocorrelations. As in Figure 4.3, the autocorrelation function of
the squared returns decays at a hyperbolic rate in the case of a long
memory.
Due to long memory, conditional volatilities are persistent. This
is in accordance with the observed volatility clustering.
Another stylised fact not so closely related to the volatility clus-
tering is an asymmetry between the stock prices and the conditional
volatility. Stock prices and conditional volatility are negatively cor-
related; if the returns are negative, the uncertainty increases and
therefore the volatility increases as well. This effect is known as the
leverage effect.
Since the seminal paper by Engle (1982) introducing the Arch
models to describe time-varying stochastic volatilities, numerous
variants of this model have been developed in order to account
for the diversity of stylised facts, particularly because long mem-
ory and asymmetry cannot be captured by standard Arch models.
However, it is not clear so far how neglecting any of these stylised
facts inuences forecasts of the VaR.
VOLATILITY MODELLING
In 1982 Engle published his paper on the modelling of volatility
by introducing the class of models of autoregressive conditional
heteroscedasticity. Assume that a process
y
t
= +
t
exists, where the model shocks are
t
= z
t
_
2
t
and z
t
is a mean-zero independent and identically distributed (iid)
process with var(z
t
) = 1, which can either be normally distributed,
Student-t distributed or distributed otherwise. Then, the volatility
inperiodt depends onthe shocks that affectedy
t
inprevious periods
2
t
= +
q
_
i=1
2
ti
where all
i
are non-negative. Bollerslev (1986) generalised this
model to the Generalised Arch(p, q) (Garch), where the conditional
79
MODEL RISK
variance is modelled as a function of the lagged error as well as the
past conditional variance. The Garch(p, q) model sufces to
2
t
= +
q
_
i=1
2
ti
+
p
_
j=1
2
tj
where againall parameters
i
and
j
are non-negative. Furthermore,
the sum of the parameters has to be less than 1, that is
max(p,q)
_
i=1
(
i
+
i
) < 1
The model can be rearranged in lag notation as
[1 (L) (L)]
2
t
= +[1 (L)](
2
t

2
t
)
where (L) and (L) are the autoregressive and moving-average
polynomials, respectively.
Within these models, shocks have a symmetric impact on the vari-
ance. However, it is empirically observable that negative returns are
often followed by a period of high volatility. One popular model
where
t
has an asymmetric impact on the process equation is the
exponential Garch (EGarch) introduced by Nelson (1990). In order
to account for problems with the existence of higher moments the
EGarch model is formulated for the logarithm of the volatility. With
(L) and (L) as above and
1
and
2
being parameters which drive
the asymmetryof the model, BollerslevandMikkelsen(1996) rewrite
the EGarch(p, q) as
log
2
t
= +[1 (L)] 1[1 +(L)]g(z
t1
)
where
g(z
t
)
1
z
t
+
2
[z
t
Ez
t
]
the augend is the sign effect and the addend is the magnitude effect.
Another widely used asymmetric model is the Asymmetric Power
Arch (APArch) by Ding et al (1993)
t
= +
q
_
i=1
i
(
ti

i
ti
)
+
p
_
j=1
tj
where > 0 is the parameter that controls the persistence and 1 <
i
< 1 is the asymmetry parameter. Thus, this model is capable of
80
modelling persistence as well. The APArch model nests many other
models, one of which is the GJR-Garch by Glosten et al (1993)
2
t
= +
1
2
t1
+
2
t1
I(
t1
< 0) +
1
2
t1
where the indicator variable I takes a value of 1 for a negative shock
and 0 if the shock is positive.
Empirically, the impact of a shock decays at a much slower
rate than the previous models are able to model. The Fractionally
Integrated Garch (FIGarch) of Baillie et al (1996) can be obtained by
2
t
= [1 (L)]
1
. , .
+1 [1 (L)]
d
2
t
. , .
(L)
(4.1)
with (L) as above.
Davidson (2004) pointed out that the memory parameter d has
some paradoxical behaviour: persistence increases if the memory
parameter tends towards 0 and decreases if d 1. But, if d = 0,
the Garch model results (which has short memory), and if d = 1,
the IGarch results. One model that overcomes this problem is the
Hyperbolic Garch (HYGarch), which is capable of modelling both
persistence and asymmetry at the same time. If (L) in Equation 4.1
is replaced by
(L) = 1 [1 (L)]
1
(L)1 +[(1 L)
d
]
the HYGarch results.
SIMULATION STUDY
Simulation procedure
In order to evaluate the VaR, the forecast distribution of a nancial
asset is of interest. It is well known that different models can lead
to similar forecast distributions. In this situation, using a potentially
incorrect model does not lead to any model risk, in the sense that
the forecast distribution of the correct and incorrect model are sim-
ilar and therefore also the respective VaRs are similar. All models
described in the previous section aimto model various stylised facts
to t the data at hand as optimally as possible. The models therefore
aimfor an optimal in-sample t. Agood in-sample t does not auto-
matically go along witha goodout-of-sample t. Inthis Monte Carlo
study we consider the out-of-sample t in terms of VaR forecasts for
81
MODEL RISK
diverse models of the Garch family, in order to see which stylised
facts do signicantly inuence the forecast distribution.
To do this we simulated 2,000 observations of six models of
the Garch family as described in the previous section. We further-
more modelled the error distribution to be either a normal or a t-
distributionwithve degrees of freedom. The sample is dividedinto
an in-sample period of 1,600 observations used for the estimation of
the model and an out-of-sample period of 400 observations for eval-
uating the forecast accuracy. One-step-ahead, ve-step-ahead and
ten-step-ahead forecasts (h = 1, 5, 10) are considered in order to see
the performance for short, middle and long forecast horizons. Alto-
gether, there is an input of 12 simulated time series. In total, we
estimated and forecasted 432 models.
The idea of our Monte Carlo study is as follows. We use each
of the 12 models as a data-generating process (DGP) in order to
know which stylised facts are really present. We then t each of
the aforementioned models to the process. Explicitly, in a rst step
we simulate a Garch(1,1) model, say. Then, in a second step, we
t all of our Garch-type models to this simulated process and esti-
mate the parameters from the generated sample. In the next step
we forecast this process and compute the forecast error. If the DGP
is a Garch(1,1) process, we know that it does not contain any long
memory or asymmetry. If the practitioner, however, believes that
the data exhibits long-range dependencies and therefore ts, say, an
HYGarch model to the data, we can by this study see whether this
belief is largely erroneous. We therefore get a feeling for the model
risk involvedby using misspeciedGarch models for a VaRforecast.
We evaluate the forecasts by mean-squared errors (MSEs). Using the
mean absolute error (MAE) or the root-mean-squared error (RMSE)
leads to similar results. Therefore, they are not reported here. The
results are available from the authors on request.
In order to have a greater clarity in our presentation we number
the investigated models as in Table 4.2.
The parameters of the DGPs are given in Table 4.3.
Simulation Results
The forecasts are computed with the G@RCH software package
(Doornik and Ooms 2006). In Table 4.4 the results for a Garch(1,1)
DGP with a normal marginal distribution and a high persistence are
82
Table 4.2 Declaration of models used
Model N t
Garch (1) (2)
EGarch (3) (4)
GJR-Garch (5) (6)
APArch (7) (8)
FIGarch (9) (10)
HYGarch (11) (12)
Table 4.3 Parameters of the data-generating processes
df d
(1) 0.5 0.1 0.6
(2) 0.5 0.1 0.6 5
(3) 0.05 0.1 0.6 0.15 1
(4) 0.05 0.1 0.6 0.15 1 5
(5) 0.05 0.1 0.6 0.15 1
(6) 0.05 0.1 0.6 1 5
(7) 0.5 0.1 0.3 0.15 1
(8) 0.5 0.1 0.6 0.15 1.5 5
(9) 0.05 0.1 0.01 0.15 1.5 0.8
(10) 0.05 0.01 0.01 0.05 5 0.8
(11) 0.05 0.01 0.01 0.8
(12) 0.5 0.01 0.01 5 0.8
displayed.
1
Surprisingly, the lowest MSE for a one-step forecast is
obtainedfor a FIGarch model. This can be explainedby the high per-
sistence due to the parameter settings of the Garch model. It should
be noted that the MSE does not differ between most models.
However, the forecasting error which determines the model risk
is rather small even when an asymmetric model is used. The EGarch
model performs best even for longer forecast horizons. Similar
results can be found for asymmetric DGPs. For an EGarch model,
for example, we nd again that a long-memory model (here the
HYGarch model) has the lowest MSE for a one-step forecast. For
longer forecasting horizons, the correctly specied EGarch model
has the smallest MSE. However, again we nd that all MSEs are
rather close together. For the APArch model, asymmetric models
dominate the short-term and the long-term forecasts, as expected.
83
MODEL RISK
Table 4.4 DGP Garch(1,1)-N: mean squared error and p-values of the
DieboldMariano test
h = 1 h = 5 h = 10
, .. , .. , ..
MSE DM MSE DM MSE DM
(1) 5.2238 0.0760
5.2010 0.0000
5.1548 0.0000
(2) 5.2253 0.0571
5.2023 0.0000
5.1560 0.0000
(3) 5.2456 0.1121 5.1525
5.1069
(4) 5.2352 0.3082 5.2012 0.0000
5.1430 0.0000
(5) 5.2451 0.0053
5.2058 0.0000
5.1554 0.0000
(6) 5.2458 0.0045
5.2068 0.0000
5.1565 0.0000
(7) 5.2443 0.0051
5.2084 0.0000
5.1581 0.0000
(8) 5.2450 0.0044
5.2098 0.0000
5.1597 0.0000
(9) 5.1868
5.1620 0.3012 5.1138 0.3012

(10) 5.2341 0.1243 5.2066 0.0000
5.1943 0.0000
(11) 5.2211 0.2990 5.1954 0.0000
5.1518 0.0000
(12) 5.2155 0.0293
5.1914 0.0006
5.1426 0.0000
, signicance at 0% level;
, signicance at 5% level;
, signicance
at 1% level;

, model with lowest MSE.
Table 4.5 DGP EGarch(1,1)-N: mean squared error and p-values of
the DieboldMariano test
h = 1 h = 5 h = 10
, .. , .. , ..
(1) 5.3658 0.0142
5.3919 0.8287 5.3958 0.0000
(2) 5.3650 0.0092
5.3911 0.2961 5.3950 0.0000
(3) 5.4411 0.0003
5.2615
5.2528
(4) 5.4290 0.0004
5.3833 0.0117
5.3369 0.0000
(5) 5.4434 0.0012
5.4060 0.1078 5.4063 0.0000
(6) 5.4404 0.0012
5.4006 0.0049
5.4003 0.0000
(7) 5.4124 0.0010
5.3010 0.0097
5.2947 0.0000
(8) 5.4115 0.0010
5.2929 0.0041
5.2856 0.0000
(9) 5.3641 0.0151
5.3897 0.1876 5.3938 0.0000
(10) 5.3106
5.3299 0.0000
5.3304 0.0015
(11) 5.3641 0.0151
5.3897 0.0000
5.3938 0.0000
(12) 5.3621 0.0110
5.3872 0.0000
5.3914 0.0000
Here again the EGarch models deliver the best forecasts; interest-
ingly enough, these are even better than the correct APArch model.
For the asymmetric GJR-Garchmodel we ndthat the EGarchmodel
84
Table 4.6 DGP GJR-Garch(1,1)-N: mean squared error and p-values
of the DieboldMariano test
h = 1 h = 5 h = 10
, .. , .. , ..
(1) 1.1384 0.0669
1.1664 0.3849 1.1715 0.0000
(2) 1.1399 0.0060
1.1691 0.0761
1.1751 0.0000
(3) 1.1208 0.0050
1.0847
1.0689
(4) 1.1200
1.1587 0.9615 1.1745 0.0000
(5) 1.1387 0.0068
1.1666 0.4049 1.1716 0.0000
(6) 1.1405 0.0060
1.1695 0.0797
1.1755 0.0000
(7) 1.1390 0.0068
1.1680 0.4320 1.1736 0.0000
(8) 1.1405 0.0060
1.1696 0.4177 1.1756 0.0000
(9) 1.1668 0.0000
1.2142 0.2278 1.2323 0.0000
(10) 1.1707 0.0000
1.2214 0.0863
1.2425 0.0000
(11) 1.1386 0.0076
1.1660 0.3849 1.1712 0.0000
(12) 1.1403 0.0070
1.1688 0.0940
1.1750 0.0000
Table 4.7 DGP APArch(1,1)-N: mean squared error and p-values of the
DieboldMariano test
h = 1 h = 5 h = 10
, .. , .. , ..
(1) 9.3512 0.0206
8.7667 0.0000
8.5554 0.0000
(2) 9.3358 0.0347
8.7552 0.0000
8.5438 0.0000
(3) 9.3289 0.0022
8.5249
8.2514
(4) 9.2752
8.5953 0.0000
8.5155 0.0000
(5) 9.4162 0.0012
8.8029 0.0000
8.5580 0.0000
(6) 9.4033 0.0017
8.7928 0.0000
8.5483 0.0000
(7) 9.4161 0.0012
8.8032 0.0000
8.5588 0.0000
(8) 9.4033 0.0017
8.7939 0.0000
8.5500 0.0000
(9) 9.3298 0.0312
8.7326 0.0000
8.5151 0.0000
(10) 9.3213 0.0457
8.7320 0.0000
8.5163 0.0000
(11) 9.3544 0.0153
8.7613 0.0000
8.5628 0.0000
(12) 9.3187 0.0478
8.7278 0.0000
8.5113 0.0000
has the best forecasting performance, even for small forecast hori-
zons. If the true DGPexhibits longmemory, thenlong-memorymod-
els also have the best forecasting performance. For the HYGarch
model, for example, the FIGarch models have the smallest MSEs,
85
MODEL RISK
Table 4.8 DGP FIGarch(1,1)-N: mean squared error and p-values of
h = 1 h = 5 h = 10
, .. , .. , ..
(1) 3.8341 3.5716 0.0000
3.3362 0.0134
(2) 3.8310 0.0000
3.5653 0.0004
3.3288 0.0256

(3) 3.8047 0.0000
3.5469
3.3256
(4) 3.7024
3.5581 0.0000
3.3574 0.0000
(5) 3.8117 0.0000
3.5641 0.2112 3.3353 0.0131

(6) 3.8083 0.0000
3.5571 0.0014
3.3270 0.0313

(7) 3.8966 0.0000
4.1964 0.3900 4.4736 0.0000
(8) 3.8960 0.0000
4.2450 0.0022
4.5838 0.0000
(9) 4.2558 0.0000
4.1522 0.0007
4.0772 0.0000
(10) 4.2574 0.0000
4.1542 0.0482
4.0795 0.0000
(11) 4.1357 0.0000
4.1004 0.0000
3.8775 0.0000
(12) 4.2142 0.0000
4.1004 0.0000
4.0096 0.0000
Table 4.9 DGP HYGarch(1,1)-N: mean squared error and p-values of

h = 1 h = 5 h = 10
, .. , .. , ..
(1) 0.9044 0.0000
0.9122 0.0000
0.9189 0.0000
(2) 0.9072 0.0000
0.9123 0.0000
0.9190 0.0000
(3) 0.9018 0.0149
0.9103 0.0000
0.9165 0.0000
(4) 0.9075 0.0038
0.9346 0.0000
0.9407 0.0000
(5) 0.9036 0.0000
0.9109 0.0000
0.9176 0.0000
(6) 0.9066 0.0041
0.9129 0.0000
0.9190 0.0000
(7) Not converged

(8) Not converged
(9) 0.8945
0.9026
0.9103
(10) 0.8946 0.0000
0.9027 0.0000
0.9104 0.0000
(11) 0.8999 0.0000
0.9089 0.0000
0.9167 0.0000
(12) 0.9072 0.0000
0.9130 0.0000
0.9190 0.0000
although we must admit that this does not hold true for the FIGarch
model, which seems to be best predicted by means of the EGarch
model. However, this might be due to well-known problems with
the memory structure of FIGarch models.
86
It should be noted that these results are independent of the
marginal distributionandof the chosensignicance level. It seems to
be important to adequately t the stylised facts, such as long mem-
ory or the asymmetry present in the data. The heavy tails of the
return distributions are not that important in terms of model risk.
All models have the ability to model these heavy tails adequately.
The longer the forecastinghorizon, the more important the adequacy
of the in-sample model becomes, because a shorter forecasting hori-
zon, wrongly specied, can deliver goodforecasts as well. For exam-
ple, long-memory models also have a good forecasting quality even
for short-memory asymmetric models. For long-term forecasts it is
important to adequately model the present stylised facts. It should
be emphasised here that it is not important to nd the exact model
of the DGP. Asymmetric APArch or GJR-Garch models are forecast
well by asymmetric EGarch models. Thus, it is important to model
the general structure of the underlying DGP, but not to nd exactly
the true DGP.
For evaluating the accuracy of the forecasts we compute the test
statistic of Diebold and Mariano (1995): let e
it
and e
jt
be the forecast
errors of two computed forecasts for which we would like to test
the comparative predictive accuracy and let g(e
it
) and g(e
jt
) be the
functions of the forecast errors. Then, the loss differential series can
be expressed by
d
t
= [g(e
it
) g(e
jt
)]
The null hypothesis of equal forecast accuracy is rejected if the loss
differential between the two forecast series deviates from zero.
It canbe seenfromthe tables that the DieboldMarianotest almost
always rejects the null of equal forecast accuracy. An exception is
the case of a Garch(1,1) DGP where the null cannot be rejected for
the FIGarch and EGarch model. However, as we aimto quantify the
inducedmodel risk, this nding is interesting; nonetheless, the abso-
lute value of the MSE is of greater interest for us as it gives a feeling
for the inducedmodel risk when misspecifying the underlying DGP.
CONCLUSION
When evaluating the market risk by means of the VaR, the practi-
tioner usually models the distribution of the underlying nancial
assets by means of models of conditional heteroscedasticity, such
87
MODEL RISK
as Garch-type models, and then derives the forecasting distribution
from these models. The aim of nding a suitable model for the data
at hand is to obtain a good in-sample t obtained by applying var-
ious specication tests. However, for risk measurement, the out-of-
sample t is of greater importance. It is knownthat wronglyspecied
models could also lead to good forecasts. Therefore, the aim of this
chapter was to evaluate the model risk involved when the model for
the underlying nancial asset is misspecied. In order to do this, we
conducted a Monte Carlo study where we generated various DGPs
modelling the various stylised facts which may be found in nan-
cial data. We considered a short-memory symmetric Garch model
as the simplest DGP, and various asymmetric models, such as the
EGarch, APArch andGJR-Garch models, to cover the leverage effect,
and FIGarch and HYGarch models as long-memory models. Fore-
casts are made for short-, medium- and long-term forecasting hori-
zons. As marginal distributions, we allowed for a standard normal
distribution and a t-distribution with ve degrees of freedom.
It turned out that modelling asymmetry and long memory is cru-
cial for the forecast performance. As long as it is accounted for (for
either asymmetry or long memory) it is of minor importance which
specic model is specied. However, neglecting these stylised facts
leads to a serious forecasting error and therefore to high model risk.
Specifying the correct marginal distribution is of minor importance.
All models seem to be able to cover the heavy tails of the return
distribution well enough.
Altogether, we nd that a correct specication of the underly-
ing data structure is essential to reduce the induced model risk. It
is important to have a good in-sample t. Thus, a correct model
specication is of major importance when evaluating the market
risk.
Our work concentrated on univariate stylised facts. For future
research it is of interest which multivariate stylised facts, such as
time-varying correlations between returns (as described by Longin
and Solnik (1995)), have an impact on risk forecasts.
1 The results for the other data-generating processes canbe foundinTables 4.54.9. More results
are available upon request.
88
REFERENCES
Baillie, R. T., T. Bollerslev and H. O., Mikkelsen, 1996, Fractionally Integrated
Generalized Autoregressive Conditional Heteroscedasticity, Journal of Econometrics 74,
pp. 330.
Bollerslev, T., 1986, GeneralizedAutoregressive Conditional Heteroscedasticity, Journal
of Econometrics 31, pp. 30727.
Bollerslev, T., and H. O. Mikkelsen, 1996, Modeling and Pricing Long Memory in Stock
Market Volatility, Journal of Econometrics 73, pp. 15184.
Davidson, J., 2004, Moment andMemory Properties of Linear Conditional Heteroscedas-
ticity Models, and a New Model, Journal of Business and Economic Statistics 22, pp. 1629.
Diebold, F. X., and R. S. Mariano, 1995, Comparing Predictive Accuracy, Journal of
Business and Economic Statistics 13, pp. 25363.
Ding, Z., C. W. J. Granger, and R. F. Engle, 1993, A Long Memory Property of Stock
Market Returns and a New Model, Journal of Empirical Finance 1, pp. 83106.
Doornik, J. A., and M. Ooms, 2006, Introduction to Ox, URL: http://www.doornik.
com/ox/oxintro.pdf.
Engle, R., 1982, Autoregressive Conditional Heteroscedasticity with Estimates of the
Variance of United Kingdom Ination, Econometrica 50, pp. 9871008.
Ghysels, E., A. Harvey, E. Renault and G. S. Maddala, 1996, Stochastic Volatility, in
G. S. Maddala and C. R. Rao (eds), Statistical Methods in Finance, Handbook of Statistics,
Volume 14, pp. 11992 (Elsevier Science).
Glosten, L. R., R. Jagannathan, and D. E. Runkle, 1993, On the Relation between the
Expected Value and the Volatility of the Nominal Excess Return on Stocks, The Journal of
Finance 48, pp. 1779801.
Hansen, P. R., and A. Lunde, 2005, A Forecast Comparison of Volatility Models: Does
Anything Beat a Garch(1, 1)?, Journal of Applied Econometrics 20, pp. 87389.
Longin, F., and B. Solnik, 1995, Is the Correlation in International Equity Returns
Constant: 19601990? Journal of International Money and Finance 14, pp. 326.
Mandelbrot, B., 1963, The Variation of Certain Speculative Prices, The Journal of Business
36, pp. 394419.
Nelson, D. B., 1990, Stationarity and Persistence in the Garch(1, 1) Model, Econometric
Theory 6, pp. 31834.
Sibbertsen, P., G. Stahl and C. Luedtke, 2009, Measuring Model Risk, The Journal of
Risk Model Validation 2(4), pp. 6581.
89
Part II
MACROECONOMIC AND
CAPITAL MODELS
5
Monetary Policy, Asset Return
Dynamics and the General
Equilibrium Effect
Kuang-Liang Chang; Nan-Kuang Chen;
Charles KaYui Leung
National Chiayi University; National Taiwan University;
City University of Hong Kong
The 20089 nancial crisis threatened not only the global economy,
but also the reputation of the eld of economics and nance. Afew
quotations are sufcient to illustrate the point:
Economists mostly failed to predict the worst economic crisis since
the 1930s. Nowthey cant agree howto solve it. People are starting
to wonder: What good are economists anyway? Acommenter on a
housingblogwrote recentlythat economists didaworse jobof fore-
casting the housing market than either his father, who has no for-
mal education, or his mother, who got up to second grade. If you
are an economist anddidnot see this coming, youshouldseriously
reconsider the value of your education and maybe do something
with a tangible value to society, like picking vegetables.
(Business Week 2009)
It should be noted that Business Week used to report many opinions
and comments by economists on different issues, and quotes like
the one above in that magazine are very unusual. In fact, nega-
tive views can also be found among well-established economists.
For instance, London School of Economics Professor Willem Buiter
commented that
Most mainstreammacroeconomic theoretical innovations since the
1970s have turned out to be self-referential, inward-looking dis-
tractions at best. Research tended to be motivated by the internal
logic, intellectual sunk capital and esthetic puzzles of established
research programmes rather than by a powerful desire to under-
stand howthe economy works let alone howthe economy works
93
MODEL RISK
during times of stress and nancial instability. So the economics
profession was caught unprepared when the crisis struck.
(Buiter 2009)
In this context, it is not surprising that the rescue plan of the crisis
quickly turned into a heated debate.
In light of that, this chapter attempts to summarise some recent
academic efforts which attempt to take the interaction of the asset
markets and the aggregate economy more seriously. As summarised
by Leung (2004), there is a historical disconnection between the liter-
ature on the real estate market and macroeconomics. As the 20089
crisis is deeply intertwined with the real estate market (the hous-
ing market, the mortgage-backed securities, etc), it is no surprise
that some early diagnoses of the crisis may be not exactly accurate.
Thus, an ideal approach would be to explicitly embed the asset
markets, including both the real estate and stock markets, into a
dynamic general equilibriumframework, and allowfor interactions
between the two. Clearly, this is not a simple task.
Kiyotaki and Moore (1997) were perhaps the rst to include the
collateral constraint in a dynamic general equilibriummodel. In that
environment, the (relative) price of capital would affect the distribu-
tion of productivity among producers with heterogeneous produc-
tivities, and hence the aggregate output. Chen (2001) shows that the
changes in asset price can also inuence the net worth of banks, and
hence their lending capacity. As a result, the real economic activities
will be affected.
Ortalo-Magne and Rady (2006) study how the collateral con-
straints affect the equilibrium trading volume and housing price
in a life-cycle economy. Iacoviello (2005) abstracts fromthe life-cycle
complications and studies how the monetary policy would affect
the house price and the economy when some agents face the collat-
eral constraint. These papers, among others, are all important con-
tributions, helping us to ll the historical gap of the relationship
between the housing market and the aggregate economy.
However, reality is often a cruel master and sometimes demands
more than the academic circle can deliver, at least in the short run.
In particular, many earlier contributions abstract from the consid-
eration of the stock market, making their analysis not immedi-
ately applicable to the 20089 crisis. Recently, some scholars have
attempted to pursue that direction. For instance, Leung (2007) has
94
MONETARY POLICY, ASSET RETURN DYNAMICS ANDTHE GENERAL EQUILIBRIUM EFFECT
built perhaps the rst dynamic general equilibriummodel, in which
the asset prices (housing and stock) and the various kinds of invest-
ment (business capital and housing) are all endogenous. In a nut-
shell, Leung (2007) has combined the asset pricing model of Lucas
(1978) and the household capital economy of Greenwood and Her-
cowitz (1991) with the additional assumption that the price of hous-
ing relative to business capital can deviate fromunity. He has found
that asset prices should be serially correlated as well as cross-
correlated. Jaccard (2007) has studied the relationships among the
aggregate economy, the house price and the commercial real estate
price. He has found that it is indeed important to distinguish the
two kinds of real estate, as well as to explicitly model the construc-
tion sector. Jaccard (2008) moved a further step forward and studied
the possible relationship between household production and stock
price. To put it in another way, the utility delivered to the house-
holds depends not only on the size (and other physical features)
of the housing units, but also on how much time the households
are able to spend in the house. And the time the household can
spend at home clearly depends on the labour market condition,
which is in turn inuenced by the aggregate economy. He found
that the explicit consideration of the household production process
helps us to explain both the house price movement and the equity
premium.
Building on the insight of all these works, Leung and Teo (2008)
were perhaps the rst to explicitly consider the role of monetary
policy when the agents can invest among housing, bond and stock.
They found that the correlation between the stock price and housing
price (both measured in real terms) varies signicantly across differ-
ent monetary regimes. Thus, the change of monetary regimes could
have important implications for the optimal portfolio, even when
the change in monetary regimes may not have a very large effect
on the aggregate consumption and investment. This could justify
why Wall Street is so concerned about the monetary policy regimes.
It is also consistent with the fact that the correlation between the
stock price and housing price varies signicantly across countries,
as different central banks may practise different monetary policies.
Stimulated by these ndings, Leung and Teo (2009) investigated
the possibility that the same monetary policy may have different
impacts on different regions in the same country. Clearly, this is also
95
MODEL RISK
inspiredby a series of empirical works onthe difference inreal estate
market in different cities. For instance, Green et al (2005) found that
the supply elasticities of housing can vary from 30 (such as Dallas)
to virtually zero (such as San Jose, San Francisco or New Orleans).
Given such discrepancies in the supply elasticities, it should not
be surprising that the house price dynamics also varies across
metropolitanareas. Infact, Malpezzi (2008) showedthat, while some
metropolitan areas in the US experienced very signicant cycles in
the past few decades, there are some cities whose housing prices
were essentially at during the same sampling period.
Motivated by these empirical ndings, Leung and Teo (2009)
investigated the situation in which the only difference between two
regions in ex ante terms is the adjustment cost of housing investment
(which translates into a difference in housing supply elasticities).
Since the productivity shocks are region specic in their models,
several stylised facts are reproduced. For instance, the correlation
between the national GDPand the regional housing price is on aver-
age weaker than that between the regional GDP and the regional
housing price. The regions with higher adjustment costs in housing
investment also naturally have a smaller volatility of the housing
investment. Perhaps more interestingly, theyndthat the correlation
between the national stock market price and the regional housing
price can differ greatly between regions, suggesting the possibility
of a region-specic portfolio. Their paper successfully demonstrates
that it is possible to study some regional economic question with
some nancial economic question in a unifying framework. They
also conrmthe intuition that the disaggregate house price, if avail-
able, may enhance a better evaluation for different economic poli-
cies. These results are also consistent with the empirical ndings of
Carlino and Dena (1998, 1999, 2006), among others, that monetary
policy has differential impacts in different regions, and the frame-
work of Leung andTeo (2009) suggests that one possible explanation
is the differential elasticities of housing supply in different regions.
While the work of Leung and Teo (2009), among others, has deep-
ened our understanding of the cross-sectional effect of monetary
policy, it still does not address one of the key policy concerns: the
potentially time-varying effect of the monetary policy. Casual obser-
vations seemto suggest that this may be an important issue. During
a boom, the central banks typically will try to increase the interest
96
rate inorder tosuppress the irrational exuberance inthe asset mar-
kets, yet they have mixed results. During a downturn, however, the
interest rate (cut) seems to have much a bigger impact on the asset
markets (see, for example, Chan and Leung (2008) for more discus-
sion on this point). More generally, it is crucial for the policy makers
to acquire information on the following issues, among others.
Is it true that the real economy and the asset markets would
react to the monetary policy differently during different
phases of the boombust cycle?
Should such differences exist, is it possible to assess which
phase the current economy is in? Moreover, is it possible to
predict whenthe economyandthe asset market wouldswitch
to another phase?
Basedon the historical data, is it possible to estimate the effects
of the monetary policy on different asset markets in different
phases of the boombust cycle?
The experience of the lost decade
1
in Japan, the Nordic and East
Asian crises during late 1980s and 1990s and the 20089 subprime
crisis, among other similar episodes, seems to support the idea that
the depth of knowledge of the policy makers regarding these ques-
tions could make a difference to the nal outcome. Therefore, while
theoretical works are important and should be encouraged, there is
no substitute for the empirical works along these lines.
Henceforth, we focus on two papers, both using regime-switch-
ing structural vector auto-regressive (SVAR) models to examine the
effects of monetary policy on the asset markets. Needless to say, we
need to rst explain why such an approach is adopted.
WHY REGIME-SWITCHING SVAR MODELS?
Chang et al (2008, 2009), among others, take an initial step in this
direction by building various versions of the SVAR models in order
to study the interaction between the monetary policy, the real econ-
omy and the asset markets. In particular, they differentiate their
efforts fromthe previous literature byallowingtwo-waycausality,
ie, not only that the asset market returns be affected by the mone-
tarypolicyandthe real economy, but also that the asset returns could
have an inuence on the monetary policy andthe real economy. This
is a well-known advantage of the SVAR model.
97
MODEL RISK
The second innovation is to allow for the possibility of regime-
switching, ie, allowing the relationship among the monetary policy,
the real economy and the asset markets to vary across different
time periods. Such a consideration is motivated by several strands
of the literature. On the theoretical front, regime switching can
occur in a general equilibrium setting. Azariadis and Smith (1988)
may have built the rst general equilibrium in which the market
economy would permanently uctuate between the regime with
credit rationing and the regime without it. Chen and Leung (2008),
among others, apply their insight in the asset market analysis. They
consider a situation where the land and structure (real estate) can
be used for both residential and commercial purposes. Both the
households and some constrained producers are leveraged in their
purchase of this real estate. Chen and Leung show that under some
conditions a negative shock will initiate a re sale, ie, both the
household side and the producer side are trying to sell real estate,
which leads to a further drop in the asset price and more liquidation
(spillover effect). Perhaps more importantly, they show analytically
that the correlations between the real output and the asset price
are different under different regimes; the same amount of output
decline is associated with a larger decrease in asset price in the bad
regime (ie, when the spillover effect of the asset price occurs). In
other words, a linear vector auto-regressive (VAR) model between
the real economy and the asset price could deliver misleading
results. This clearly calls for an econometric framework which will
allow for some degree of non-linearity.
On the empirical front, it has long been suggested that economic
time series may be better characterised by a Markov regime-switch-
ing process, rather than a smooth autoregressive moving-average
(ARMA) process. For instance, Hamilton (1989) shows that the
aggregate output in the US can be characterised by such a pro-
cess. Regime-switchingmodels have since beenwidelyusedinmod-
ellingdifferent classes of asset prices, includingstock, option, foreign
exchange, interest rate, etc (see, for example, Hansen and Poulsen
2000; Maheu and McCurdy 2000).
The empirical literature on market efciency provides an addi-
tional reason to employ the regime-switching model. It has long
been recognised that the change in housing price, or the housing
return (we will use these two terms interchangeably throughout),
98
is predictable, at least in the short run (see, for example, Case and
Shiller 1990; Shiller 2008). It has been cited as evidence for market
efciency. Although the logical link between predictability and mar-
ket efciency is weak,
2
it is still puzzling from an intuitive point of
view: if the short-run house price can be predicted, then people may
leverage to buy andprot fromit, as much as the leverage will allow.
In other words, there is room for arbitrage for prot, which seems
to contradict the notion of an efcient market. In fact, this intu-
ition applies only for the single-regime econometric model, such as
the conventional ordinary-least-squares or linear VARmodels. With
multiple regimes and the randomness in the occurrence of regime
switching, the situation can be very different. To x this idea, let us
consider the following simple example
y
t+1
=
_
_
_
a
1
y
t
+u
t+1
when y
t+1
is in regime 1
a
1
y
t
+u
t+1
when y
t+1
is in regime 2
where a
1
a
1
, and the distributions of the error terms u
t+1
, u
t+1
may
not be identical. (The language here is somewhat informal. More
formal treatment on the regime-switching model will be provided
in later sections.) And we further assume that the probability for
y
t+1
is in regime j, given that y
t
is in regime i, is p
ij
, 0 < p
ij
< 1,
i, j = 1, 2. Now consider the case when the regime tends to be very
persistent, meaning that p
ii
, i = 1, 2, is very close to unity. For the
sake of argument, assume that the system is now in regime 1. Then,
in the short run, this system behaves like a simple AR(1) process
y
t+1
= a
1
y
t
+u
t+1
andhence becomes very predictable. Onthe other hand, as long as
p
ii
< 1, regime switching occurs and those who bet based on a sim-
pleAR(1) process will suddenly realise a capital loss. In other words,
the persistence of regime can lead to short-run predictability, while
the randomoccurrence of regime switching rules out long-run prof-
itability. Short-run predictability can be compatible with long-run
non-protability in a multiple regime setting. The regime-switching
model brings in not only new econometric tools, but also new eco-
nomic intuition. We therefore consider that the employment of a
regime-switching SVAR model is appropriate in this stage of the
empirical study of the policy effect on asset returns.
99
MODEL RISK
Table 5.1 Statistical summary of federal funds rate, interest rate spread,
housing market returns, and equity REIT returns (1975 Q22008 Q1)
FFR SPR HRET REIT
Mean 6.464 1.490 1.401 1.519
Median 5.618 1.581 1.345 1.852
Maximum 17.780 3.611 4.425 18.523
Minimum 0.997 2.182 0.406 18.174
Std. Dev. 3.493 1.341 0.947 6.849
Skewness 1.040 0.604 0.564 0.182
Kurtosis 4.307 2.904 3.284 3.173
Observations 132.000 132.000 132.000 132.000
FFR, federal funds rate; SPR, interest rate spread; HRET, housing market
returns; REIT, equity REIT returns.
DATA ANDTHE ECONOMETRIC MODEL
Chang et al (2008, 2009) (henceforth CCL 2008 and CCL 2009) study
the effect of monetary policy on asset markets with US data. The
justication is clear: the US is the largest economy in the world and
arguably a goodapproximationof the typical textbook closedecon-
omy case. Moreover, many empirical studies on the effect of mon-
etary policy are based on US data (see, for example, Sargent et al
2006; Sims 1980a,b; Sims and Zha 2006). Choosing the same country
as these studies will facilitate the comparison.
The sampling period of both papers begins with 1975 Q2, when
the rst ofcial house price index is available. Other time series
typically begin much earlier (see, for example, Leeper et al 1996) and
hence the availability of the house price seems to be the major data
constraint. Table 5.1 is adapted fromChang et al (2008) and provides
a summary of the data (see also Figures 5.1). Aquick look at the data
will give the impression that the data series do not behave like a
normal distribution, but rather displaysomefat-tailedproperties.
3
This is consistent with the ndings of some earlier literature. For
instance, signicant differences between the real estate investment
trusts (REITs) listedinthe1990s andthoselistedbefore(includingthe
liquidity, size, the degree of focus by property type, nancing policy,
capital structure, etc) have beenfoundandthese mayleadtochanges
in the return distribution of REIT (see, for example, Capozza and
Seguin 1998, 1999). Also, the conduct of monetary policy might have
changed over time, along with different chairmanship of the Federal
100
Figure 5.1 (a) Federal funds rate (FFR), (b) interest rate spread (SPR),
(c) housing market returns (HRET) and (d) equity REIT returns (REIT)
1980 1990 2000 1980 1990 2000
1980 1990 2000 1980 1990 2000
4
3
2
1
0
1
2
3
20
10
0
10
20
(a) (b)
(c) (d)
20
16
12
8
4
0
4
3
2
1
0
1
5
Reserve Board and during several dramatic episodes of aggregate
shocks. Therefore, it might be useful to compare the case of single
regime case (ie, the conventional linear VARmodel) withthe regime-
switching case.
Followingthe literature, the three-monthFederal FundRate (FFR)
is employed as a proxy of the monetary policy. The housing return
(HRET) is computed from the housing price index provided by the
Ofce of Federal Housing Enterprise Oversight (OFHEO). The REIT
return(abbreviatedtoREIT) is the returnof the equityREIT, whichin
2009 constitutes more than90%of the REITcapitalisation. The data is
takenfromthe National Associationof Real Estate Investment Trusts
(NAREIT). For the term spread (or term structure, SPR), CCL 2008
follows Estrella and Trubin (2006) by choosing the spread between
10-year Treasury bond yield and three-month Treasury bill rate,
and both are released by the Federal Reserve Board of Governors.
As for the three-month Treasury bill rate, since the constant matu-
rity rates are available only after 1982, the secondary-market three-
month rate expressed on a bond-equivalent basis is used. Estrella
and Trubin (2006) argue that this spread provides an accurate and
101
MODEL RISK
Table 5.2 Correlation coefcients (1975 Q22008 Q1)
FFR SPR HRET REIT
FFR 1.000
SPR 0.554 1.000
HRET 0.055 0.101 1.000
REIT 0.108 0.146 0.211 1.000
FFR, federal funds rate; SPR, interest rate spread; HRET, housing market
returns; REIT, equity REIT returns.
robust measure in predicting US real activity over long periods of
time.
In a sense, Table 5.2 veries a version of the long-run neutrality
of money with the asset market data (see, for example, King and
Watson 1994, 1997). The idea is simple. If the central bank cuts the
short-terminterest rate and continuously injects money into the eco-
nomic system, the market will expect an increase in the ination rate
in the future. To compensate for such ination risk, the long-term
bond would deliver a higher interest rate in equilibrium, resulting
in a higher term spread. In other words, the short-term interest rate,
represented by FFR in Table 5.2, is expected to be negatively corre-
lated with the term spread SPR. Table 5.2 also shows that the cor-
relation is far from perfect (with absolute value smaller than 0.6). It
indicates that there maybe factors affectingthe two interest rates dif-
ferently, or that the market may not be perfect. Interestingly, other
pairwise correlation coefcients are generally low. A more careful
investigation of the data will show that these variables are indeed
signicantly related, and now we will introduce the econometric
tools used in CCL 2008.
THE ECONOMETRIC SET-UP OF CCL 2008
The econometric model is simple. The structural form of the time-
varying vector autoregression model with lag length p for a process
y
t
is given by
A
0
y
t
= +A
1
y
t1
+A
2
y
t2
+ +A
p
y
tp
+u
t
(5.1)
where we allow for all parameters, including intercept coefcients,
autoregressional coefcients and the covariance matrix of stochas-
tic terms to be contingent on the unobservable state variable s
t
S
102
(to ease the burden of notation we suppress the state-dependent
subscripts). The vector autoregression model is chosen because
it imposes (relatively) fewer presumptions on the data structure,
and it also conveniently parameterises the dynamic interactions
within a system.
4
The time-varying coefcients capture possible
non-linearities or time variation in the lag structure of the model.
The stochastic volatility allows for possible heteroscedasticity of the
stochastic terms.
The variables of interest y
t
= (y
1,t
, y
2,t
, . . . , y
m,t
)
are anm1 vector.

The stochastic intercept term = (
1
(s
t
),
2
(s
t
), . . . ,
m
(s
t
))
cap-
tures the difference in the intercept under different states. A
0
is an
m m state-dependent matrix which measures the contemporane-
ous relationship between variables, and the econometric identica-
tion of the model is obtained through restrictions on A
0
. A
k
is an
mm matrix with state-dependent elements a
(ij)
k
(s
t
), i, j = 1, . . . , m,
k = 1, . . . , p. The stochastic error term u
t
will be explained below.
The corresponding reduced form of the above model can be
obtained by pre-multiplying Equation 5.1 by A
1
0
, which yields
y
t
= d +
1
y
t1
+
2
y
t2
+ +
p
y
tp
+
t
(5.2)
where d = A
1
0
,
k
= A
1
0
A
k
and
t
= A
1
0
u
t
, k = 1, 2, . . . , p.
k
is an
mm matrix with state-dependent elements
(ij)
k
(s
t
), i, j = 1, . . . , m,
k = 1, . . . , p. We further dene d(s
t
) c + (s
t
), which will be
explained below. The stochastic error-term vector
t
can be further
expressed as
t
= A
1
0
u
t
= (s
t
)H
1/2
v
t
(s
t
)
where H is an m m diagonal matrix with diagonal elements
2
j
,
j = 1, . . . , m, (s
t
) is anmmdiagonal matrixwithdiagonal elements
j
(s
t
), j = 1, . . . , m
(s
t
) =
_
_
_
_
_
_
_
1
(s
t
) 0 0
0
2
(s
t
) 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0
m
(s
t
)
_
_
_
_
_
_
_
which captures the difference in the intensity of volatility, and v
t
(s
t
)
is a vector of standard normal distribution, v
t
(s
t
) N(0, (s
t
)),
103
MODEL RISK
where the covariance matrix is given by
(s
t
) =
_
_
_
_
_
_
_
1 r
21
(s
t
) r
m1
(s
t
)
r
12
(s
t
) 1 r
m2
(s
t
)
.
.
.
.
.
.
.
.
.
.
.
.
r
1m
(s
t
) r
2m
(s
t
) 1
_
_
_
_
_
_
_
(5.3)
In CCL 2008, two cases are considered. One is a three-variate
time-varying SVAR(p) model, ie, m = 3. The three variables of
interest are y
t
= (FFR, SPR, RET)
, where FFR denotes the federal

funds rate, SPR is the interest rate spread and RET denotes either
REIT returns (REIT) or housing market returns (HRETs). Essentially,
they study the asset returns one at a time. They also compare
the case of the stock return based on the S&P 500 Index (SRET).
They then extend this to the case of the four-variate model, where
y
t
= (FFR, SPR, GDP, RET)
; GDP denotes GDP growth. It turns out

the results are pretty robust despite the introduction of the GDP
growth variable. It is consistent with the notion that asset prices are
forwardlookingandtherefore informationabout (past) GDPgrowth
has already been reected in the asset price data. Hence, the intro-
duction of the GDP variable will only marginally affect the SVAR of
asset returns. For econometric identication, restrictions on the ele-
ments of A
0
need to be imposed. Following the discussion in Leeper
et al (1996) and Christiano et al (1998, 1999), A
0
is specied to be a
lower triangular matrix. In the three-variable case, it means that
A
0
=
_
_
_
_
1 0 0
a
21
0
(s
t
) 1 0
a
22
0
(s
t
) a
23
0
(s
t
) 1
_
_
_
_
(5.4)
and the four-variable case is similar. As shown in Equation 5.4, we
have imposed a recursive restriction so that y
1,t
(FFR) affects y
2,t
(SPR), and both y
1,t
and y
2,t
affect y
3,t
(RET) contemporaneously, but
not vice versa. On the other hand, it is still possible for RET to affect
FFR and SPR, but with a time lag. Thus, the restriction may not be
as stringent as it seems.
Two-state Markov process
Following the literature on Markov switching, and being limited
by the sample size, we assume that there are only two states, ie,
s
t
S = 1, 2. The procedure for the identication of the regime of
104
theeconomyfor agivenperiodwill bediscussedbelow. TheMarkov-
switching process relates the probability that regime j prevails in t
to the prevailing regime i in t 1, Pr(s
t
= j s
t1
= i) = p
ij
. The
transition probability matrix is then given by
P =
_
p
11
1 p
11
1 p
22
p
22
_
The persistence can be measured by the duration 1/(1 p
ii
), and
hence the higher the value of p
ii
, the higher the level of persistence.
Given that the economy can be in either state 1 or state 2, the term
j
(s
t
), j = 1, . . . , m, dened above, captures the difference in the
intercept under different states. For convenience, we set
j
(1) = 0for
s
t
= 1; thus,
j
(2) measures the difference in the intercept between
state 2 and state 1. Furthermore, we set the diagonal element of
(s
t
) at state 1 to be unity, ie,
j
(1) = 1, so that if
j
(2) > 1, then the
intensity of volatility in state 2 is larger than that in state 1, and vice
versa.
Since v
t
(s
t
) is a vector of standard normal distribution and
j
(1)
is set to be 1, the variance of y
j,t
, j = 1, . . . , m, at state 1 is
2
j
, and the
variance is
2
j
(2)
2
j
.
Identication of regimes
Finally, we discuss the identication of regimes in this model. Since
the state of the economy is unobservable, we identify the regime for
givena time periodbyHamiltons (1989, 1994) smoothedprobability
approach, in which the probability of being state s
t
at time t is given
by (s
t

T
), where
T
= y
1
, y
2
, . . . , y
t
, . . . , y
T
. The idea is that we
identify the state of the economy from an ex post point of view, and
thus the full set of information is utilised. Notice that we only allow
for two regimes, ie, s
t
S = 1, 2. Thus, if (s
t
= j
T
) > 0. 5,
then we identify the economy most likely to be in state j, j = 1, 2.
Stationarity of the Markov regime-switching model
The stationarity test of Markov regime-switching model is provided
by Francq and Zakoian (2001). To illustrate the idea, take a three-
variable, VAR(2) model as an example. Let
(s
t
) =
_
_
_
_
1
(s
t
)
2
(s
t
) 0
3
I
3
0
3
0
3
0
3
0
3
0
3
_
_
_
_
105
MODEL RISK
Figure 5.2 Smoothed probabilities for the SVAR(1) model of (FFR,
SPR, GDP, REIT)
1.0
0.8
0.6
0.4
0.2
0
1.0
0.8
0.6
0.4
0.2
0
1980 1990 2000 1980 1990 2000
(a) (b)
(a) Regime 1; (b) regime 2.
SPR, GDP, HRET)
1.0
0.8
0.6
0.4
0.2
0
1980 1990 2000
(a)
1.0
0.8
0.6
0.4
0.2
0
1980 1990 2000
(b)
(a) Regime 1; (b) regime 2.
where I
3
is a 33 identity matrix, 0
3
is a 33 null matrix, and
1
(s
t
)
and
2
(s
t
) are the autoregression matrixes in Equation 5.2. We then
dene the following matrix
=
_
p
11
( (1) (1)) p
21
( (1) (1))
p
12
( (2) (2)) p
22
( (2) (2))
_
(5.5)
and let () be the spectral radius of . Francq and Zakoian (2001)
show that a sufcient condition for second-order stationarity of a
Markov-switching VAR(2) model is () < 1.
Empirical results
This sectionhighlights someof theempirical results of CCL2008. Fig-
ures 5.25.4 provide a visualisation of the identication of regimes
106
SPR, GDP, SRET)
1.0
0.8
0.6
0.4
0.2
0
1980 1990 2000
(a)
1.0
0.8
0.6
0.4
0.2
0
1980 1990 2000
(b)
Table 5.3 AIC values for various three-variable VAR(p) models of the
REIT system
VAR model State-contingent parameters P = 1
Single-regime model None 10.625
Two-regime model (A) c(s
t
), (s
t
) 9.951
Two-regime model (B) c(s
t
), (s
t
), v
t
(s
t
) 9.928
Two-regime model (C) c(s
t
), (s
t
),
k
(s
t
) 9.945
Two-regime model (D) c(s
t
), (s
t
),
k
(s
t
),v
t
(s
t
) 9.916
Note: The three variables are FFR, SPR and REIT.
Table 5.4 AIC values for various three-variable VAR(p) models of the
HRET system
VAR model P = 1 P = 2
Single-regime model 6.120 6.084
Model A (two-regime model) 5.403 5.059
Model B (two-regime model) 5.402 4.972
Model C (two-regime model) 6.330 4.907
Model D (two-regime model) 5.308 4.781
The three variables are FFR, SPR and HRET.
in the four-variable case. Regime 1 is always the (relatively) high-
volatility regime. It is clear that the identication of regimes is
similar and yet they are not exactly identical. It is consistent with
the notion that the asset returns tend to co-move and, at the same
time, there are instances or periods in which some asset returns are
107
MODEL RISK
Figure 5.5 Impulse responses of REIT to innovations in FFR when the
effect of SPR or GDP is shut off (FFR, SPR, GDP, REIT)
1.0
0.5
0
0.5
1.0
1.5
2.0
2.5
5 10 15 20 25 30
(a)
1.0
0.5
0
0.5
1.0
1.5
2.0
2.5
5 10 15 20 25 30
1.0
0.5
0
0.5
1.0
1.5
2.0
2.5
5 10 15 20 25 30
(b)
(c)
(a) Single regime; (b) regime 1; (c) regime 2. Solid line: full effect; dashed line:
SPR shut-off; dot-dashed line: GDP shut-off.
affected differently by the shocks. It should be noted that, given the
current formulation, there is an asset-specic innovation and hence
it is possible that the systemof different asset returns would identify
different periods. Tables 5.3 and 5.4 also conrm that the statistical
model D, which allows all the parameters (including those in the
variancecovariance matrix) to be regime dependent, performs best.
This conrms that regime switching is indeed an important feature
of both the REIT and housing return during the sampling periods.
Figures 5.55.7 summarise a series of counter-factual analyses.
Notice that in this SVARsystemthe monetary policy (FFR) can affect
the asset return in different ways. On top of the direct effect from
FFR to the asset return, it can also inuence the asset return through
the terms spread or the GDPgrowth (the indirect channel). To dis-
entangle the two effects, CCL 2008 rst report the impulse response
of the full effect, assuming that the system does not switch to an
alternative regime. In other words, all reported impulse responses
are conditional. They then shut down the effect from the term
108
Figure 5.6 Impulse responses of HRET to innovations in FFR when the
effect of SPR or GDP is shut off (FFR, SPR, GDP, HRET)
0.1
0
0.1
0.2
0.3
0.4
5 10 15 20 25 30
(a)
0.1
0
0.1
0.2
0.3
0.4
5 10 15 20 25 30
0.1
0
0.1
0.2
0.3
0.4
5 10 15 20 25 30
(b)
(c)
spread to the asset return, and allow only a once-and-for-all inno-
vation in the monetary policy variable (FFR). Hence, that second
impulse response only records the direct effect andthe indirect effect
through GDP. The third case is when the effect from GDP to the
asset return is shut down. Again, only a once-and-for-all innova-
tion is allowed in the monetary policy variable (FFR). Hence, that
third impulse response only records the direct effect and the indi-
rect effect through the term spread. Figure 5.5 reports the case of
REIT when a linear SVAR is estimated (single regime) and when a
regime-switching SVAR is estimated. In the latter, CCL 2008 further
distinguish the case when the system is under regime 1 from that
under regime 2. The result is clear. When the term spread channel
is shut down, the impulse response of REIT return dies out much
more quickly under the single regime and under regime 2. In other
words, the term spread acts like a multiplier in the case of REIT
return.
Figure 5.6 repeats the exercise for the housing return. Interest-
ingly, the case of full effect and that of GDP effect shut off are
109
MODEL RISK
Figure 5.7 Impulse responses of SRET to innovations in FFR when the
effect of SPR or GDP is shut off (FFR, SPR, GDP, SRET)
1
0
1
2
3
5 10 15 20 25 30
1
0
1
2
3
5 10 15 20 25 30
1
0
1
2
3
5 10 15 20 25 30
(a)
(b)
(c)
remarkablysimilar, suggestingthat the GDPplays a minor role inthe
propagation mechanism. On the other hand, under regime 1 (ie, the
high-volatility regime), the impulse response is much more volatile
when the term spread channel is shut down. In other words, the
term spread acts like a stabiliser in the case of housing return.
Figure 5.7 repeats the exercise for the stock return (based on the
S&P 500 Index). Interestingly, the case when the GDP growth is shut
down and the case when the termspread is shut down provide very
similar pictures to the case with the full effect, suggesting the very
minor role of both for the monetary policy variable to affect the stock
return. After all, the effect of the monetary policy on the stock return
seems to be very small and short lived in all cases.
While some of the results in CCL 2008 are not discussed here due
to limited space, there are still a few lessons we can learn from their
exercises. First, the regime-switching nature seems to be very impor-
tant in the data, especially for the asset return. Second, GDP does
110
not seem to play an important role in the propagation of the mone-
tary policy to the asset market returns. Third, while the term spread
seems to play an important role in the propagation of the monetary
policy to the asset market returns, such a role seems to be asset spe-
cic and regime dependent. One possible explanation is that, when
the central bank cuts the (short-term) interest rate, the market par-
ticipants anticipate a possible change in the long-term interest rate
and hence the long-term interest rate tends to increase. The house-
hold sector seems to be very aware of the long-term interest rate
increase, while the corporate sector seems to focus on the benet
of the short-term interest rate, possibly reecting the difference in
source of nancing. Thus, the policy makers may wish to be mindful
of such a difference in their policy operation.
FORECASTING: THE FOCUS OF CCL 2009
While CCL 2008 takes an initial step towards exploring the effect of
the monetary policy in a system-dynamics context, there are several
issues to be addressed. First, CCL2008 study the asset returns one-
at-a-time and hence this precludes the possibility of interactions
among different assets. Second, CCL 2008 did not touch on the fore-
casting issue, which is one of the major concerns of both the public
and the policy makers. Third, CCL 2008 consider only the condi-
tional impulse response and hence ignore the uncertainty generated
by the possible stochastic regime switching along the path of adjust-
ment. In light of these shortcomings, CCL 2009 takes a preliminary
step in attempting to address some of these issues.
Econometrically speaking, CCL 2009 also employs a regime-
switching structural VAR (RSSVAR) model, as in the case in CCL
2008. However, they make several important changes. First, they
focus on the housing return and stock return and include both vari-
ables in all the models they estimate. Second, they include a much
longer list of variables and estimate many different versions of those
RSSVAR models. Third, after the estimations of all those models,
they use the calculated probabilities of regime switching for evalu-
ating the forecasting performances of house and stock prices across
various models, and then examine both in-sample (1975 Q12005
Q4) andout-of-sample (2006 Q12008 Q3) forecasting performances.
2005 Q4 is naturally chosenas the cut-off point because the rise inthe
house price growth rate starting in the 1990s peaked around the end
111
MODEL RISK
of 2005. As will become clear, CCL 2009 actually allow the models
to learn, ie, they update the data starting at 2006 Q1 and see how
these models performwhen the growth of US house prices began to
decline and, consequently, the subprime crisis unfolded.
While detailed discussion of the literature is beyond the scope of
this chapter, a few points are worth attention.
5
First, CCL use multi-variate regime-switching SVAR models,
while manyexistingstudies onforecastingeither use a single-variate
(ie, the variable to be forecasted) model or employ linear VAR (ie,
single regime) models. The former approach suffers from an endo-
geneity problem (see Sims (1980b) for a detailed discussion), while
the latter implicitly rules out the possibility of regime switching (see,
for example, Hamilton 1994).
Second, studies on the possible regime-switching nature of the US
monetary policy typically adopt the Bayesian econometric method
(Sargent et al 2006; Sims 1980a,b; Sims and Zha 2006). In order for
this book to complement the literature, it may be a merit for this
chapter to adopt the classical econometric method.
Third, CCL2009 conduct out-of-sample forecasting using two dif-
ferent approaches: conditional expectations and simulation-based
methods. While the former approach is easier to implement, it does
not really track the regime that occurs for each forecasting step, and
the condence intervals are not available. Following Sargent et al
(2006), CCL 2009 adopt the simulation-based approach to calculate
the median path and the condence interval. CCL provide more
discussion on this issue.
Fourth, CCL 2009 perform the forecasting in a system-dynamics
manner and hence can deliver its prediction on both housing and
stock returns simultaneously. In fact, from the investors point of
view, since the returns of the two assets are imperfectly correlated, it
is natural for agents toinclude bothassets under some dynamic port-
folio consideration (see, for example, Yao and Zhang 2005; Leung
2007). Moreover, some recent works identify channels in which
the housing markets and stock returns are closely related (Lustig
and Van Nieuwerburgh 2005; Piazzesi et al 2007). Sutton (2002)
presents evidence that a signicant part of house price uctuations
can be explained by stock prices in six countries (US, UK, Canada,
Ireland, the Netherlands and Australia). A study by the Bank for
International Settlements (2003) also shows that, for a large group
112
Table 5.5 Statistical summary of federal funds rate, term spread, gross
domestic production growth rate, external nance premium, market
liquidity, stock index return and housing market return (1975 Q22008
Q3)
FFR SPR GDP EFP TED SRET HRET
Mean 6.397 1.502 0.759 1.087 0.883 1.968 1.344
Median 5.563 1.604 0.731 0.957 0.637 2.263 1.313
Max. 17.780 3.611 3.865 2.513 3.307 18.952 4.511
Min. 0.997 2.182 2.038 0.560 0.097 26.431 2.713
Std. Dev. 3.508 1.335 0.750 0.422 0.742 7.659 1.040
Skewness 1.037 0.627 0.127 1.220 1.552 0.664 0.040
Kurtosis 4.283 2.941 6.150 4.229 4.917 4.070 4.691
Oberv. 134.000 134.000134.000134.000134.000 134.000 134.000
FFR, federal funds rate; SPR, interest rate spread; GDP, gross domes-
tic production growth rate; EFP, external nance premium; TED, market
liquidity; SRET, stock index return; HRET, housing market return.
of countries, house prices tend to follow the stock market with a
two- to three-year lag. Kakes and End (2004) nd that stock prices in
the Netherlands signicantly affect house prices. On the other hand,
Lustig and Van Nieuwerburgh (2005) nd that the US housing col-
lateral ratio predicts aggregate stock returns and investors seem to
demand a larger risk compensation in times when the housing col-
lateral ratio is low. Yoshida (2008) nds that the housing component
serves as a risk factor in the pricing kernel of equities and this miti-
gates the equity premiumpuzzle and the risk-free rate puzzle. Thus,
it would be important to take into account the interactions of stock
returns and housing returns by studying them at the same time.
Data
It may be instructive to quickly review the data here. Table 5.5 sum-
marises the data series that are used in CCL 2009. They include the
(three-month) FFR, which are a measure of the US monetary pol-
icy, the SPR, which are the discrepancy between the long-term (10-
year) interest rate and the short-term(three-month) counterpart, the
external nance premium (EFP), which is equal to corporate bond
spread (Baa-Aaa), the TED spread which is the difference between
the three-month Eurodollar deposit rate and the three-month Trea-
sury bill rate, growth rates of GDP, the SRETandthe HRET, covering
the period 1975 Q12008 Q3.
6
All these variables are widely used in
113
MODEL RISK
Table 5.6 Correlation coefcients (1975 Q22008 Q3)
FFR SPR GDP EFP TED SRET HRET
FFR 1.000 0.557 0.104 0.544 0.833 0.009 0.015
SPR 1.000 0.145 0.037 0.437 0.021 0.115
GDP 1.000 0.179 0.165 0.030 0.111
EFP 1.000 0.650 0.057 0.151
TED 1.000 0.049 0.076
SRET 1.000 0.055
HRET 1.000
Numbers in bold denote best t models.
Table 5.7 List of models
Model Model structure Variables
A Linear FFR, SPR, TED, EFP, GDP, SRET, HRET
B Two-regime FFR, GDP, SRET, HRET
C Two-regime FFR, SPR, SRET, HRET
D Two-regime FFR, EFP, SRET, HRET
E Two-regime FFR, TED, SRET, HRET
F Two-regime EFP, SPR, SRET, HRET
G Two-regime EFP, TED, SRET, HRET
H Two-regime SPR, TED, SRET, HRET
Key: (unless specied, all variables refer to quarterly data) FFR, Fed-
eral Fund Rate; SPR, term spread, which is equal to the 10-year bond
rate minus FFR; TED spread, which is equal to the interbank rate minus
the T-bill rate, a measure of market liquidity; EFP, external nance pre-
mium, which is equal to the lending rate minus the T-bill rate, a measure of
EFP; GDP denotes GDP growth rate; SRET, stock market return; HRET,
housing market return.
the literature.
7
A comparison of their forecasting ability in a unify-
ing framework, with emphasis on the interactive (through the use
of VAR) and the regime-switching nature, seems to be a novelty of
CCL 2009. Figure 5.6 shows that several of these variables seem to
be signicantly correlated. For instance, the correlationbetweenFFR
and TED is above 0.8 and that between TED and EFP is above 0.6.
Thus, it seems appropriate to study their movement in a VAR-type
model rather thana single-equationcontext. Inaddition, while some
of those variables are correlated, it is still possible to compare (and
rank) their forecasting performance formally.
114
Table 5.7 shows clearly how models are constructed in a way
that would facilitate the comparison. For instance, models A to D
would have FFR involved, which can highlight the potential role of
monetary policy in the asset return dynamics. Models F to H dif-
fer from the previous ones as the monetary policy variable FFR is
removed. Instead, an additional nancial market variable is intro-
duced to the system. Thus, model F can be interpreted as model C
with FFRreplacedby EFP, model Gas model Ewith FFRreplacedby
EFP, and model H as model E with FFR replaced by SPR. As it will
become clear, in despite of all these similarities, models with only
one variable difference may have a very different performance in
forecasting.
Econometric procedures
CCL 2009 conduct out-of-sample forecasting starting 2006 Q1, and
thus we divide the sample into an in-sample period (1975 Q2
2005 Q4) and an out-of-sample period (2006 Q12008 Q3). We then
proceed with out-of-sample forecasting in two different approaches.
The rst approach is the conventional conditional moment meth-
od. Giventhe estimationwindow1975 Q22005 Q4 anda forecasting
horizon h = 1, . . . , 4, the estimated parameters are used to forecast
house and stock prices h-steps ahead outside the estimation win-
dow, using the smoothed transition probabilities. The h-step-ahead
forecasted value of z
t+h
based on information at time t,
t
, is given
by
E(z
t+h

t
) =
2
_
i=1
E[z
t+h
s
t+h
= i,
t
] p(s
t+h
= i
t
)
where z
t
y
t
. The estimation window is then updated one obser-
vation at a time and the parameters are re-estimated. Again the h-
step-ahead forecasts of house and stock prices are computed out-
side the new estimation window. The procedure is iterated up to
the nal observation, 2008 Q3. The forecasts based on this method
basically compute the h-step-ahead conditional expectations of the
variable beingpredicted. Most existing(non-Bayesian) works follow
this method.
The secondapproach is the simulation method. The idea is simply
that the path of the forecasted values is obtained by simulating the
model repeatedly. The procedure is as follows.
115
MODEL RISK
Step 1. The RSSVAR model is estimated from the sampling
period 1975 Q22005 Q4, and the parameters, transition prob-
abilities and variancecovariance matrix, etc, are obtained.
Given the estimation results, the smoothed probabilities for
identifying the regime at the period 2005 Q4 are computed.
Step 2. Given the regime at 2005 Q4, the path of h-step-ahead
regimes by random drawing is simulated (h = 1, . . . , 4).
8
Given this particular path of h-step-ahead regimes, the path
of predicted values of z
t
y
t
is obtained from Equation 5.2.
Step3. Steps 1and2are repeated50,001times inorder toobtain
the median of the h-step-ahead forecasted values during 2006
Q12006 Q4 and their corresponding condence intervals.
Step 4 The sampling period is then updated with four more
observations (ie, data for another year) and steps 13 are
repeated to simulate the path of predicted values for the next
four quarters.
This procedure is repeated up to the end of our sample.
An advantage of this method over the computation of the mean of
possible future values in the rst approach is that this method takes
full account of the regime-switching model by determining the path
of future regimes using random drawing, rather than simply taking
expectations over transition probabilities. Another advantage is that
a condence interval with which to evaluate its forecasting perfor-
mances is generated naturally. It should be noted that the regime-
switching nature of the model implies that the future forecast is path
dependent and hence the conventional way to construct condence
interval becomes invalid.
To evaluate the performances of in-sample and out-of-sample
forecasts, we compute two widely used measures for forecasting
a variable z
t
y
t
: root-mean-square errors (RMSEs) and mean
absolute errors (MAEs), which are dened respectively by
RMSE(h) =
_
1
T h
Th
_
t=1
(z
t+h
z
t+ht
)
2
_
1/2
MAE(h) =
1
T h
Th
_
t=1
z
t+h
z
t+ht
where z
t+ht
E(z
t+h

t
).
116
Table 5.8 A summary of goodness of t for all eight models
Model Models AIC
A Single-regime model (FFR, SPR, TED, 11.230
EFP, GDP, SRET,HRET)
B Two-regime (FFR, GDP, SRET, HRET) 13.472
C Two-regime (FFR, SPR, SRET, HRET) 12.450
D Two-regime (FFR, EFP, SRET, HRET) 10.159
E Two-regime (FFR, TED, SRET, HRET) 11.134
F Two-regime (EFP, SPR, SRET, HRET) 9.747
G Two-regime (EFP, TED, SRET, HRET) 8.404
H Two-regime (SPR, TED, SRET, HRET) 11.274
Estimation results
As in the case of CCL 2008, we can only highlight some of the
ndings of CCL 2009. Table 5.8 provides a summary of the esti-
mation results. In general, a model allowing for regime switching
attains a lower value of Akaikes information criterion (AIC) and a
higher log-likelihood value. Among all these models, the regime-
switching model (EFP, TED, SRET, HRET) has the best t, ie, a sig-
nicantly lower value of AIC than other models, suggesting that
the credit market frictions and asset returns are indeed signicantly
inter-related.
CCL 2009 identify regime 1 as the high-volatility regime and
regime 2 as the low volatility regime. And, as in the case of CCL
2008, both regimes tend to be highly persistent. For instance, the
transition probability matrix of the model (EFP, TED, SRET, HRET)
is given by
P =
_
p
11
p
12
p
21
p
22
_
=
_
0. 854 0. 146
0. 068 0. 932
_
whichsuggests that the expecteddurationof regime 1 is 1/(1p
11
) =
6. 8 quarters and the expected duration of regime 2 is 1/(1 p
22
) =
14. 7 quarters.
In-sample forecasting
Table 5.9 summarises the in-sample forecasts of asset returns. For the
stock returns, model C (FFR, SPR, SRET, HRET) has the best perfor-
mance. For housing return, however, it is model D (FFR, EFP, SRET,
117
MODEL RISK
Table 5.9 A summary of in-sample forecasting performance
(four-quarter-ahead forecasts)
Stock returns Housing returns
, .. , ..
Model RMSE MAE RMSE MAE
A Single-regime model 7.5842 5.6699 0.8226 0.6499
(FFR, SPR, TED, EFP,
GDP, SRET,HRET)
B Two-regime 7.6411 5.6640 0.8286 0.6508
(FFR, GDP, SRET, HRET)
C Two-regime 7.5103 5.5922 0.7974 0.6361
(FFR, SPR, SRET, HRET)
D Two-regime 7.6460 5.6561 0.7801 0.6129
(FFR, EFP, SRET, HRET)
E Two-regime 7.6232 5.6959 0.7984 0.6207
(FFR, TED, SRET, HRET)
F Two-regime 7.7767 5.7204 0.7940 0.6331
(EFP, SPR, SRET, HRET)
G Two-regime 7.7917 5.8092 0.8397 0.6468
(EFP, TED, SRET, HRET)
H Two-regime 7.6169 5.7064 0.8161 0.6313
(SPR, TED, SRET, HRET)
HRET) that out-performs all others. The results hold regardless of
whether the RMSE or MAE is used as the criterion. Notice that both
models contain the monetary policy variable FFR. It is interesting to
see that neither the linear model with seven variables nor model B,
which contains GDP growth, give the best performance. In other
words, the commonly used approach of using a linear VAR model
withas manyvariables as possible maynot be advised. Instead, care-
ful selection of variables with the regime-switching consideration
may be very important in understanding the reality.
Out-of-sample forecasting
We nowturn to the out-of-sample forecasts results based on the con-
ditional mean on housing and stock returns in CCL 2009. Table 5.10
summarises the results. In general, the regime-switching model H
(SPR, TED, SRET, HRET) performs very well. In terms of forecasting
stock returns, it out-performs all other models under the criteria of
RMSEs. Under the criteria of MAEs, it is extremely close to model A,
118
Table 5.10 A summary of out-of-sample forecasting performance
(four-quarter-ahead forecasts)
, .. , ..
A Single-regime model 7.9841 5.6808 2.1292 1.8424
(FFR, SPR, TED,
EFP, GDP, SRET,HRET)
B Two-regime 7.2027 5.8760 2.1303 1.8739
C Two-regime 7.3392 6.0156 1.9161 1.7198
D Two-regime 7.3122 5.9867 1.9977 1.7797
E Two-regime 7.0037 5.7126 2.0761 1.7754
F Two-regime 8.2423 6.7808 1.8184 1.6078
G Two-regime 7.2071 5.7972 2.0430 1.7617
H Two-regime 6.9225 5.6933 1.8284 1.5201
which is the top performer. In terms of forecasting housing returns,
it out-performs all other models under the criteria of MAEs. Under
the criteria of RMSEs, it is extremely close to model F, which is the
top performer. Notice that, unlike the case for in-sample forecasting,
model Hdoes not contain either the monetary policy variable FFRor
the fundamental variable GDP. It seems tosuggest that the in-sample
and out-of-sample forecasting powers can be quite different.
We next turn to the simulation-based forecasting results in CCL
2009. They consider a forecasting window of four quarters start-
ing in 2006 Q1, with h-quarter-ahead forecasts, h = 1, . . . , 4. After
simulating the out-of-sample path 2006 Q12006 Q4 based on obser-
vations up to 2005 Q4, the data is updated with four observations
and the parameters are re-estimated. The procedure is iterated up
to the nal observation, 2008 Q3. The purpose of this exercise is to
see how the performances of the models change when information
is updated. The simulated paths together with their 80% condence
119
MODEL RISK
Figure 5.8 Simulation-based out-of-sample forecasts of stock returns
with 80% CI from 2006 Q1 to 2006 Q4 based on information available at
2005 Q4
15
10
5
0
5
10
15
05Q2 05Q4 06Q2 06Q4
15
10
5
0
5
10
15
05Q2 05Q4 06Q2 06Q4
15
10
5
0
5
10
15
05Q2 05Q4 06Q2 06Q4
15
10
5
0
5
10
15
05Q2 05Q4 06Q2 06Q4
15
10
5
0
5
10
15
05Q2 05Q4 06Q2 06Q4
Data
Model A
CI for Model A
Data
Model B
CI for Model B
.
.
Model C
CI for Model C
Data
Model H
CI for Model H
Data
Model D
CI for Model D
.
.
Model E
CI for Model E
Data
Model F
CI for Model F
.
.
Model G
CI for Model G
Model A: single-regime (FFR, SPR, TED, EFP, GDP, SRET, HRET); Model B: two-
regime (FFR, GDP, SRET, HRET); Model C: two-regime (FFR, SPR, SRET, HRET);
Model D: two-regime (FFR, EFP, SRET, HRET); Model E: two-regime (FFR, TED,
SRET, HRET); Model F: two-regime (EFP, SPR, SRET, HRET); Model G: two-
regime (EFP, TED, SRET, HRET); Model H: two-regime (SPR, TED, SRET, HRET).
intervals can be visualised in Figures 5.85.10 for stock returns and
in Figures 5.115.13 for housing returns. Tables 5.11 and5.12 provide
a summary of the performances of different models.
For the predictions of stock returns, the predictedpaths of the rst
two forecasting windows (Figures 5.8 and 5.9) and actual data are
well withinthe boundaries of the 80%condence intervals for all ve
models. Inasense, althoughthe models didnot predict what actually
120
2006 Q4
15
10
5
0
5
10
15
06Q2 06Q4 07Q2 07Q4
15
10
5
0
5
10
15
06Q2 06Q4 07Q2 07Q4
15
10
5
0
5
10
15
06Q2 06Q4 07Q2 07Q4
15
10
5
0
5
10
15
06Q2 06Q4 07Q2 07Q4
15
10
5
0
5
10
15
06Q2 06Q4 07Q2 07Q4
Data
Model B
CI for Model B
.
.
Model C
CI for Model C
Data
Model D
CI for Model D
.
.
Model E
CI for Model E
Data
Model F
CI for Model F
.
.
Model G
CI for Model G
Data
Model A
CI for Model A
Data
Model H
CI for Model H
happened in 2006 and 2007, the models predictions are not that far
off the mark. But the last window (2008 Q12008 Q3 in Figure 5.8
performs much worse: except for the regime-switching model H
(SPR, TED, SRET, HRET), all models have at least one period (ie, a
quarter) which lies outside the condence region. Notice that, for
overall out-of-sample forecasting performance (four-quarter-ahead
forecasts), it is also model H that performs best overall. While more
121
MODEL RISK
2007 Q4
Data
Model A
CI for Model A
15
10
5
0
5
10
15
20
07Q1 07Q3 08Q1 08Q3
15
10
5
0
5
10
15
20
07Q1 07Q3 08Q1 08Q3
15
10
5
0
5
10
15
20
07Q1 07Q3 08Q1 08Q3
15
10
5
0
5
10
15
20
07Q1 07Q3 08Q1 08Q3
15
10
5
0
5
10
15
20
07Q1 07Q3 08Q1 08Q3
Data
Model B
CI for Model B
.
.
Model C
CI for Model C
Data
Model D
CI for Model D
.
.
Model E
CI for Model E
Data
Model F
CI for Model F
.
.
Model G
CI for Model G
Data
Model A
CI for Model A
research is clearly needed, the results here seem to suggest that
the interest rate spread (SPR) and the TED spread are indeed very
important inthe forecastingof stockreturns, eveninanancial crisis.
For the predictions of housing returns, the forecasting perfor-
mances of all the models in a sense deteriorate much faster than
the predictions for stock returns. Figure 5.11 shows that most models
basically capture the downward trend of the housing return in 2006
122
Figure 5.11 Simulation-based out-of-sample forecasts of housing
returns with 80% CI from 2006 Q1 to 2006 Q4 based on information
available at 2005 Q4
Data
Model A
CI for Model A
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0
05Q2 05Q4 06Q2 06Q4
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0
05Q2 05Q4 06Q2 06Q4
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0
05Q2 05Q4 06Q2 06Q4
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0
05Q2 05Q4 06Q2 06Q4
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0
05Q2 05Q4 06Q2 06Q4
Data
Model B
CI for Model B
.
.
Model C
CI for Model C
Data
Model D
CI for Model D
.
.
Model E
CI for Model E
Data
Model F
CI for Model F
.
.
Model G
CI for Model G
Data
Model H
CI for Model H
within their 80% condence intervals, although model A (the lin-
ear model with all seven variables) and model D (FFR, SPR, SRET,
HRET) are not totally successful even for the forecasting of 2006.
Unfortunately, Figure 5.12seems tosuggest that the models are mis-
led bythe boundback of housingreturnin2006 Q4, whichresults
inbasicallyat predictionsfor the 2007returns. The realityis much
worse, and hence the data for 2007 is basically outside the con-
dence intervals of all models, except for 2007 Q1. Figure 5.13 shows
123
MODEL RISK
Data
Model A
CI for Model A
2.8
2.4
2.0
1.6
1.2
0.8
0.4
0
0.4
0.8
06Q2 06Q4 06Q2 06Q4
2.8
2.4
2.0
1.6
1.2
0.8
0.4
0
0.4
0.8
06Q2 06Q4 06Q2 06Q4
2.8
2.4
2.0
1.6
1.2
0.8
0.4
0
0.4
0.8
06Q2 06Q4 06Q2 06Q4
2.8
2.4
2.0
1.6
1.2
0.8
0.4
0
0.4
0.8
06Q2 06Q4 06Q2 06Q4
2.8
2.4
2.0
1.6
1.2
0.8
0.4
0
0.4
0.8
06Q2 06Q4 06Q2 06Q4
Data
Model B
CI for Model B
.
.
Model C
CI for Model C
Data
Model D
CI for Model D
.
.
Model E
CI for Model E
Data
Model F
CI for Model F
.
.
Model G
CI for Model G
Data
Model H
CI for Model H
that there is another bound back of housing return in 2007 Q4.
This time, all the models even predict that the housing returns will
increase and the condence intervals are increasing in value over
time. The reality again disappoints. As a result, for the forecasting
window 2008 Q12008 Q3, the data lie completely outside the con-
dence interval. In other words, all models fail, as summarised by
Table 5.12.
124
Data
Model H
CI for Model H
4
3
2
1
0
1
2
3
07Q1 07Q3 08Q1 08Q3
4
3
2
1
0
1
2
3
07Q1 07Q3 08Q1 08Q3
4
3
2
1
0
1
2
3
07Q1 07Q3 08Q1 08Q3
4
3
2
1
0
1
2
3
07Q1 07Q3 08Q1 08Q3
4
3
2
1
0
1
2
3
07Q1 07Q3 08Q1 08Q3
Data
Model A
CI for Model A
Data
Model B
CI for Model B
.
.
Model C
CI for Model C
Data
Model D
CI for Model D
.
.
Model E
CI for Model E
FORECASTINGTWO ASSET RETURNS: IS IT NECESSARY?
We may wonder is this system-dynamics approach necessary?
Does this approach perform better? While a throughout compari-
son of different approaches of forecasting asset returns is beyond
the scope of this chapter, we partly address the issue by conduct-
ing the following exercise. Recall that we put both the stock return
and the housing return in the same RSSVAR model in all cases. We
125
MODEL RISK
Table 5.11 Is the forecasted stock return within the 80% condence
interval?
Predicting Predicting Predicting
2006 2007 2008
based on based on based on
Models 19752005 19752006 19752007
A Single-regime (FFR, SPR, TED, Yes Yes Partly
EFP, GDP, SRET, HRET)
B Two-regime Yes Yes Partly
C Two-regime Yes Yes Partly
D Two-regime Yes Yes Partly
E Two-regime Yes Yes Partly
F Two-regime Yes Yes Partly
G Two-regime Yes Yes Partly
H Two-regime Yes Yes Yes
now ask: Is the forecasting of stock return better in the presence
of housing return? Is the forecasting of the housing return better in
the presence of stock return? To answer these two questions, we
simply remove the asset returns, one at a time, and then compare
the performance of those one-asset-return models with the original
two-asset-return models. As in the case of CCL 2009, we use the cri-
teria of MSAEs and MAEs. To impose some self-discipline in this
comparison, we require that a model performs better only when
it beats the alternative in both MSAE and MAE criteria. Tables 5.13
and 5.14 provide a summary of the results and Tables 5.15 and 5.16
provide more details.
Interestingly, in terms of in-sample forecasting, the presence of
the housing return does not lead to an unambiguous improvement
over the alternatives. However, in terms of out-of-sample forecast-
ing, the presence of the housing return does help to predict the stock
return in the case of model E (ie, FFR, TED, SRET, HRET) and in the
case of model H (ie, SPR, TED, SRET, HRET). The case for housing
return is even more encouraging. In terms of both in-sample and
126
Table 5.12 Is the forecasted housing return within the 80% condence
interval?
Predicting Predicting Predicting
2006 2007 2008
based on based on based on
Models 19752005 19752006 19752007
A Single-regime Partly Partly No
(FFR, SPR, TED,
EFP, GDP, SRET, HRET)
B Two-regime Yes Partly No
C Two-regime Yes Partly No
D Two-regime Partly Partly No
E Two-regime Yes Partly No
F Two-regime Yes Partly No
G Two-regime Yes Partly No
H Two-regime Yes Partly No
Table 5.13 Do models forecast stock return better in the presence of
housing return?
In-sample Out-of-sample
B predicts stock return better than B1 No No
C predicts stock return better than C1 No No
D predicts stock return better than D1 No No
E predicts stock return better than E1 No Yes
F predicts stock return better than F1 No No
G predicts stock return better than G1 No No
H predicts stock return better than H1 No Yes
out-of-sample forecasting, the presence of stock return does lead
to a clear improvement in the prediction of housing return in the
case of model C (ie, FFR, SPR, SRET, HRET), model E (ie, FFR, TED,
SRET, HRET) and model H (ie, SPR, TED, SRET, HRET). Recall that
model Cis the best model in the in-sample forecasting, and model H
127
MODEL RISK
Table 5.14 Do models forecast housing return better in the presence of
stock return?
In-sample Out-of-sample
B predicts housing return better than B2 No No
C predicts housing return better than C2 Yes Yes
D predicts housing return better than D2 No No
E predicts housing return better than E2 Yes Yes
F predicts housing return better than F2 No No
G predicts housing return better than G2 No No
H predicts housing return better than H2 Yes Yes
Yes means the model is better in both RMSE and MAE criteria.
is the best in out-of-sample forecasting. In other words, among the
winner models (ie, the models with the best forecasting ability),
the presence of the other asset return does play a positive role. These
results are consistent with recent theoretical works that emphasise
the interactions of the two asset returns in a dynamic general equi-
librium context (see, for example, Leung 2007; Jaccard 2007, 2008;
Leung and Teo 2008, 2009). They also provide support to the gen-
eral philosophy of CCL (2008 and 2009) that it is indeed important
to understand the asset return dynamics in a system-dynamics
context.
9
CONCLUDING REMARKS
The public has very serious concerns about the dramatic movements
in asset prices andthe advent of nancial crises. They expect the eco-
nomics and nance profession to provide some formof early warn-
ing system. Yet forecasting asset prices and returns is always dif-
cult, especially at a time of nancial crisis. As Sanders (2008, p. 261)
admits, The sudden paradigm shift in 2005 and 2006 demonstrates
that markets can change dramatically and the most sophisticated
models can be taken by surprise. This chapter reviews some recent
efforts, on both the theoretical and empirical fronts, which may shed
light onour understandingand, hopefully, helpus tobuildthe early
warning system demanded by society.
The following lessons have emerged fromthis review. It is indeed
important to explicitly model the asset returns in the existing frame-
work (both theoretically and empirically). There are important, and
128
Table 5.15 A summary of in-sample forecasting performances
(four-quarter ahead forecasts)
, .. , ..
B (FFR, GDP, SRET, HRET) 7.6411 5.6640 0.8286 0.6508
B1 (FFR, GDP, SRET) 7.7030 5.6441
B2 (FFR, GDP, HRET) 0.8205 0.6444
C (FFR, SPR, SRET, HRET) 7.5103 5.5922 0.7974 0.6361
C1 (FFR, SPR, SRET) 7.5674 5.5665
C2 (FFR, SPR,HRET) 0.8126 0.6488
D (FFR, EFP, SRET, HRET) 7.6460 5.6561 0.7801 0.6129
D1 (FFR, EFP, SRET) 7.6261 5.6389
D2 (FFR, EFP, HRET) 0.7725 0.6068
E (FFR, TED, SRET, HRET) 7.6232 5.6959 0.7984 0.6207
E1 (FFR, TED, SRET) 7.6309 5.6751
E2 (FFR, TED, HRET) 0.8004 0.6237
F (EFP, SPR, SRET, HRET) 7.7767 5.7204 0.7940 0.6331
F1 (EFP, SPR, SRET) 7.7579 5.7245
F2 (EFP, SPR, HRET) 0.7976 0.6272
G (EFP, TED, SRET, HRET) 7.7917 5.8092 0.8397 0.6468
G1 (EFP, TED, SRET) 7.7226 5.7090
G2 (EFP, TED, HRET) 0.8387 0.6441
H (SPR, TED, SRET, HRET) 7.6169 5.7064 0.8161 0.6313
H1 (SPR, TED, SRET) 7.6332 5.6946
H2 (SPR, TED, HRET) 0.8263 0.6479
All models are two-regime models.
empirically detectable, interactions among different asset returns.
Some recent empirical works seem to suggest that the regime-
switching nature of the data can no longer be denied. Whether it
is for in-sample or out-of-sample forecasting, before or during the
time of nancial crisis, the explicit consideration of the stochastic
regime switch is important. It has been shown that the failure to
recognise such a feature of the data will lead to signicant deteriora-
tion of the model performance. There are some signicant portions
of the asset price movements not captured by the existing models.
Thus, more work needs to be done. Among many options, one pos-
sibility is to compare the experiences of other countries during the
129
MODEL RISK
Table 5.16 A summary of out-of-sample forecasting performances
(four-quarter ahead forecasts)
, .. , ..
B (FFR, GDP, SRET, HRET) 7.2027 5.8760 2.1303 1.8739
B1 (FFR, GDP, SRET) 7.1082 5.8411
B2 (FFR, GDP, HRET) 2.1220 1.8478
C (FFR, SPR, SRET, HRET) 7.3392 6.0156 1.9161 1.7198
C1 (FFR, SPR, SRET) 7.1075 5.8296
C2 (FFR, SPR, HRET) 1.9379 1.7388
D (FFR, EFP, SRET, HRET) 7.3122 5.9867 1.9977 1.7797
D1 (FFR, EFP, SRET) 7.1105 5.8246
D2 (FFR, EFP, HRET) 1.9793 1.7648
E (FFR, TED, SRET, HRET) 7.0037 5.7126 2.0761 1.7754
E1 (FFR, TED, SRET) 7.2950 5.9108
E2 (FFR, TED, HRET) 2.0820 1.7832
F (EFP, SPR, SRET, HRET) 8.2423 6.7808 1.8184 1.6078
F1 (EFP, SPR, SRET) 7.9820 6.5703
F2 (EFP, SPR, HRET) 1.7998 1.5891
G (EFP, TED, SRET, HRET) 7.2071 5.7972 2.0430 1.7617
G1 (EFP, TED, SRET) 7.1682 5.7683
G2 (EFP, TED, HRET) 2.0199 1.7586
H (SPR, TED, SRET, HRET) 6.9225 5.6933 1.8284 1.5201
H1 (SPR, TED, SRET) 7.1912 5.8319
H2 (SPR, TED, HRET) 1.8508 1.5642
All models are two-regime models.
20089 crisis and previous nancial crises. And this is a direction
being pursued.
The authors are grateful to Jun Cai, Wing Hong Chan, Andrew
Filrado, Mico Loretan, Douglas Rolph, Wing Leong Teo, Xueping
Wu, seminar participants at the Asian Real Estate Society and
ASSA-AREUEAmeetings, BIS and City University of Hong Kong
for many useful comments and suggestions, and the National
Science Council in Taiwan and the RGC Earmark Grant (No.
9041515) in Hong Kong for nancial support. The usual disclaimer
applies.
130
1 The Japanese asset price bubbles collapse from 1991 to 2000.
2 See, for example, Leung (2007) for a dynamic general equilibrium analysis.
3 Fat-tailed properties in asset-return distributions have been documented repeatedly in the
literature; see, for example, Rachev et al (2005) for a survey.
4 See, for example, Sims (1980) for more discussion on these issues and the potential biases that
could be eliminated by the VAR method.
5 See, for example, CCL 2009 for a (partial) review of the literature.
6 Notice that we use asset returns rather than asset prices in the empirical investigation.
The reason is clear. Asset prices are typically non-stationary, while asset returns may
exhibit mean-reversion.
7 See CCL 2009 for more detailed justication and the related literature.
8 For example, suppose the regime identied at the time 2005 Q5 is state 1; we use the transition
probabilities p
11
and p
12
to generate the state at the period 2006 Q1. Specically, we draw a
value v froma uniformdistribution U[0, 1]. The state at 2006 Q1 is state 1 if v (0, p
11
), and is
state 2 otherwise. Suppose that we have identied the state at 2006 Q1 as state 2. Then we use
the transition probabilities p
21
and p
22
to generate the state at the period 2006 Q2. Therefore,
we will be able to simulate the path of h-step-ahead regimes.
9 Notice that while this chapter focuses the discussion on CCL 2008 and CCL 2009, which
employ VAR models, they are by no means the only form of system dynamics. In fact,
future research may want to explore alternative econometric models that may capture the
system-dynamics nature of the data even better than the VAR models.
REFERENCES
Azariadis, C., and B. Smith, 1988, Financial Intermediation and Regime Switching in
Business Cycles, American Economic Review 88, pp. 51636.
Bank for International Settlements, 2003, 73rd Annual Report.
Bernanke, B., andM. Gertler, 1995, Inside the BlackBox: The Credit Channel of Monetary
Policy Transmission, Journal of Economic Perspectives 9, pp. 2748.
Bernanke, B., and M. Gertler, 2001, Should Central Banks Respond to Movements in
Asset Prices?, American Economic Review 91, pp. 2537.
Blinder, A. S., and R. Reis, 2005, Understanding the Greenspan Standard, Working
Paper, Princeton University.
Bohl, M. T., P. L. Siklos andT. Werner, 2007, Do Central Banks React to the Stock Market?
The Case of the Bundesbank, Journal of Banking and Finance 31, pp. 71933.
Bossaerts, P., and P. Hillion, 1999, Implementing Statistical Criteria to Select Return
Forecasting Models: What Do We Learn?, Review of Financial Studies 12, pp. 40528.
Buiter, W., 2009, The Unfortunate Uselessness of Most State-of-the-Art Academic
Monetary Economics, Financial Times, March 3.
Business Week, 2009, What Good Are Economists?, April 27.
Campbell, J. Y., 1987, StockReturns andthe TermStructure, Journal of Financial Economics
18, pp. 37399.
Campbell, J. Y., andJ. F. Cocco, 2007, Howdo house prices affect consumption? Evidence
from micro data, Journal of Monetary Economics 54, pp. 591621.
Campbell, J. Y., and R. J. Shiller, 1988, The DividendPrice Ratio and Expectations of
Future Dividends and Discount Factors, Review of Financial Studies 1, pp. 195227.
131
MODEL RISK
Campbell, J. Y., and M. Yogo, 2006, Efcient Tests of Stock Return Predictability, Journal
of Financial Economics 81, pp. 2760.
Capozza, D., andP. Seguin, 1998, Managerial Style andFirmValue, Real Estate Economics
26, pp. 13150.
Capozza, D., and P. Seguin, 1999, Focus, Transparency and Value: The REIT Evidence",
Real Estate Economics 27, pp. 587619.
Carlino, G., and R. Dena, 1998, The Differential Regional Effects of Monetary Policy,
Review of Economics and Statistics 80, pp. 57287.
Carlino, G., and R. Dena, 1999, The Differential Regional Effects of Monetary Policy:
Evidence From the US States, Journal of Regional Science 39, pp. 33958.
Carlino, G., and R. Dena, 2006, Macroeconomic Analysis Using Regional Data: An
Application to Monetary Policy, in R. Arnott and D. McMillen (eds), A Companion to
Urban Economics (Oxford: Blackwell).
Case, K. E., and R. J. Shiller, 1990, Forecasting Prices and Excess Returns in the Housing
Market, Real Estate Economics 18, pp. 25373.
Case, K. E., J. Quigley and R. J. Shiller, 2005, Comparing Wealth Effects: The Stock
Market versus the Housing Market, Advances in Macroeconomics 5, p. 1235.
Chami, R., T. F. Cosimano and C. Fullenkamp, 1999, The Stock Market Channel of
Monetary Policy, Working Paper 99/22, International Monetary Fund.
Chan, W. H., andC. K. Y. Leung, 2008, State-Dependent Jumps inStock Market Returns,
Mimeograph, City University of Hong Kong.
Chang, K. L., N. K. Chen and C. K. Y. Leung, 2008, Monetary Policy, Term Structure
and Asset Return: Comparing REIT, Housing and Stock, Working Paper (presented at
the DePaul Workshop on REIT).
Chang, K. L., N. K. Chen and C. K. Y. Leung, 2009, Would Some Model Please Give Me
Some Hints? An Empirical Investigation on Monetary Policy and Asset Return Dynam-
ics, Working Paper (presented at the Asian Real Estate Society Meeting), URL: http://
economic.ccu.edu.tw/2009/conference/1E1.pdf.
Chen, N. K., 2001, Bank Net Worth, Asset Prices and Economic Activities, Journal of
Monetary Economics 48, pp. 41536.
Chen, N. K., and C. K. Y. Leung, 2008, Asset Price Spillover, Collateral and Crises: With
an Application to Property Market Policy, Journal of Real Estate Finance and Economics 37,
pp. 351385.
Christiano, L. J., M. Eichenbaum and C. L. Evans, 1998, Modeling Money, Working
Paper 6371, NBER.
Christiano, L. J., M. Eichenbaum and C. L. Evans, 1999, Monetary Policy Shocks: What
Have We Learned and To What End?, in J. B. Taylor and M. Woodford (eds), Handbook of
Macroeconomics, Volume 1A(Amsterdam: Elsevier Scence).
Clapp, J. M., andC. Giaccotto, 1994, The Inuence of Economic Variables onLocal House
Price Dynamics, Journal of Urban Economics 36, pp. 16183.
Cochrane, J., 2001, Asset Pricing (Princeton University Press).
Emiris, M., 2006, The Term Structure of Interest Rates in a DSGE Model, Mimeograph,
National Bank of Belgium.
Estrella, A., 2005, Why Does the Yield Curve Predict Output and Ination?, Economic
Journal 115, pp. 72244.
132
Estrella, A., and G. Hardouvelis, 1991, The Term Structure as a Predictor of Real
Economic Activity, Journal of Finance 46, pp. 55576.
Estrella, A., and F. Mishkin, 1997, The Predictive Power of the TermStructure of Interest
Rates in Europe and the United States: Implications for the European Central Bank,
European Economic Review 41, pp. 13751401.
Estrella, A., and M. R. Trubin, 2006, The Yield Curve as a Leading Indicator: Some
Practical Issues, Current Issues in Economics and Finance 12, July/August, No. 5, URL:
http://www.newyorkfed.org/research/current_issues/ci12-5.pdf.
Fama, E., 1990, Term-Structure Forecasts of Interest Rates, Ination, and Real Returns,
Journal of Monetary Economics 25, pp. 5976.
Fama, E., and K. French, 1988, Dividend Yields and Expected Stock Returns, Journal of
Financial Economics 22, pp. 327.
Fama, E., and K. French, 1989, Business Conditions and Expected Returns on Stocks and
Bonds, Journal of Financial Economics 25, pp. 2349.
Francq, C., and J. M. Zakoian, 2001, Stationarity of Multivariate Markov-Switching
ARMAModels, Journal of Econometrics 10, pp. 33964.
Frankel, J. A., and A. Rose, 1996, Currency Crashes in Emerging Markets: An Empirical
Treatment, Journal of International Economics 41, pp. 35166.
Gertler, M., and C. Lown, 1999, The Information in the High-Yield Bond Spread for the
Business Cycle: Evidence and Some Implications, Oxford Review of Economic Policy 15,
pp. 13250.
Goetzmann, W. N., andP. Jorion, 1993, Testing the Predictive Power of DividendYields,
Journal of Finance 48, pp. 66379.
Goodhart, C., 2001, What Weight Should Be Given to Asset Prices in the Measurement
of Ination?, Economic Journal 111, pp. 33556.
Goodhart, C., and B. Hofmann, 2007, House Prices and the Macroeconomy: Implications for
Banking and Price Stability (Oxford University Press).
Goyal, A., and I. Welch, 2003, Predicting the Equity Premium with Dividend Ratios,
Management Science 49, pp. 63954.
Green, R., S. Malpezzi and S. Mayo, 2005, Metropolitan-Specic Estimates of the Price
Elasticity of Supply of Housing, and Their Sources, American Economic Review 95(2),
pp. 3349.
Greenwood, J., and Z. Hercowitz, 1991, The Allocation of Capital and Time over the
Business Cycle, Journal of Political Economy 99, pp. 11881214.
Hamilton, J., 1994, Time Series Analysis (Princeton University Press).
Hansen, A. T., andR. Poulsen, 2000, ASimple Regime Switching TermStructure Model,
Finance and Stochastics 4, pp> 40929.
Himmelberg, C., C. Mayer and T. Sinai, 2005, Assessing High House Prices: Bubbles,
Fundamentals and Misperceptions, Journal of Economic Perspective 19, pp. 6792.
Hodrick, R., 1992, Dividend Yields and Expected Stock Returns: Alternative Procedures
for Inference and Measurement, Review of Financial Studies 5, pp. 35786.
Hott, C., and P. Monnin, 2008, Fundamental Real Estate Prices: An Empirical Estimation
with International Data, Journal of Real Estate Finance and Economics 36, pp. 42750.
Hjalmarsson, E., 2008, Predicting Global Stock Returns, International Finance Discus-
sion Papers 933, Board of Governors of the Federal Reserve System.
133
MODEL RISK
Jaccard, I., 2007, Real Estate Prices in Production Economies, Mimeograph, European
Central Bank.
Jaccard, I., 2008, Asset Pricing in Home Production Economies, Mimeograph, European
Central Bank.
Jarocinski, M., and F. Smets, 2008, House Prices And The Stance of Monetary Policy,
ECB Working Paper 891.
King, R., and M. Watson, 1994, The Post-War US Phillips Curve: A Revisionist Econometric
History, CarnegieRochester Conference Series on Public Policy, Volume 41, pp. 157219
(Amsterdam: Elsevier).
King, R., and M. Watson, 1997, Testing Long-Run Neutrality, Federal Reserve Bank of
Richmond Economic Quarterly 83, pp. 69101.
Kiyotaki, N., and J. Moore, 1997, Credit Cycles, Journal of Political Economy 105, pp. 211
48.
Iacoviello, M., 2005, House Prices, Borrowing Constraints, and Monetary Policy in the
Business Cycle, American Economic Review 95, pp. 73964.
Leeper, E. M., C. A. Sims and T. Zha, 1996, What Does Monetary Policy Do?, Brookings
Papers on Economic Activity 2, pp. 163.
Lettau, M., and S. C. Ludvigson, 2005, Expected Returns and Expected Dividend
Growth, Journal of Financial Economics 76, pp. 583626.
Leung, C. K. Y., 2004, Macroeconomics and Housing: a Reviewof the Literature, Journal
of Housing Economics 13, pp. 24967.
Leung, C. K. Y., 2007, EquilibriumCorrelations of Asset Price and Return, Journal of Real
Estate Finance and Economics 34, pp. 23356.
Leung, C. K. Y., andW. L. Teo, 2008, TheWall Street Concerns Might BeRight: Equilibrium
Correlations of Asset Price do Change with Monetary Policy, Working Paper (presented
at the Asian Real Estate Society Meeting).
Leung, C K. Y., and W. L. Teo, 2009, Should the Optimal Portfolio Be Region-Specic?
A Multi-Region Model with Monetary Policy and Asset Price Co-Movements, Working
Paper (presented at the Far East Econometric Society Meeting).
Lucas, R., 1978, Asset Prices in an Exchange Economy, Econometrica 46, pp. 142645.
Lustig, H., andS. VanNieuwerburgh, 2005, HousingCollateral, ConsumptionInsurance
and Risk Premia: An Empirical Perspective, Journal of Finance 60, pp. 1167219.
Maheu, J., and T. McCurdy, 2000, Identifying Bull and Bear Markets in Stock Returns,
Journal of Business and Economic Statistics 18, pp. 10012.
Malpezzi, S., 2008, Housing Supply, Mimeograph, University of Wisconsin (Madison).
Menzly, L., T. Santos and P. Veronesi, 2004, Understanding Predictability, Journal of
Political Economy 112, pp. 147.
Mishkin, F., 2001, The Transmission Mechanismandthe Role of Asset Prices in Monetary
Policy, NBER Working Paper 8617.
Mishkin, F., 2007, Housing and the Monetary Transmission Mechanism, Working
Paper (presented at the Annual Economic Symposium organised by the Kansas City
Fed in Jackson Hole), URL: http://www.federalreserve.gov/Pubs/feds/2007/200740/
200740pap.pdf.
Ortalo-Magne, F., and S. Rady, 2006, Housing Market Dynamics: On the Contribution
of Income Shocks and Credit Constraints, Review of Economic Studies 73, pp. 45985.
134
Piazzesi, M., M. Schneider and S. Tuzel, 2007, Housing, Consumption, and Asset
Pricing, Journal of Financial Economics 83, pp. 53169.
Plosser, C., and K. G. Rouwenhorst, 1994, International Term Structures and Real
Economic Growth, Journal of Monetary Economics 33, pp. 13355.
Rachev, S. T., F. T. Fabozzi and C. Menn, 2005, Fat-Tailed and Skewed Asset Return Distri-
butions: Implications for Risk Management, Portfolio Selection, and Option Pricing (New York:
Wiley).
Rigobon, R., and B. Sack, 2003, Measuring the Reaction of Monetary Policy to the Stock
Market, Quarterly Journal of Economics 118, pp. 63969.
Rosenberg, J. V., andS. Maurer, 2008, Signal or Noise? Implications of the TermPremium
for Recession Forecasting, Economic Policy Review 14, July, No. 1.
Sanders, A., 2008, The Subprime Crisis and Its Role in the Financial Crisis, Journal of
Housing Economics 17, pp. 25461.
Sargent, T., N. Williams and T. Zha, 2006, Shocks and Government Beliefs: The Rise and
Fall of American Ination, American Economic Review 96, pp. 1193224.
Shiller, R., 2008, The Subprime Solution (Princeton University Press).
Sims, C., 1980a, Comparison of Interwar and Postwar Business Cycles: Monetarism
Reconsidered, American Economic Review 70, pp. 25057.
Sims, C., 1980b, Macroeconomics and Reality, Econometrica 48, pp. 148.
Sims, C., and T. Zha, 2006, Were There Regime Switches in US Monetary Policy?,
American Economic Review 96, pp. 5481.
Sutton, G. D., 2002, Explaining Changes in House Prices, BIS Quarterly Review,
September, pp. 4655.
Teo, W. L., 2009, Asset Return Dynamics under Different Monetary Policy Regimes,
Mimeograph, National Taiwan University.
Tsatsaronis, K., and H. Zhu, 2004, What Drives Housing Price Dynamics: Cross-Country
Evidence, BIS Quarterly Review, March, pp. 6578.
Yao, R., and H. Zhang, 2005, Optimal Consumption and Portfolio Choices with Risky
Housing and Borrowing Constraints, Review of Financial Studies 18, pp. 197239.
Yoshida, J., 2008, Technology Shocks and Asset Price Dynamics: The Role of Housing in
General Equilibrium, Mimeograph, University of Tokyo.
Zhou, Z., 1997, Forecasting Sales and Price for Existing Single-Family Homes: A VAR
Model with Error Correction, Journal of Real Estate Research 14, pp. 15567.
135
6
Capital Divergence: Sensitivity of
Economic and Regulatory
Capital under Stress
Oleg Burd
KfW IPEX-Bank GmbH
The 20089 nancial crisis raised questions about central issues of
credit risk management and the management of regulatory and eco-
nomic capital, amongothers. The discussiontriggeredamongsuper-
visory authorities, academia andindustry has so far focusedon poli-
cies on the management of regulatory capital and their implications
for the whole nancial system.
1
However, regulatory capital is not
the only scarce resource to be managed; economic capital, which
is calculated by internal models with a similar philosophy to that
of regulatory capital, also needs to be managed. In this chapter we
analyse how both constraints (the economic and regulatory capital
requirements) react to an increase in input parameters (default prob-
ability and correlation) which corresponds to a transformation from
a benign to an adverse economic environment.
The impact of stress testing on credit portfolios has been studied
by, among others, Rsch and Scheule (2007), Mager and Schmieder
(2008) and Dllmann and Erdelmeier (2009). We focus, however,
not only on additional capital requirements arising in stress tests,
but also on the different sensitivities of regulatory and economic
models.
Our ndings showthat economic capital is signicantly more vul-
nerable to stress than regulatory capital. We have found that matu-
rity adjustment and asset correlation, which are both modelled in
risk weight formulas as decreasing functions of default probabili-
ties, are responsible for this divergence. Our results have implica-
tions for the capital management of credit portfolios and suggest
that, as long as models with different sensitivities are in place, this
137
MODEL RISK
divergence has to be taken into account by capital management (eg,
by different capital buffers for economic and regulatory capital).
The remainder of the chapter is organised as follows. We begin
the next section by reviewing the general risk-factor model and
its implementation for the regulatory and economic framework.
We then describe the benchmark portfolio adopted for the study.
2
In the following section, we introduce different stress scenarios
and calculate their impact on capital requirements. The last section
summarises and concludes this chapter.
GENERAL FRAMEWORK
In this section we recall the factor model for portfolio credit risk,
which forms a foundation for regulatory and economic capital
methodologies. Subsequently, we present these frameworks and
highlight the difference between them.
Risk-factor model for portfolio credit risk
We consider a credit portfolio of n exposures, where the ith exposure
has a principal amount A
i
over a specic time horizon (eg, one year).
All amounts A
i
are assumed to be known and non-random. The
weight w
i
of exposure i in the portfolio is given by a ratio of its
principal to the total principal of the portfolio
w
i
=
A
i
n
i=1
A
i
For simplicity, weassumethat everyobligor has onlyasingleliability
and, in the case of the obligors default, the fraction of principal
amount which is lost (loss given default, LGD) is deterministic and
equals LGD
i
. We consider a book-value approach to risk, with losses
arising from a default event only. Obligor i defaults if its asset value
V
i
falls below some threshold c
i
. The total portfolio loss fraction
(hereafter, portfolio loss) L is expressed as
L =
n
_
i=1
w
i
LGD
i
1
V
i
c
i
(6.1)
Asset value V
i
is modelled as a random variable, which depends
on a vector X of d systemic risk factors and on random variable
i
,
which represents the idiosyncratic risk
V
i
=
i
X +
i
i
, (6.2)
138
CAPITAL DIVERGENCE: SENSITIVITY OF ECONOMIC AND REGULATORY CAPITAL UNDER STRESS
where a d-dimensional vector of factor loadings
i
R
d
with
0 < |
i
| < 1 denes the dependency of V
i
on systemic risk factors.
To stay within the Gaussian framework, we assume that systemic
risk X is a d-dimensional standard normal vector with a correlation
matrix , and idiosyncratic risk factors
i
are independent, identi-
cally distributed standard normal randomvariables independent of
X. By setting the factor loading
i
of the idiosyncratic risk factor to
be equal to
i
=
_
1
T
i

i
we ensure that V
i
are also standard normal distributed.
Under the independency assumption (
i
is independent of X),
the conditional probability p
i
(x) that obligor i defaults for a given
realisation x of X is given by
p
i
(x) = Pr[V
i
< c
i
X = x]
= Pr[
i
X +
i
i
< c
i
X = x]
=
_
c
i
i
x
i
_
(6.3)
where denotes the cumulative distribution function of standard
normal distribution.
Regulatory capital
The regulatory capital requirements are calculated according to the
advanced Internal Rating Based Approach (IRBA) for corporate
loans as suggested by the Basel Committee on Banking Supervi-
sion (2004). The IRBAassumes a single risk factor for systemic risk
(d = 1) andinnitelyne-grainedcredit portfolio, whichreduces the
general multi-factor risk model outlined in the previous section to
the asymptotic single risk factor (ASRF) model. Gordy (2003) shows
that, in an ASRF framework, portfolio loss L conditional on X con-
verges in probability to its conditional expectation E[L X = x],
ie
L[X = x]
almost surely
E[L X = x] =
n
_
i=1
w
i
LGD
i
p
i
(x)
with p
i
(x) specied by Equation 6.3
We dene the mapping l : R (0, 1) as
l(x) = E[L X = x] =
n
_
i=1
w
i
LGD
i
p
i
(x)
139
MODEL RISK
Loss function l(x) is strictly decreasing and continuous, and thus we
express a distribution of L by
Pr[L x] = 1 (l
1
(x)) with x (0, 1) (6.4)
The -quantile of loss distribution L, which corresponds to the
value-at-risk (VaR) is given by
VaR
(L) = q
(L) = inf
_
l R : Pr[L x]
= l(
1
(1 ))
=
n
_
i=1
w
i
LGD
i
_
c
i
+
i
1
()
_
1
2
i
_
(6.5)
In accordance with results of Tasche (1999), the contribution of the
ith exposure VaR
i,
to portfolio VaR
(L) can be attained by partial

differentiation
VaR
i,
= w
i
LGD
i
VaR
(L)
w
i
= w
i
LGD
i
_
c
i
+
i
1
()
_
1
2
i
_
where threshold c
i
is determined by the unconditional default
probability pd
i
as c
i
=
1
(pd
i
).
The supervisory capital charges K
i
for unexpected losses are mea-
sured as a difference between VaR at a condence level of 99.9%
and expected losses EL
i
(where EL
i
= LGD
i
pd
i
) amended by the
maturity adjustment function MA(pd
i
, M
i
). They are given by
K
i
= (VaR
i,99.9%
EL
i
)MA(pd
i
, M
i
) (6.6)
The Basel maturity adjustment function MA(pd, M) addresses the
higher risk of loans with longer maturities, which is ignored by
the VaR in a one-period risk-factor framework. It is modelled as an
increasing function of maturity M, taking a value of 1 for one-year
maturities and decreasing with respect to pd.
3
Figure 6.1 shows the
behaviour of the IRBAmaturity adjustment function.
To make Equation 6.6 operative and comparable among differ-
ent institutions, the Basel Committee sets factor loadings
i
to be
modelled as an exponentially decreasing function of pd
i
, bounded
between 12% and 24%
4
i
= (pd
i
) = 24%(1 a
i
) +12%a
i
with a
i
=
1 e
50pd
i
1 e
50
(6.7)
140
Figure 6.1 IRBA maturity adjustment as function of pd and maturity
0.0
0.1
0.2
0.3
0.4
0.5
pd
1
2
3
4
5
Maturity
1.0
1.5
2.0
2.5
M
a
t
u
r
i
t
y
a
d
j
u
s
t
m
e
n
t
Economic capital
We calculate economic capital with a multi-factor CreditMetrics-
type portfolio model for a one-period time horizon and default-only
mode. The loss distribution of the portfolio is generated by Monte
Carlo simulations of idiosyncratic and systemic risk factors. In addi-
tion to the portfolio described in the previous section, there is only
one input parameter to be specied: the correlation matrix of sys-
temic risk factors . We derive this correlation fromdaily returns of
corresponding MSCI equity indexes for the period of one year from
June 2008 to June 2009.
To establish comparability with the IRBA framework, the eco-
nomic capital of the portfoliois measuredas VaRat acondence level
of 99.9%. The allocation of capital to sub-portfolios in a multi-factor
framework is a more subtle matter than in the ASRF framework:
VaR is not subadditive and as such is not a coherent risk measure in
the sense of Artzner et al (1999). With expected shortfall (ES), Acerbi
andTasche (2001) introduceda coherent alternative toVaR. Expected
shortfall is a risk measure that describes an average loss in the worst
cases. It can be calculated for smooth loss distribution L as
ES
(L) =
1
1
_
1
VaR
u
(L) du (6.8)
Toallocate the total portfolioVaRVaR
(L) we choose a condence

level , resulting in the following relation
ES
(L) = VaR
(L) (6.9)
Finally, we allocate
ES
(P) =
_
iP
ES
(L
i
)
141
MODEL RISK
Table 6.1 Asset correlation in IRBA and multi-factor models
Regulatory Economic
approach (%) approach (%)
Maximum 23.8 65
Average 21.2 10.6
Weighted average
21.3 15.3
Minimum 12.0 10.9
Netto-exposure weighted average correlation.

units of capital (adjustedfor expectedlosses) toa sub-portfolioP. For
the benchmark portfolio P used in this study, we found numerically
that equals 99.7%, satisfying Equation 6.9 for the regulatory xed
condence level of 99.9%.
5
Differences between both frameworks
The most striking difference between economic and regulatory
frameworks, multi-factor versus single-factor modelling, appears to
be an innocent one. Thus, Tarashev and Zhu (2007) showed that
a single risk-factor model performs sufciently well by a proper
parameter calibration.
We examined other input parameters: maturity adjustment and
correlation. In the default mode of the CreditMetrics-type model,
longer maturities play hardly any role, as the model is only a one-
period model. The second parameter, correlation, is dened as a cor-
relation of systemic risk factors and is independent of the obligors
default probabilities.
Recent studies on credit risk (eg, Tarashev and Zhu (2007) and
Rosenow et al (2007)) have shown that calibration of correlation
parameter has a far-reaching impact on portfolio risk.
6
Table 6.1
illustrates the properties of asset correlations ina portfoliocalculated
within regulatory and economic frameworks.
DATA DESCRIPTION
For illustrative purposes, we adopted the portfolio used in the
joint ISDA/IACPM study on the convergence of capital models
(IACPM/ISDA 2006). The study was conducted by 28 nancial
institutions (members of IACPM) with the following purpose: to
developanunderstanding of the various data or model assumptions
142
Table 6.2 Rating distribution of portfolio
Rating One-year Exposure
grade pd (%) (millions, %) LGD (%)
AAA 0.010 3,332 39.7
AA 0.021 6,640 43.0
A 0.021 20,667 40.6
BBB 0.178 38,862 40.7
BB 1.271 20,513 40.1
B 6.639 7,259 39.2
CCC 25.499 2,727 40.5
that would reconcile differences in capital estimates. The structure
of the portfolio is not specic to certain operations and therefore is
well suited to various studies on loan portfolios (see, for example,
Algorithmics 2008).
The portfolio, with a total amount of US$100 billion, consists of
3,000 borrowers. Eachborrower has twooutstandingexposures with
a specic LGD in the range 2258%.
7
To analyse concentration and
diversications effects in the portfolio, each borrower is assigned to
an industry, a rating bucket and a country.
Each of the 3,000 borrowers in the portfolio has a rating corre-
sponding to Standard&Poors letter grade scale fromAAA (high-
est) to CCC (lowest).
8
To maintain the comparability of the results,
we adopted the default probabilities used in the IACPM/ISDA
study. Table 6.2 presents the distribution of the portfolio among rat-
ing grades accompanied by default probabilities and by weighted
average LGDs for each grade.
The maturities of exposures vary between six months and seven
years with a distribution centred around the mean of 2.5 years, as
shown in Figure 6.2.
With exposures between US$1 million and US$1,250 billion, the
single name concentration in the portfolio is low. We measured it
with the HerndahlHirschman Index (HHI) and the share of the 5,
30 and 300 biggest names in the whole portfolio. Table 6.3 reports
both concentration measures calculated with respect to exposure
and to LGD-adjusted exposure (so-called netto-exposure).
Table 6.4 describes the sectoral distribution and provides an eco-
nomic and regulatory capital allocation among the sectors. Gener-
ally, the discrepancies between both frameworks cause deviations
143
MODEL RISK
Table 6.3 Single name concentration in portfolio
Concentration Exposure Netto-exposure
measure weighting (%) weighting (%)
Herndahl indexes 0.17 0.18
Top 5 6 7
Top 30 24 26
Top 300 70 70
Table 6.4 Regulatory and economic capital requirements as
percentage of total exposure
ISDA/IIF Obligor Average Regulat. Econ.
industry name Exposure no. pd (%) capital capital
Automobiles 7,985 64 1.98 339 372
Banking and 11,331 179 0.74 403 511
nance
Broadcasting 3,745 32 0.34 117 96
Chemicals 4,095 108 1.20 161 163
Construction 3,685 125 1.34 123 122
Electronics 1,933 95 7.53 150 217
Energy 4,667 127 0.72 158 135
Entertainment 979 163 1.47 48 21
Food 2,138 170 1.42 94 38
General 4,943 201 0.84 234 190
Health care 1,049 103 1.30 52 46
Hotels 1,221 122 0.94 51 28
Insurance 2,946 133 0.90 117 104
Machinery 2,895 194 1.96 174 218
Manufacturing 5,684 151 2.15 357 615
Metals and 2,960 64 2.05 182 212
mining
Oil and gas 7,715 112 1.24 363 314
Paper and 3,669 133 1.38 170 149
forest products
Publishing 2,459 128 2.58 138 129
Technology 5,377 135 1.33 205 211
Telecoms 9,350 112 0.58 210 111
Textiles 575 44 4.53 39 24
Transportation 4,439 135 0.88 158 173
Utilities 4,160 170 1.83 159 110
Portfolio 100,000 3,000 1.38 4,202 4,309
144
Figure 6.2 Maturity distribution of portfolio
E
x
p
o
s
u
r
e

(
U
S
$

b
n
)
30
25
20
15
10
0
5
0.5 1 2 3 5 7
Maturity (years)
between corresponding capital requirements. Even if the total port-
folio capital requirements are close to each other, the deviations
between economic and regulatory capital contributions at sector
level are substantial and range from 62% in defensive (ie, less
cyclical) industries, suchas the foodindustry, to+63%for the cyclical
manufacturing industry.
Finally, Figure 6.3 shows the regional distribution, which appears
to be balanced, albeit slightly biased towards the US.
EMPIRICAL RESULTS OF STRESSTESTING
In this section we describe the stress scenario and apply it to the
benchmark portfolio in order to analyse its impact on regulatory
and economic capital.
Stress scenario design
The stress scenario consists of stressing default probabilities andcor-
relations, whichhave a non-linear effect oneconomic capital require-
ments. To make our results comparable with other studies (see, for
example, IACPM/ISDA 2006) to a certain extent, we included an
increase in default probabilities by 20% and 40%. We also analysed
a more severe (but still realistic) scenario, with the stress of default
probabilities by 100%.
Because the IRBAcorrelation formula is xed by supervisors, the
correlation stress has an impact only on economic capital calcula-
tion. In practice, the correlation stress of systemic risk factors is the
most delicate part of stress scenario design. Generally, there are two
145
MODEL RISK
Figure 6.3 Regional distribution of portfolio
US 36%
Switzerland
11%
Japan
18%
Germany
12%
France
8%
Australia
7% UK 8%
ways to obtain the stressed correlation matrix. The rst (historical)
one consists of a selection of historically observed correlation matri-
ces which correspond to a distressed time period or fulll certain
criteria (eg, high average correlation, high correlation between cer-
tain factors). The other approach remodels the original correlation
matrixbychangingsome(or all) of its elements, andif theresulting
matrix
stressed, interim
is not a valid correlation matrix,
9
the appropri-
ate methods (see, for example, Dash 2004; Higham 2002; Rebonato
and Jckel 2000) are applied in order to nd the valid correlation
matrix
stressed
, which is the closest (ie, of Frobenius norm on R
d
)
to
interim
. We applied the second approach, being the more exible
one, and increased in our study all correlations of in such a way
that the average correlation dened by
stressed
is 40% higher then
the average correlation of .
Results of the stress test
We studied the sensitivity of regulatory and economic capital
requirements in univariate and multivariate scenarios. Table 6.5
presents the relative increases in capital requirements in each
scenario.
The most striking observation is that economic capital appears to
be more sensitive than regulatory capital in all scenarios. To explain
the divergence of sensitivities in the univariate pd-stress scenarios, it
is important to remember that maturity adjustment and correlation
function in the IRBA framework are both modelled as decreasing
146
Table 6.5 Regulatory and economic capital requirements in various
stress scenarios
Regulat. Econ.
capital capital
Scenario increase increase
type Parameter stressed (%) (%)
Univariate pd +20% 6.4 10.6
Univariate Correlation+40% 0.0 23.3
Multivariate Correlation+40% and pd +20% 6.4 32.7
Table 6.6 Increase in regulatory capital requirements with constant
maturity and constant correlation
pd IRBA Fixed Fixed correlation/
increase standard maturity maturity Economic
(%) (%) (%) (%) capital (%)
+20 6.4 7.5 8.2 10.6
+40 11.9 14.0 15.3 16.6
+100 25.6 30.0 32.7 33.3
IRBA standard, IRBA formula standard; Fixed maturity, IRBA formula
with pd-independent maturity; Fixed correlation/maturity, IRBA formula
with pd-independent maturity and pd-independent correlation.
functions of default probabilities. Hence, the question arises as to
howregulatory capital would react if these features were switched
off.
To answer this question, we rst xed the maturity of all expo-
sures to one year, which makes the maturity adjustment function
independent of pd and constant.
10
Hereafter, we set the asset corre-
lation to a maximum IRBA value of 24%. Table 6.5 shows how the
regulatory capital requirements change if maturity and correlation
functions are set to be constant. Taking the pd-dependency out of
both maturity adjustment and the correlation function results in an
IRBA capital increase in univariate pd-scenarios, which are similar
to those in an economic multi-factor model.
147
MODEL RISK
Table 6.7 Sensitivity (%) of regulatory capital requirements with xed
maturity and stressed correlation of 28.4%
IRBA Economic
Scenario adjusted (%) capital (%)
Stressed correlation 17.9 23.3
Stressed correlation+pd 20% 27.1 32.7
IRBAadjusted, IRBAformula with maturity xed to one year and constant
correlation of 28.4%.
Table 6.6 demonstrates that the sensitivity of the IRBA function
approaches under univariate pd-stress the sensitivity of economic
capital if IRBAmaturities and IRBAcorrelations are independent of
pd and constant. Nevertheless, the increase in regulatory capital of
32.7% remains far beyond the increase in economic capital of up to
61.9%in the multivariate scenario with both parameters, correlation
and pd, stressed. To achieve such a capital increase in the regulatory
framework, we have to raise the asset correlation parameter .
In the ASRF model the average correlation between any two bor-
rowers equals
2
. Thus, the correlation stress of +40% is achieved
through an increase in by 28. 4% =

40%. Table 6.7 compares
an increase in regulatory capital requirements (calculated with a
stressed correlation of 28.4% and a xed maturity of one year) with
the sensitivity of economic capital. Although we found numerically
that increasing the correlation up to 30% leads to similar increases
in economic and regulatory capital requirements at portfolio level,
the divergence between contributions at the sector level persists.
CONCLUDING REMARKS
The results presented in this chapter show that the IRBA formula
for regulatory capital has several built-inassumptions (ie, maturities
and correlations are decreasing functions of default probability) that
cause the regulatory capital requirements to be less sensitive than
economic capital requirements.
Because both regulatory and economic capital are bound con-
straints, the difference in sensitivity has a signicant impact on how
148
credit portfolios are managed and thus should be at least taken into
account for consistent risk management.
The question remains whether such a divergence is a desirable
feature of regulatory capital or whether it is a side effect of the IRBA
function. The latter appears to be the case. For our benchmark port-
folio, the pd-increase of 100% caused the average asset correlation
to fall by 5% from 21.2% to 20.1%. However, empirical evidence by
Longin and Solnik (1995, 2001), Ang and Chen (2002) and Campbell
et al (2002) suggests the opposite: in an adverse economic environ-
ment (eg, recession, crisis), the correlation tends to rise. Thus, in a
typical stress situation both the correlation and default probabilities
soar.
Hitherto, neither economic capital models nor regulatory models
have paid any attention to this characteristic. To avoid unexpected
capital increases, the immediate solution is to analyse the sensitiv-
ity of capital requirements to correlation and default-probability
stresses as presented in this chapter and to establish such capital
buffers that factor in the different sensitivities. In order to be consis-
tent withinbothframeworks andamongempirical ndings, a model
correlation as a non-constant and probably even non-deterministic
function of the economy state is required. While stochastic corre-
lation modelling approaches have already been made in a credit
risk context (see, for example, Burd 2009), further research is nec-
essary for more complex stochastic processes and, in particular, for
the multi-factor framework.
1 See Borio et al (2001) and Borio and Drehmann (2009) for a survey of the literature on the issue
of procyclicality.
2 We thank IACPM and ISDAfor providing the data on benchmark portfolios.
3 Basel Committee on Banking Supervision (2005) provides detailed information on derivation
and calibration of MA(pd, M).
4 The IRBAformula for also contains the size adjustment, whichequals zero for big borrow-
ers (50 million annual sales and above). We assume that benchmark portfolio contains only
such big borrowers and therefore we neglect a size adjustment for the correlation function
(pd).
5 We refer the reader to Kalkbrener and Overbeck (2004) for a study on capital allocation with
ES.
6 Intuitively, the correlation determines the shape of the portfolio loss distribution, particularly
the thickness of the distribution tail. Therefore, a portfolio risk dened as a high quantile of
loss distribution should depend strongly on correlation.
7 In addition to the value of the LGD, which can be treated as the expected value of the LGD
distribution, the data set contains also its standard deviation.
8 Thus, all loans in the portfolio are performing loans.
149
MODEL RISK
9 A real-valued square matrix is a valid correlation matrix if it is symmetric, positive semi-
denite and has diagonal elements equalling 1.
10 The IRBAformula is linear in respect to maturity. Thus, results regarding the relative changes
of capital requirements pertain to any constant value of maturity.
REFERENCES
Acerbi, C., and D. Tasche, 2001, Expected Shortfall: A Natural Coherent Alternative to
Value-at-Risk, Economic Notes 31 pp. 379388.
Algorithmics Software LLC, 2008, Studies on Default and Migration Risks, Technical
Report, Algorithmics Software LLC.
Ang, A., and J. Chen, 2002, Asymmetric Correlations of Equity Portfolios, Journal of
Financial Economics 63(3) pp. 443494.
Artzner, P., F. Delbaen, J.-M. Eber and D. Heath, 1999, Coherent Measures of Risk,
Mathematical Finance 9(3), pp. 20328.
Measurement and Capital Standards: A Revised Framework, Technical Report, Basel
Committee on Banking Supervision.
Basel Committee on Banking Supervision, 2005, An Explanatory Note on the Basel II
IRB Risk Weight Functions, Technical Report 15, Basel Committee on Banking Supervi-
sion, March.
Borio, C., and M. Drehmann, 2009, Towards an Operational Framework for Financial
Stability: Fuzzy Measurement and Its Consequences, Working Paper.
Borio, C., C. Furne and P. Lowe, 2001, Procyclicality of the Financial System and
Financial Stability: Issues and Policy Options, Working Paper.
Burd, O., 2009, Breaking Correlation Breakdowns: Non-Parametric Estimation of Down-
turn Correlations and Their Application in Credit Risk Models, The Journal of Risk Model
Validation (Special Issue on Implementing Stress Testing) 2(4), pp. 5164.
Campbell, R., K. Koedijk and P. Kofman, 2002, Increased Correlation in Bear Markets:
ADownside Risk Perspective, Financial Analysts Journal 58, pp. 8794.
Dash, J. W., 2004, Quantitative Finance and Risk Management: A Physicists Approach (World
Scientic).
Dllmann, K., and M. Erdelmeier, 2009, Stress Testing German Banks in a Downturn in
the Automobile Industry, Discussion Paper, Deutsche Bundesbank.
Gordy, M., 2003, ARisk-Factor Model Foundationfor Ratings-BasedBankCapital Rules,
Journal of Financial Intermediation 12(3), pp. 199232.
Higham, N. J., 2002, Computing the Nearest Correlation Matrix: A Problem from
Finance, IMA Journal of Numerical Analysis 22(3), pp. 32943.
IACPM/ISDA, 2006Convergence of Economic Capital Models, Technical Report, Rutter
Associates LLC (conducted on behalf of the International Association of Credit Portfolio
Managers and International Swaps and Derivatives Association).
Lotter, H., M. Kalkbrener and L. Overbeck, 2004, Sensible and Efcient Capital
Allocation for Credit Portfolios, Risk 17, pp. 1924.
Constant: 19601990?, Journal of International Money and Finance 14, pp. 326.
150
Longin, F., and B. Solnik, 2001, Extreme Correlations of International Equity Markets,
Journal of Finance 2(56) pp. 64974.
Mager, F., and C. Schmieder, 2008, Stress Testing of Real Credit Portfolios, Discussion
Paper, Deutsche Bundesbank.
Rebonato, R., and P. Jckel, 2000, The Most General Methodology for Creating a Valid
Correlation Matrix for Risk Management andOption Pricing Purposes, The Journal of Risk
2(2), pp. 1727.
Rsch, D., and H. Scheule, 2007, Stress-Testing Credit Risk Parameters: An Application
to Retail Loan Portfolios, The Journal of Risk Model Validation 1(1), pp. 5575.
Rosenow, B., R. Weibach and F. Altrock, 2007, Modelling Correlations in Credit
Portfolio Risk, Discussion Paper, Deutsche Bundesbank.
Tarashev, M., and H. Zhu, 2007, Modelling and Calibration Errors in Measures of
Portfolio Credit Risk, Working Paper.
Tasche, D., 1999, Risk Contributions and Performance Measurement, Working Paper.
151
Part III
Credit Portfolio Risk
Models
7
Diversied Asset Portfolio
Modelling: Sources and
Mitigants of Model Risk
Sean Keenan, Stefano Santilli, Sukyul Suh;
Andrew Barnes, Huaiyu Ma, Colin McCulloch
GE Capital; GE Global Research Center
Quantitative risk analytics has become increasingly important for
nancial institutions either that seek to match best practices or who
face evolving regulatory requirements. This is especially true in the
area of capital adequacy, where regulatory standards require insti-
tutions to have some defensible capability. With the rapid pace of
regulatory change, and with formulaic approaches (including the
Basel II formula) facing eroding support, institutions have little
choice but to invest in internal capital adequacy modelling capabili-
ties to stay ahead of the curve. Of course, while capital adequacy
measures (partly for external consumption) may be the primary
motivation, rms expect these modelling exercises to produce, as a
by-product, capital allocationcapabilities that canbe usedtomanage
the business more efciently. Rapid modelling advances combined
with increasing complexity in nancial products and institutional
design create huge opportunities for modellers to err, oversimplify
or otherwise produce answers that will turn out to be wrong. An
increased awareness of this model risk by already sceptical man-
agers and supervisors is helping to impose more discipline on the
modelling community and a more serious recognition of model risk
throughout the model development and use process.
For rms developing and managing complex portfolio loss mod-
els, two classes of concern about model risk need to be considered
and, to the extent possible, addressed. The rst is simply the avoid-
ance of what we may call gross model risk, which is the potential
for the user or developer to fail to recognise inconsistencies between
155
MODEL RISK
the structure of the model and the phenomena that the model is
being used to represent. Such inconsistencies can create large single-
point risks that may skew or completely invalidate model results.
Mitigation of these risks involves a careful review of model fea-
tures against the features of the problem under investigation, and
an attempt to recognise inconsistencies, evaluating their potential
impact and adjusting either the model itself, or the interpretation of
the model outputs. Since modelling always involves compromises
(as complexities of the real world are stripped away in favour of
feasibility), many such inconsistencies are intentional design fea-
tures and modellers have pre-assessed the magnitude and direc-
tion of potential distortions. But inconsistencies may slip through
in the design process, and though subtle, may have a huge impact
on model outputs. Moreover, when vendor models are employed,
users do not have access to the entire thought process that was fol-
lowed as the model specications were determined and thus they
must attempt to rethink it carefully, looking for such fundamental
inconsistencies.
The second concern relates to the potential for biases to develop
through accumulation, as model outputs are used as inputs to
higher-level models, perhaps creating many layers of dependency
and increasing complexity.
In this chapter we describe some of the experiences we have had
in developing and maintaining an integrated portfolio loss model
designed both for capital adequacy evaluation (economic capital)
and capital allocation purposes across a diverse loan, lease, equity
and physical asset portfolio. The chapter is not intended to be com-
prehensive, nor did we attempt to maintain a balance of theory and
practice across topics. Rather, we present the issues we have faced
and continue to wrestle with in their various stages of evolution
and resolution. Due to space constraints, and to keep our observa-
tions meaningful andrelevant, wefocus primarilyonthecommercial
portfolio andon the aggregation approach, touching more briey on
the consumer credit portfolio and other sub-portfolios.
GENERAL OVERVIEW OFTHE MODEL
Generally, modellingportfolioriskis acomplextaskthat mirrors, toa
large extent, the tiered operating structure of the business. Complex
156
DIVERSIFIED ASSET PORTFOLIO MODELLING: SOURCES AND MITIGANTS OF MODEL RISK
Figure 7.1 Risk streams requiring aggregation
Consumer
Commercial
CRE
Aviation
Insurance
RD
RV
RD, equipment redeployment; RV, residual value; CRE, commercial real estate.
portfolios are composed of diverse product types in multiple sectors
and geographies including commercial lending, leasing activities
that range from very small ticket ow-type businesses to large
ticket businesses such as commercial aviation and energy project
nance. They may include commercial real estate investments (both
debt and equity), leasing and lending activities in speciality sectors
such as health care and transportation, and a global consumer port-
folio that includes car loans, mortgages, credit cards and personal
loans. For leasing portfolios, contracts may include the return of
equipment post-lease for a pre-specied residual value (RV), creat-
ing a forward asset value risk. These RV risks need to be modelled
at the collateral-class level, and then aggregated into a single total
RV risk estimate.
To model the joint loss potential across these diverse components,
losses are modelled at the like-asset level, and then aggregated
upwards. Finally, a set of some 27 loss components, represented
by expected one-year forward-loss distributions and time-series
loss histories, are aggregated using a copula approach. These indi-
vidual risk streams represent seven major portfolio components,
summarised as commercial credit, consumer credit, commercial
real estate (CRE), commercial aviation services, insurance assets,
RV exposure and equipment redeployment (RD) risk as shown in
Figure 7.1.
157
MODEL RISK
The largest and arguably most complex part of the portfolio is
the commercial credit component that includes leasing and lend-
ing. This sub-portfolio is modelled using a commercially available
portfolio-modelling tool. Henceforth, we will refer to this vendor
model as the commercial credit model (CCM). All of the other
components are modelled using proprietary, internally developed
models. These either derive the component-level risk prole using
granular, bottom-up methodologies similar to the CCM, or apply
a top down model that accounts for the internal diversication in
the component portfolio to derive its loss distribution. Finally, these
sub-portfolio model outputs are aggregated at the highest level of
the organisation, creating a one-year-forward estimated-loss dis-
tribution used to assess the capital required to withstand increas-
ing amounts of stress. A second, disaggregation, process attributes
the diversied capital back down to the individual transaction
level, to help understand the diversication and diversied risk of
businesses, products, sectors, etc.
The remainder of this chapter focuses on the modelling of com-
mercial credit exposures, on the aggregation/attribution method,
and on the interpretation of the model outputs.
MODELLING COMMERCIAL EXPOSURES
To mark, or not to mark
One of the biggest challenges often faced is to deploy a mod-
elling system that contained an economically sensible denition
of loss given the different accounting treatments and management
approaches for different parts of a portfolio. In particular, we have
some investments held for sale as well as other assets that require
mark-to-market accounting and are managed as more-or-less liquid
assets, with actual or implied market pricing used to assess losses
and gains. This corresponds to a commonly used concept of loss
available in the CCM, so there seemed to be little gross model risk
created by using the vendor model for these assets.
However, if the bulk of a loan and lease portfolio is held to matu-
rity by design, losses are essentially all credit losses. By convention,
unearned income is excluded, and losses of principal and accrued
interest are accounted for as post-default charge-offs. The use of a
mark-to-market loss concept for these assets is thus immediately at
variance with generally accepted accounting principles (GAAP), as
158
well as with business practice. Also, because in general no pricing
exists for these illiquid exposures, mark-to-market is by necessity
converted to mark-to-model (MtM), with potentially huge model
risk implications. Therefore, our analysis of the applicability of the
CCM for the bulk of our commercial portfolio (assets other than
securities held for sale), involved the following three steps:
(i) gauge the capital implications of using a mark-to-market loss
concept;
(ii) assess the model risk associated with the CCMs MtM ap-
proach;
(iii) if not comfortable with either (i) or (ii), determine if the model
can be remediated.
In short, we concluded that the mark-to-market loss concept was
creating gross model risk that could invalidate the model from
both a capital adequacy perspective and a capital allocation per-
spective. We therefore jumped to (iii), and found that by replacing
the vendors MtM methodology with one designed to neutralise
the entire mark-to-market loss concept, we could obtain a result that
was sufcientlyconsistent withour internal denitions andbusiness
practices to be useful.
To clarify somewhat, a typical feature of mark-to-market-type
models is that they begin by estimating a current value for each
exposure. They then project the value forward to a specic horizon
under a no-default assumption, and this becomes the loss refer-
ence point (LRP). Future values above the LRP generate no loss,
while future values belowthe LRPgenerate losses. There are twokey
assumptions associated with this approach that we were unwilling
to accept. Firstly, the use of an estimated market value as a starting
point for dening loss implies that past losses/gains have already
been fully recognised and passed through the income statement as
such. A transaction that increased greatly in value since inception
would have embedded gains. But from the models perspective,
these either would have already been realised so that any diminu-
tion in embedded gain will constitute a loss, or were the result of
having purchased the exposure at a premium, so that, again, any
value decrease would constitute a loss.
Secondly, (issues of model risk for MtMaside) the model is based
on the notion that all assets are booked at a value that can be
159
MODEL RISK
described as par with respect to the market price of risk. As
described above, deviations from such a price are either losses or
gains. This may be a useful simplication for a bond fund that pur-
chases liquid assets with minimal transaction costs; but for an insti-
tution that originates transactions, offering specialised products in
illiquid markets, the feespread characteristics of those transactions
will needtocompensate the rmfor transactioncosts, illiquidityand
other risks born by the lender, as well as a premiumthat stems from
the institutional and informational asymmetries that exist in such
markets. These premiums are the reason rms like ours originate
credit exposures in these markets, and represent one of the main
value-creating activities of the rm. However, fromthe models per-
spective, since above market returns have already been realised,
theysimplyincreasethepotential for futurelosses andattract capital.
Finally, this concept of loss leads inexorably to a replacement
concept of capital. By associating loss with market value changes,
capital held to cover potential loss means capital held to purchase
additional assets to ll the gap created by market-value losses in the
risky portfolio. This differs substantially from the concept of cap-
ital held to repay investors in the event of a loss in principal on
assets stemming fromobligor defaults. Two distortions emerge, one
that affects capital adequacy measures and one that affects capital
allocation: when most transactions carry origination premiums like
those described above, the replacement cost typically overstates the
economic loss and (generally speaking) overstates the risk cushion
required for investors given a specied condence level (value-at-
risk (VaR) cut-off); an untoward implication is that for two trans-
actions identical in every characteristic except spread, the one with
the higher spread is viewed as having greater loss potential and
therefore attracts more capital. This creates a perverse incentive for
underwriters if they are aware of potential capital requirements at
the time of underwriting.
Key features and settings for the commercial model
Before considering specic model adjustments, we enumerated fea-
tures that the model would need to exhibit to satisfy both our
intuition and the expressed intuition of senior management. These
related primarily to capital allocation and included the following.
160
(i) Capital should be an increasing function of duration (average
time to repayment).
1
(ii) Size should matter, ie, capital for a given exposure should be
higher than for an exposure identical in every respect except
smaller beyond some reasonable size threshold.
(iii) Diversication benets should be readily interpretable at
least directionally.
The CCM offers a variety of alternative settings designed to allow
the user to change the basic interpretations and assumptions of
the model. Available literature suggested deploying the model in
default only mode (see, for example, Rosen 2004) would neu-
tralise the mark-to-market loss concept by ignoring changes in the
value of assets due to changes in obligors credit worthiness short
of default. The CCM achieves this by setting the value of the assets
at the horizon in the case on non-default (see next paragraph) to be
the same (the average) in any iteration of the Monte Carlo simula-
tion that randomly draws states of the world. However, removing
the credit migration effects means that the transactions maturities
would not play a role and there would be no signicant relationship
between duration and economic capital. This is a violation of the
rst requirement listed above.
Recasting the commercial model: a simple mark-to-par
approach
The model turns out to be extremely sensitive to some of these set-
tings and relatively insensitive to others. From the point of view of
mitigating model risk, it is important to understand which model
parameters and settings lead to large sensitivities in the outputs
and which settings have relatively small consequences. Before dis-
cussing the details of the vended model settings, we briey outline
the primary quantities being modelled, to see if the translation of the
real-life problem to an analytical problem introduces gross model
risk.
The CCM considers the detailed term structure of cashows for
every transaction in the portfolio (the timing and amount of all prin-
cipal and coupon payments), as well as the stochastic characteristics
of eachtransaction(default probability termstructure andloss given
default). Based on these characteristics, the model denes a quantity
161
MODEL RISK
called the value of the transaction at the horizon and assumes that
this quantity is the primary random variable of interest. Although
therearesources of model riskburiedinthis denitionof value, we
think that the concept of transaction value as the primary stochastic
modelling entity is philosophically sound. Furthermore, the essen-
tial methodology of computing the probability distribution of the
transaction value is also sound: each cashow at each point in time
is considered, along with the likelihood of the realisation of that
cashow (based on the term structure of default probability) and,
in the case of default, the loss given default is factored into the cal-
culation. In all these calculations, the time value of money is also
incorporated into the calculation via an appropriately chosen term
structure of risk-free interest rates, which are used for discounting
cashows.
Asource of model risk in the calculation of the probability distri-
butionof transactionvaluearises fromthefact that thevendedmodel
uses a risk-neutral default probability measure that is derived from
the user input physical default probability measure. The vended
model comes with some calibration tools to specify this risk-neutral
default probability measure, but it is nevertheless a source of model
risk because observeddata on illiquidinstruments poses many chal-
lenges. More fundamentally, it is worth questioning whether we
ought to use a risk-neutral default probability measure in the rst
place for a portfolio that is largely illiquid and held to maturity.
Going from value to loss
The CCM denes loss as
L = C V (7.1)
where the constant C is called the loss reference point, V is the value
of the transaction at the horizon and L is the corresponding loss.
At the transaction level, our vended model offers a single choice of
C, namely the expected value of the transaction at horizon, given
non-default. This choice of C corresponds to a rather strict mark-to-
market interpretation of loss; hence, we found this choice of C to
be problematic for the portfolio, as discussed earlier (see page 158).
Here we illustrate the nature of the issue with a very simple, stylised
example.
162
Suppose that we have two transactions that have the following
payouts
V
1
=
_
_
_
90 if default occurs
105 if no default occurs
V
2
=
_
_
_
90 if default occurs
110 if no default occurs
Weassumethat all other characteristics of thetransactions (including
probability of default (PD)) are identical.
So the transactions have identical payouts in case of default, but
V
1
pays less interest than V
2
in the case of non-default. The CCM
sets the loss reference point C = $105 for V
1
, but sets C = $110 for
V
2
. In the case of default, the CCM will assign a loss of $15 to V
1
and a loss of $20 to V
2
. So even though V
2
is always preferable to
V
1
, the CCM assigns higher capital to V
2
and the better deal draws
higher capital. For the vast majority of our portfolio, we want our
risk measurement methodology to strictly respect the monotonicity
property that worse deals (where worse means the values are
lower) draw higher capital.
Within the context of the CCM, we realised that we could not
easily change the methodology for setting the loss reference point
without drastically reworking or completely overhauling the CCM.
Since we wanted to harness the other benets of the CCM (widely
used external model, etc), we decided to apply the mark-to-par
methodology described below.
Operating within constraints of vended model
Given two transactions that are identical except for their spread
characteristics (same commitment, PD, LGD, maturity, correlation
drivers, etc), we would like themto drawthe same capital. Away of
attempting to neutralise the effect of the choice of the loss reference
point, C, is to give both transactions the same spread. This spread
is now no longer the actual spread, but an articial amount that is
designed to equalise both transactions. In theory, this equalisa-
tion shouldworkwithanyspreadvalue as longas bothtransactions
have the same spread. Inpractice, we needa spreadamount that will
not cause anyarticial skews. The questionis: howmuchspread? We
need something that adequately compensates each transaction for
163
MODEL RISK
the default risk (and LGD), while taking into account the maturity
and type of transaction.
Ideally, we would ex ante compute the capital for each transaction
and ensure that the two identical transactions have the same capital
and that if some transaction characteristics are changed, then the
capital moves in the right direction. In practice, we cannot do an
ex ante calculation of transaction level capital (certainly not within
a non-deterministic Monte Carlo simulation model like the vended
one). However, we can calculate a MtMvalue of each transaction by
averaging the cashows of the transaction (using the term structure
of PD, the LGD and considering the time value of money via the
term structure of the risk-free rate). So we could set the spread to
be the amount that ensures that the MtM value of the transaction
equals the exposure amount (equal to the exposure at the analysis
date). We will call the process of computing the spread such that the
MtM value equals par mark-to-par, and will describe this in the
next section.
Within the CCM, interest for a given transaction is specied via
a percentage spread that is used for computing coupon payments
(in addition to any principal payments). The calculation of the MtM
value of a transaction depends on these coupon payments and the
associated principal payments, as well as a host of other variables
(transaction characteristics like PD, LGD, amortisation characteris-
tics, maturity date, dollar commitment amount, term structure of
risk-free interest rates). Although this calculation of the MtM value
is somewhat involved, we can conceptualise it as a process of com-
puting an expectation of the randomvariable specifying transaction
value under a probability measure that accounts for the likelihood
of occurrence of any given transaction value (at the horizon) (this
probability measure depends on the risk-neutral default probability
measure, LGD, as well as the other variables listed above)
MtM= E[V] under a probability measure P
(depending on PD, LGD, etc) (7.2)
In practice, the calculation boils down to a functional relationship
between the transaction characteristics and the consequent MtM
value
MtM= f (PD,LGD, EAD, maturity, spread, . . . , risk-free rate, . . . )
(7.3)
164
The technical problemconsists of nding the spread such that the
MtM value equals par. We solved this problem by implementing
an iterative solver outside the CCM: the idea is that we compute
the MtM value via Equation 7.3 for several different spreads and
choose the spread that brings the MtM value closest to par. In the
course of implementing this solution, we found that we could actu-
ally do much better than a brute force iterative solver. We were able
to explicitly write down the function f in Equation 7.3 so as to make
the dependence of the MtM on the spread transparent, and were
then able to explicitly solve the inverse problem of computing the
spread corresponding to a given MtM.
We found that it was relatively easy to compute the required
spread, even for extremely large portfolios where it is mathemat-
ically impossible to reach par. There are two such cases.
The PD and LGD are so low that the required spreads turn
out to be negative. Although it is helpful to track these trans-
actions, they are probably not a cause for concern from a risk
management standpoint. In this case, we set the spread to zero
(constraint imposed by the CCM).
The PD and LGD are so high that even when the spread is
set to the maximum value (100%), the MtM value is still less
than par. It turned out that these cases were exclusively due to
transactions already in default, i.e. PD = 100%.
Although the steps in specifying and simplifying Equation 7.3 are
somewhat involved, the end result turned out to be an equation of
the form
MtM=
Aspread
2
+Bspread +C
Dspread +E
(7.4)
This tells us that inthe CMMthe MtMis simplya rational function
of the input spread. The constants A, B, C, D and E are complicated
functions of the PD, LGD, etc, but these are easily computed by
a dedicated code (separate from the vended tool). The solution of
the mark-to-par problem then turns out to require the solution of a
quadratic equation for each transaction in the portfolio. In theory, it
is possible to have pairs of imaginary roots, but in practice we found
that all roots were real, and one of the roots always corresponded to
the nancially meaningful spread.
165
MODEL RISK
Figure 7.2 Accuracy of the mark-to-par algorithm
10
7
10
6
10
5
10
4
10
3
10
2
|
M
T
M

p
a
r
|

/
p
a
r
0.2 0.0 0.4
Spread
After computingthe mark-to-par spread, we input this spreadinto
the CMM and ran the tool to check that the MtM value was indeed
close to par. The accuracy of the results is shown in Figure 7.2 for a
test portfolio. Each point on the scatter plot in Figure 7.2 represents a
single transaction; the spreadis plottedonthe horizontal axis andthe
relative error of the MtM value (from par) is plotted on the vertical
axis. We see that, except for the cases where it is mathematically
impossible to achieve par, the relative error (of the MtM from par)
is always less than 10
4
= 1 basis point, and more typically this
relative error is approximately 10
5
= 0. 1bp or lower.
Recovery: a critical modelling question
Whether it uses a simulation-based approach or an analytic ap-
proach, a commercial credit-portfolio model must make certain
choices about howto predict transaction-level recoveries under var-
ious levels of stress as manifest in higher default incidence. The two
key issues to address are:
How should the conditional uncertainty (conditioned on the
level of stress) in recovery be modelled?
How should the uncertainty in recovery be conditioned on
default incidence?
In our case, we answered the second question rst, which simplied
the rst considerably. In general, we have little evidence that LGDs
166
are in fact cyclical (Keenan et al 2008) although we are aware that for
certain products and sub-portfolios strong correlations do exist. As
described above, we tend to hold exposures to maturity and to work
out recoveries on defaulted exposures ourselves, so we are not inter-
ested in modelling recoveries as estimates of trade value under
stressed market conditions. We are therefore comfortable modelling
LGD as stochastic and independent of the default rate. Fortunately,
the CCM features independence between stochastic default and
LGD.
With respect to the rst question, our main model risk mitigant
is the LGD point estimates themselves and the models that produce
them. We maintain a family of over thirty internally developed LGD
models that covers the entire credit portfolio. The diversity of this
model suite is in itself a model risk mitigant, since their errors are
likely to be less than perfectly correlated. Standardised performance
statistics for these models are also maintained (Li et al 2009), and
these give us a direct measure of the uncertainty of point estimates
that can be used to parameterise the portfolio model.
Again, even though we have some condence in our LGD esti-
mates at the transaction level, we need to be sure that the portfolio
model will interpret them in the appropriate way. The issue here is
one of getting denitional clarity on the meaning of recovery in the
CCM, evaluating the model options and then ensuring that it is con-
sistent with the LGD data being input. The issue may be elucidated
by a simple example. Suppose the LGD for a particular transaction
is 20% (and the recovery is 80%). What does the 20% LGD refer to
(ie, 20% of what quantity)? The CCM offers various choices for this
quantity, and the model users are required to pick the right setting.
We found that some of the recovery settings offered within the
CCM were too complicated or indirect to be practically meaningful
and eliminated these at the outset. Of the remaining options, we
selected the setting that best matched our requirements, as follows.
The recovery is a percentage of exposure at the time of default.
The quantity of this exposure had to best match the corre-
sponding quantity (measuring exposure) in the data collec-
tionexercise for the LGDmodels. This data-collectionstepwas
preceded by a carefully written denition of LGD.
We also found that misspecication of the recovery setting can have
a strong impact on important outputs from the CCM. In fact, the
167
MODEL RISK
impact of user misspecication of this setting is so strong that it
couldinvalidate the results of the model, andhence represent amajor
source of gross model risk. Here is a case where the model users have
to be very careful to understand the design of the model builders,
and ensure that the data and interpretation of the model conforms
to its design and extensively test alternative settings.
INTERNAL ALTERNATIVETOTHE COMMERCIAL CREDIT
MODEL
As stated above, the CCM that we have been discussing is an exter-
nal vendor model and it is used to model the loss potential for the
largest sub-component of our overall portfolio. To help mitigate cer-
tain model risks associatedwith using a model that is not fully trans-
parent andto satisfyaninternal desire for anindependent lookat the
commercial credit book, we developed an internal portfolio model
(IPM). The philosophy of the IPMis to recognise the inherent uncer-
tainties in risk modelling and produce a parsimonious, exible and
fast modelling framework. The key for the IPMwas simplicity in the
model structure; we did not intend to duplicate the complexity of
the external CCM, but we did wish to examine the results of a model
that captures the main effects of co-dependency of defaults within
the commercial credit portfolio. The model was also intended to be
fast, so that users could easily run multiple scenarios of large port-
folios in a short time frame. These design requirements of simplicity
and speed led us to take an analytical mathematical approach to
aggregation and a semi-analytical approach to disaggregation. The
point was to stay away frombrute force Monte Carlo simulation and
develop the model outputs in mathematical form to the maximum
extent before performing numerical computation.
The basic assumptions of the IPM are standard in the industry.
The IPM is a default threshold-type model and correlation is driven
by a factor model. We restrict the implementation to a single-factor
model, although the mathematical machinery is more general and
permits extension to a multi-factor case. We felt that it was impor-
tant to examine the effects of a relatively simple, single-factor cor-
relation model before embarking on a complex correlation journey.
We also felt that a comparison of the CCM with the parsimonious
internal model would be helpful, particularly in the light of large
168
Figure 7.3 Comparison of the CCM and IPM loss distributions
P
r
[
l
o
s
s

<

L
]
0.2
0.0
0.4
0.6
0.8
1.0
IPM
CCM (diversified)
CCM (non-diversified)
0 5,000 10,000
L (US$ million)
100,000 obligors.
uncertainties arising from model risk in the problem formulation
and input data uncertainty.
The comparison of credit-portfolio models fromdifferent sources
can be a somewhat intricate process, where inputs and modelling
assumptions have to be carefully harmonised (Gordy 2002; Koylu-
oglu and Hickman 1998). The test results shown in Figure 7.3 were
performed on a specially constructed synthetic test portfolio where
we were careful to harmonise the inputs to insure comparability.
The test portfolio is simplied to a case where all exposures have
a maturity of a single year. The CCM was run under two differ-
ent correlation assumptions (corresponding to a highly diversied
portfolio and less diversied portfolio). The simplied correlation
structure of the IPMdid not permit this distinction, so the IPMmod-
els both these CCM cases via a single distribution. Figure 7.3 shows
that the two models are almost identical in the body of distribution
and show close agreement in the tail, with the IMP results sand-
wiched in-between the two CCM cases. From a model risk perspec-
tive, it is comforting to see such an overall correspondence between
two models having signicantly different approaches (simulation
versus analytical). It also constitutes a sound baseline for further
developments of the IPM, aimed at better tailoring the tool to our
specic needs by removing some of the constraints with the CCM
we described earlier.
169
MODEL RISK
COMBINING LOSS DISTRIBUTIONS FOR DISSIMILAR ASSETS
To the extent that sub-portfolios canbe representedbybothforward-
loss distributions andhistorical loss time series, it is straightforward,
in principle, to combine them into a joint loss distribution using a
copula,
2
where the variancecovariance matrix is determined by the
correlations among the loss time series. This risk-aggregation step
is central not only to the estimation of total risk level but also to the
decision for allocating capital.
It seems almost self-evident that combining sub-portfolio loss dis-
tributions that have been estimated using different models, and in
some cases completely different methodologies, would by itself mit-
igate model risk. The likelihood that these various sub-portfolio
estimation approaches will all err materially in the same direc-
tion would seem, by intuition, to be a decreasing function of the
number of separate methodologies employed. That said, the risk-
mitigating properties of sub-model diversity can easily be over-
whelmedby the model risk inherent in the aggregation assumptions
and techniques. Therefore, these need to be clearly understood and
addressed satisfactorily.
First of all, there is no obvious answer to the very rst question
when using a copula approach: which copula to use for a partic-
ular problem? The Gaussian copula was a favourite of practition-
ers, since it appears through the multivariate central-limit theorem.
However, it has many shortcomings that have been observed and
theoretically studied through its application, one being the lack of
tail dependence, ie, it is not able to capture the event of simultaneous
defaults. Not unreasonably, the application of the Gaussian copula
in modelling credit derivatives is said to have been one of the rea-
sons behind the nancial crisis of 20089 (Salmon 2009). Regardless
of whether this argument is right or wrong, it sharply stresses how
dangerously risky an invalid model can be.
We decide to use the Student-t copula (Demarta andMcNeil 2005)
in our risk-aggregation tool for the following reasons.
It is more general than the Gaussian copula.
The Gaussian copula has asymptotically independent tails,
while the Student-t copula has asymptotically dependent tails.
The concept of tail dependence is crucial tothe model, since the
majority of the nancial data display some tail dependence.
170
It has been shown by many empirical studies that the t copula
performs better in modelling extreme joint losses.
Compared to the Gaussian copula, Students t copula has an extra
parameter, which is the degrees of freedom. Given random vec-
tors x = (x
1
, x
2
, . . . , x
d
)
, whose marginal cumulative distribution

functions are F
1
, F
2
, . . . , F
d
, the t copula can be written as
C
t
,R
(x) =
_
t
1
(F
1
(x
1
))

_
t
1
(F
d
(x
d
))
(
1
2
( +d))
(
1
2
)
_
()
d
R
_
1 +
y
R
1
y
_
(+d)/2
dy
(7.5)
where t
1
v
denotes the quantile function of a standard univariate t
v
function with v degrees of freedom and R is the correlation matrix.
Embrechts et al (2002) discuss in detail methods for empirically
estimating v and R.
We donot claimthat the Student-t copula is the copula touse for
our portfolio. We wouldrather consider it as the part of our quantita-
tive risk assessment toolkit that addresses risk from the perspective
of stress testing. In better understanding the limitations of our mod-
els, we canstress the portfoliobychangingthe value of v. The smaller
the value of v, which potentially represents a risky market, the fatter
the tail and the higher the overall loss level.
Secondly, the estimationof the correlationstructure is not straight-
forward, evenincases where perfect historical data is available. Sug-
gested by academic research,
3
the co-movements of large losses in
nancial markets seem more highly correlated than the assumption
of constant correlation R would predict. Note that we use as the
correlation measure Kendalls tau rank correlation (Kendall 1938)
insteadof Pearsons linear correlation that most people mean by cor-
relation, because Kendalls tau is able to measure non-linear associa-
tions, is invariant to strictly monotonic transformations and is resis-
tant to outliers. If correlation increases for larger losses of the sub-
portfolios as we move further out into the tails of the sub-portfolio
loss distributions, the diversication benets would be dramatically
reduced. It is crucial that our model is able to capture this if it is
indeed present in our portfolio.
There are mainly two types of approach in the literature on mod-
elling conditional correlation: those that condition on time andthose
171
MODEL RISK
Figure 7.4 Conditional versus unconditional loss distribution
0
1991 1996 2001 2006
Rate Mean
0.5
1.0
1.5
2.0
Unconditional
Conditional
R
a
t
e
that condition on the level of the losses. The multivariate Garch-type
models are widelydiscussedandusedas a typical time-conditioning
approach. However, they suffer from two major drawbacks when
appliedto risk modelling: their correlation estimates are often found
to be too variable to use for capital allocation purposes, and they
require relatively high frequency data, which we do not have since
our risk assessment and data collection processes are done on a
quarterly basis.
We choose to use a sort of hybrid approach that combines
the unconditional joint loss distribution, estimated via the copula
approach outlined above, with the univariate time-series models
of the individual loss histories. In this framework, the correlation
wouldessentiallyconditiononthe loss levels inthe previous quarter.
However, it would also indirectly condition on time by denition,
since time is inherent in the observed loss-level data. Graphically,
the conditional loss distribution moves with the credit cycle but has
a narrower support andlower variance, while the unconditional loss
distribution evolves slowly with the accumulation of new data, is
centred around the long-term loss rate, and has a variance dened
both by the cyclical variation and the loss variance conditional on
the cycle state, as described in Figure 7.4.
Mathematically, the following describes the conditional copula.
Let z (m 1) represent a subset of x that we would like to model
with simple AR(p) models and let z
(t)
represent t-quarter lagged
variable z.
UsingAR(1) as anexample, we consider our newrandomvariable
to be x = (x
1
, x
2
, . . . , x
d
, z
1(1)
, z
2(1)
, . . . , z
m(1)
)
. And, we assume x can

172
be modelled by a t copula C
t
,R
F
( x), where
R
F
=
_
R R
12
R
21
R
22
_
R
22
is dened as the subset of Rthat corresponds to z. R
12
= R
21
is
constructedusingtheAR(1) model coefcients. For instance, without
loss of generality, let z
1
x
1
; the best AR(1) model we can get is
z
1(t+1)
r
1
z
1(t)
+ . Then, the rst row of R
12
would be written as
( r
1
, 0, . . . , 0). We do not consider cross-correlations in the R
12
(R
21
)
matrix.
Given the loss level of z = in the previous quarter, the
conditional copula function can be written as
C
t
,
R
(x z = ) =
_
t
1
(F
1
(x
1
))

_
t
1
(F
d
(x
d
))
(
1
2
( v +d))
(
1
2
)
_
( v)
d
_
1 +
(y )

R
1
(y )

_
( v+d)/2
dy
where
= R
12
R
1
22
a,

R =
+a
R
1
22
a
+m
(R R
12
R
1
22
R
21
), = +m
and
a = (t
1
(F
Z
1
(
1
)), . . . , t
1
(F
Zm
(
m
))). (7.6)
With conditional risk models, risk metrics would need different
interpretations and be used differently to address risk. For exam-
ple, in a conditional world, we may be more interested in know-
ing what to expect immediately in terms of the change on the loss
level, rather than in understanding the uncertainty on the expected
loss. Whereas, in the unconditional normal market, which is pre-
sumably represented by the through-the-cycle unconditional corre-
lation estimation, expected loss would not be commonly considered
as alternative risk metrics to VaR, expected shortfall, etc.
As an aside, this framework can be used for macro stress test.
Of course, for that purpose, R
F
would be estimated fromreasonably
synchronisedportfolio historical losses andmacro factor time series.
Lastly, the simulation engine behind the copula framework to
some extent addresses model risk in a Bayesian manner. Let the vec-
tor denote the model input parameters andlet the joint probability
distribution be written g
, where g
could be estimated empirically,

173
MODEL RISK
or constructed from non-informative priors. R and v all belong to
. Even copulas can be considered as part of . The uncertainty
in the input parameters can be studied by simulation, namely, by
generating N samples of plausible values from g
and running the

portfolio tool once for each of these samples. The relevant risk met-
rics (eg, VaR) conditional on the sampled parameter value

i
(for
i = 1, . . . , N), as well as the conditional loss distribution f
x
i
, can be
estimated accordingly. For example, to calculate the uncertainty on
VaR, we can simply tabulate the N values of VaR fromthe N simula-
tions and obtain a 90%uncertainty interval. On the other hand, after
understandingvariationina riskmetric, we might still desire tohave
a single estimate that incorporates all uncertainties. For this pur-
pose, we introduce the concept of a predictive risk measure. Instead
of generating a total loss distribution for each of the N simulations,
we calculate a single overall loss distribution from all the simulated
scenarios in all the N simulations. All the samples in this simulation
are from
f
x
=
N
_
i1
g
i
f
x
i
All risk metrics can then be drawn fromthis single loss distribution,
which addresses the model uncertainties considered.
REDUCTION AND PROVIDING CONTEXT FOR MODEL
OUTPUTS
Because of the complexity of the model suite, and because the nal
outputs of a joint loss distribution and allocated capital are difcult
to interpret in intuitive risk terms, it is helpful to be able to re-
express the portfolio model in a reduced formmeta model. In our
case, we treated the transaction level capital rate as the dependant
variable and applied a spline regression to approximate the output
of the portfolio model.
As shown in Figure 7.5, the spline regression model is capable of
tting the data quite well with an R
2
value of over 96% in this case.
We do expect that a exible, non-parametric model will t the data
well, since we expect the relationships embedded in the portfolio
to be mostly smooth and quasi-concave in nature. Oddly shaped
or excessively non-monotonic relationships are not ruled out of the
metamodel a priori; however, since most stable economic andnatural
relationships are not like that, the regression model offers us a view
174
Figure 7.5 Capital rates: portfolio model versus meta model
10
5
10
5
10
3
10
1
10
3
10
1
P
r
e
d
i
c
t
e
d
Observed
Figure 7.6 Model implied capital rates versus PD by maturity band
0.00
0.00
0.02
0.04
0.05
0.06
0.08
0.10
0.10 0.15 0.20 0.25 0.30
C
a
p
i
t
a
l

r
a
t
e
PD
1 yr maturity
3 yr maturity
7 yr maturity
of the portfolio model with some level of simplicity imposed. There-
fore, a good t is already a model risk mitigant, since it ensures that
the model results are not being driven by discontinuous or poorly
175
MODEL RISK
behaved functions, which would suggest instability of the portfolio
model. Also, to the extent that simulation noise is present in the
portfolio model results (even at low levels), the regression is apt
to eliminate this providing a slightly cleaner look at some of the
underlying relationships. Here, the regression model can highlight
the economics in a way that the portfolio model cannot, and allows
users to evaluate marginal relationships to see if they accord with
intuition.
For example, it wouldaccordwith intuition if, all else being equal,
riskier transactions attracted more capital than safer ones. But with
respect to a key risk driver such as PD, what does this relationship
look like? Does it look different for different portfolio segments?
Such questions are easily answered with the reduced formequation
in hand. As shown in Figure 7.6, capital rate curves by PD, seg-
mented by maturity, give the analyst and non-quantitative manager
something relatively easy to grasp and juxtapose against intuition.
The tted capital-ratePD relationship is plotted for one-, three-
and seven-year maturity transactions along with 95% condence
bands. As expected, the portfolio model does imply a monotoni-
cally increasing capital rate with transaction PD for all three curves.
Furthermore, for maturities of less than three years the capital rate
increases quickly as a function of maturity for any given PD. How-
ever, there is also an apparent interaction between PD and maturity
that affects longer maturity, lower rated transactions: for PD larger
than roughly 3%, moving from three years to seven years maturity
makes relatively little difference to the capital rate assigned by the
portfolio model.
The condence bands in Figure 7.6 widen or narrowbased on the
amount of data available. For instance, the condence bands on the
seven-year maturity curve widen for a PDgreater than 10%because
there is less than 5%exposure in this region. All curves shown in the
gure control for the other transaction characteristics (eg, LGD) at
their exposure-weighted averages.
The meta model can be translated easily into rules of thumb for
managerial consumption. For example, consider a one-year/5% PD
transaction. This transaction attracts roughly 1.5% capital rate in
the gure. Increasing the maturity on this transaction to three years
would attract slightly more than double the capital. On the other
hand, to attract the same increased capital while holding constant
176
Figure 7.7 US public company default rates
0.01 0.02 0.03 0.04 0.05
5
10
15
20
25
30
0
M
o
n
t
h
s

o
b
s
e
r
v
e
d
Trailing 12-month default rate
Source: GE Capital proprietary database.
the one-year maturity, the obligor wouldneedtheir PDdowngraded
from 5% to roughly 15%.
Obviously, a true validation of a probabilistic forward-loss dis-
tribution is infeasible. Observations of portfolio loss will tend to
cluster around the mean. Large deviations may occur from time to
time but, while the model associates various loss levels with prob-
abilities, empirical observations provide us with only the rst part,
leaving a thorny estimation problem that blocks us from validating
the model. How likely was the 20089 credit market crisis? How
likely was the 2001 cycle? Or, more in the spirit of economic capital
calculations, how likely is a credit crisis three times worse than the
20089 crisis?
The association of extreme tail losses with probabilities and with
describable economic and political scenarios is neither art nor sci-
ence, but simply guesswork. However, the use of a probability dis-
tribution to describe future losses can be loosely supported by inter-
preting key historical loss drivers as probabilistic outcomes andcon-
sideringthe model-basedloss distributioninthe context of the distri-
butionof the historical driver. There are twokeysteps toforgingsuch
a connection between probabilistic model outputs and observed
(historical) reality:
the analyst must develop a probabilistic view of frequency-
based historical data;
177
MODEL RISK
Figure 7.8 US public company default rates by sector
35
30
25
20
15
10
5
0
P
e
r
c
e
n
t
J
a
n

9
6
J
a
n

9
7
J
a
n

9
8
J
a
n

9
9
J
a
n

0
0
J
a
n

0
1
J
a
n

0
2
J
a
n

0
3
J
a
n

0
4
J
a
n

0
5
J
a
n

0
6
J
a
n

0
7
J
a
n

0
8
J
a
n

0
9
Air
Basic industries
Financial
Retail
Auto
Energy
Manufacturing
Telecom
probabilistic model outputs must be converted or disaggre-
gated into frequency-based analogues.
For example, Figure 7.7 presents a histogramof trailing12-monthUS
public company default rates from 1999 to 2009 Q2. We could view
these observations as representing draws froma particular probabil-
ity distribution. If we take this approach, which may include tting
a parametric distribution to the data, we have establisheda stylised
fact, against which portfolio model outputs may be compared. If,
for example, the estimated likelihood of exceeding a 5%default rate
is less than 1% based on the distribution tted directly to the empir-
ical data, than the default rates associated with portfolio model loss
levels below the 99th percentile should tend to be at or below 5%.
Adifculty with this type of model check arises when the portfolio
model cannot output descriptive statistics, such as implied default
rates, for different loss-probability ranges. As described earlier (see
page 168), we developed our internal commercial model in part so
we could engineer it to retain and report such descriptive statistics,
to provide some contextual link between high loss levels and the
likelihood that such loss levels will be reached.
Such contextual outputs can also shed light on the degree of
abstraction underlying various model specications to help gauge
whether or not they are reasonable or acceptable. For example, while
little has been established to conrmor reject popular approaches to
default correlation modelling, basic stylisedfacts do exist. Parsing
178
the data presented in Figure 7.8 by sector shows the extent to which
corporate default rates move together across credit cycles. We can
certainly presume that extreme tail events will lead to extreme lev-
els of default correlation. But are the portfolio model results even
roughly consistent with the observed co-movement of default rates
by sector for stress levels up to those associated with the most
severe historical credit cycles? Until this question can be answered
in the afrmative, gross model risk for this class of model is still
dangerously high.
CONCLUSIONS
For nancial institutions with regulatory exposure, quantitative risk
analytics has become increasingly important, particularly in the
area of capital adequacy. Portfolio loss models are now expected to
play a role in capital adequacy assessment (and capital allocation),
although managers and regulators maintain a healthy scepticism
about the relevance andreliabilityof outputs fromcomplex portfolio
models. This tension has elevated the requirement for modellers to
address model risk as part of the model development process.
In our view, managing model risk in the portfolio model context
has certain distinct components. Of primary importance is the clear
understanding of the functional dynamics of each model compo-
nent, anda matchingof those functional dynamics withthe structure
and functional relationships of the business component being mod-
elled. Where these do not match well, explicit workarounds need
to be created, vetted and subjected to their own model risk assess-
ments. The issue of matching model design to business process is
particularlyacute whenusingcommerciallyprovidedsoftware solu-
tions, and extra care should be taken to study every feature of such
models.
Triangulation can be helpful. Where possible, analogous models
that performthe same estimation in different ways can help to iden-
tify model idiosyncrasies. Again, this is particularly helpful when
an institution has an analytic dependency on a third-party-provided
software solution.
Reduction, in the form of re-estimating a complex model using a
regression or similar approach can be extremely valuable, both as a
check on intuition and as a communication device. Merely demon-
strating that a complex model suite is primarily exhibiting a few
179
MODEL RISK
core, material and intuitive relationships helps demystify the model
and frame the basic marginal risk drivers in simple terms. Such esti-
mating equations can also be helpful in deal evaluation, or in other
contexts when deploying the entire portfolio model is impractical or
infeasible.
Aggregation of risk components using a copula is now nearly
standard practice, and the selection of a copula function has been
heavily discussed in the literature. That said, when used for predic-
tion, other model-risk aspects of a copula-based joint distribution
estimation are generally analogous to model-risk issues in regres-
sion. The quantities being modelled may be various representations
or transformations of the empirical data available, and these are
choices to be made by the modeller. For us, the decision to include in
the copula certain autoregressive components as well as extraneous
macroeconomic factors was based on our desire to estimate the con-
ditional joint loss distribution. This reects the fact that our historical
losses have exhibitedautocorrelationas well as some measure of cor-
relation with certain macroeconomic variables. Lack of robustness
occurs whentransitoryhistorical phenomena or noise are embedded
in predictions, leading to forecast error. Here again, there is a direct
analogy with regression. In the regression context there are a wide
range of robust methods basedon various sub-sampling schemes. In
the copula context, robustness must be obtainedthrough the process
of estimating the variancecovariance matrix. Since, given our data
limitations, sub-sampling can play only a limited direct role, we use
a combination of intuition and brute force to condition the variance
covariance matrix directly. Putting some limitations on the amount
of diversication across asset or product types helps to build-in con-
servatism in a way that can provide some additional comfort to
sceptical managers and supervisors.
Model outputs must be evaluated against current and historical
data that can provide context. Except perhaps for a neighbourhood
around the mean, probabilistic loss distributions are challenging to
evaluate on their own. Examining implied default rates as a func-
tion of implied likelihood of loss is a good way of establishing con-
text for a commercial loss model. Comparing the pattern of implied
default rates by sector as a functionof impliedlikelihoodof loss with
observed patterns of default/loss by sector can be a good way of
gauging the efcacy of the embedded correlation model. Certainly,
180
the ability to provide granular descriptions of the underlying cir-
cumstances associated with a probabilistic loss value is key to cre-
ating a dialogue between risk managers and model managers, and
to building condence that the model suite is making sense. From
there it may be a small leap to a two-way dialogue in which mangers
hypothesise stress scenarios described in their own terms, and mod-
elling teams can use the model suite to assign specic loss estimates
to these scenarios.
Finally, it should be noted that, in an academic setting, a model
may be simply mathematics that can be written down on a page
or two. Within a nancial institution, a model requires mathematics,
numerical algorithms, data processing code, data systems and hard-
ware, and people trained to create and assemble all of these parts.
The greatest single mitigant of model risk is a culture in which a
heightened sensitivity to model risk is shared across the risk team.
The opinions expressed herein are those of the authors alone and
do not necessarily reect the opinions of GE Capital.
1 Academic literature is surprisingly sparse on this topic; see Gurtler and Hethecker (2008).
2 See, for example, Cherubini et al (2004).
3 See, for example, Anderson et al (2001) and Solnik et al (1996).
REFERENCES
Andersen, T. G., T. Bollerslev, F. X. Diebold and H. Ebens, 2001, The Distribution of
Realized Stock Return Volatility, Journal of Financial Economics 61, pp. 4376.
Cherubini, U., E. Luciano and W. Vecchiato, 2004, Copula Methods in Finance (New York:
John Wiley & Sons).
Demarta, S., and A. J. McNeil, 2005, The t Copula and Related Copulas, International
Statistical Review 73(1), pp. 11129.
Embrechts, P., A. McNeil and D. Straumann, 2002, Correlation and Dependence in Risk
Management: Properties and Pitfalls, in M. A. H. Dempster (ed.), Risk Management: Value
at Risk and Beyond, pp. 176223 (Cambridge University Press).
Gordy, M. B., 2000, AComparative Anatomy of Credit Risk Models, Journal of Banking
and Finance 24 (December), pp. 11949.
Gurtler, M., and D. Hethecker, 2008, Multi-Period Defaults and Maturity Effects on
Economic Capital in a Ratings-Based Default-Mode Model, Die Betriebswirtschaft 68,
pp. 50124.
Keenan, S., D. Li, S. Santilli, A. Barnes, K. ChalermkraivuthandR. Neagu, 2008, Credit-
Cycle Stress Testing Using a Point-in-Time Rating System, in H. Scheule and D. Rosch
(eds) Stress Testing for Financial Institutions, Chapter 3 (London: Risk Books).
181
MODEL RISK
Kendall, M., 1938, ANew Measure of Rank Correlation, Biometrika 30, pp. 819.
Koyluoglu, H. U., and A. Hickman, 1998, A Generalized Framework for Credit Risk
Portfolio Models, Research Paper, Credit Suisse Financial Products.
Li, D., R. Bhariok, S. Santilli and S. Keenan, 2009, A Common Validation Framework
for Loss Given Default Models, The Journal of Risk Model Validation, 3(3).
Rosen, D., 2004, Credit Risk Capital Calculation, in PRM Handbook, Volume III,
Chapter III.B.6, URL: http://prmia.org.
Salmon, F., 2009, Recipe for Disaster: The Formula That Killed Wall Street, Wired
Magazine, URK: http://www.wired.com.
Solnik, B., C. Boucrelle and Y. F. Le, 1996, International Market Correlation and
Volatility, Financial Analysts Journal 52, pp. 734.
182
8
Transmission of Macro Shocks to
Loan Losses in a Deep Crisis:
The Case of Finland
Esa Jokivuolle; Matti Virn; Oskari Vhmaa
Bank of Finland; University of Turku and Bank of Finland;
University of Turku
Understanding the sources of corporate credit losses continues to
lie at the heart of commercial banks risk management as well as
macro-prudential analysis conducted by nancial authorities. In the
aftermath of the nancial crisis of 20089, worldwide weakened
economic growth prospects threaten to increase corporate defaults
and credit losses, which may further burden the already troubled
banking sector and impair their lending ability.
In analysing the link between macro economy and corporate
credit losses it is useful to look at earlier historic episodes when
credit losses have occurred. A recent case in point is the Finnish
banking crisis and great depression in the early 1990s, which have
been analysed by, for example, Conesa et al (2007), Gorodnichenko
et al (2009) and Honkapohja et al (2009). The size of the Finnish cri-
sis was exceptional even by international standards (Reinhart and
Rogoff 2008). The share of non-performing loans went up to 13%
and annual default rates rose to 3%, which resulted in high loan
losses (see Figures 8.1 and 8.2). The crisis cost the Finnish govern-
ment almost 13%of annual GDP(Laeven and Valencia 2008; see also
Figure 8.1). Since the depression, economic development has been
very favourable and annual default rates have come down to less
than 1%, so that loan losses have been almost non-existent.
1
Some interestingobservations regardingthe relationshipbetween
defaults and loan losses can be made just by visual inspection of the
Finnish evidence. In particular, the defaultsloan-losses relationship
appears at least to some extent to be non-linear (Figure 8.2). Even
183
MODEL RISK
Figure 8.1 Industry-specic default rates
0.000
0.002
0.004
0.006
0.008
0.010
0.012
0.014
1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008
Agriculture
Manufacturing
Construction
Trade
Transportation
Other
D
e
f
a
u
l
t

r
a
t
e
Quarterly data. Source: Statistics Finland.
Figure 8.2 Relationship between loan losses and the aggregate default
rate
3,000
2,500
2,000
1,500
1,000
500
0
500
L
o
s
s
e
s
Losses
Default rate
D
e
f
a
u
l
t

r
a
t
e
0.0
0.2
0.4
0.6
0.8
1.0
1
9
8
8
1
9
9
0
1
9
9
2
1
9
9
4
1
9
9
6
1
9
9
8
2
0
0
0
2
0
0
2
2
0
0
4
2
0
0
6
2
0
0
8
The (seasonally adjusted) default rate corresponds to the whole economy. Loan
losses are interpolated from an annual Statistics Finland series.
184
TRANSMISSION OF MACRO SHOCKSTO LOAN LOSSES IN A DEEP CRISIS
in good times there are a fair few business failures (in the Finnish
data, the annual default rate has not gone below 0.8%) but they
do not seem to have caused many loan losses. There are several
possible reasons for this regularity. First, some fraction of newrms
will always fail, eg, because of entrepreneurial incompetence or the
initial lackof resources or demand. Most suchrms are fortunately
small, so that the effect on banks loan losses is limited. The second
reason is probably related to collateral values. In good times loan-to-
value ratios are reasonably lowand thus most losses can be covered
by collateral. In deep recessions things are different: some very big
rms also fail and falling market values of collateral reinforce the
negative impact.
2
Thus, loan losses tend to be severe only in deep
recessions or depressions. That is why, in this chapter, we also wish
to further analyse the behaviour of corporate credit losses in deep
and long-lasting recessions and depressions.
Our work follows the branch of literature that has focused on the
transmission of macro shocks, notably output, real interest rate and
aggregate corporate indebtedness, to corporate failures and banks
loan losses. We adopt the framework of Sorge and Virolainen (2006),
which rst models industry-specic corporate default rates with
the macro variables and then simulates loan losses by using the
industry-specic default rates as proxies for corporate probabilities
of default (PD) in the respective industries and by assuming a con-
stant loss given default (LGD) across all companies. Our aim is to
extend their analysis of the Finnish case in several ways. Firstly, we
consider the following extensions to their model.
We allow LGD to depend on the state of the business cycle.
We consider industry-specic output shocks instead of the
aggregate shock. We also take a step back and make a basic
comparison by running a single-equation model for the aggre-
gate corporate default rate, in order to investigate the impor-
tance of disaggregating to the industry-specic default rates
in the rst place.
We try an alternative times-series specication of the macro
shocks.
As inthe Sorge andVirolainen(2006) model, we thenperformMonte
Carlo simulations of the macro-based loan-loss model to produce
185
MODEL RISK
loan-loss distributions. We investigate how the above model exten-
sions affect these distributions. Taken together, these results con-
stitute a set of robustness checks to the basic model of Sorge and
Virolainen (2006), which may provide some guidance as to what
may be the most crucial areas of model risk in the basic model.
The most important result here concerns the LGD, endogenous to
the business-cycle state, which we measure as the annual average
of the entire corporate sector and which we are able to estimate by
using aggregate data on the number of defaults, loan losses and the
distribution of corporate debt within the corporate sector.
Secondly, we extend loan-loss scenario analyses to further study
the effects of a deep, prolonged recession. Interestingly, it appears
that a constant GDP shock that persists over several years has a
slightly convex effect on cumulative loan losses. This non-linear
effect is further reinforced when we add a simplistic feedback mech-
anism from loan losses to the GDP growth. We also demonstrate a
non-linear effect from macro shocks to loan losses, in that differ-
ent simultaneous shocks appear to reinforce each anothers effects
on loan losses versus an individual macro shock of commensurate
size.
Thirdly, we consider the impact on potential aggregate loan losses
of single dominant industrial cluster, for which Finland also pro-
vides an interesting case as a result of the central role in its economy
of the newlydevelopedinformationandcommunicationtechnology
(ICT) cluster.
Fourthly, we complement the analysis of Sorge and Virolainen
(2006) by providing the t of the macro-based model of loan losses
tothe actual loanlosses experiencedinFinland, especiallyduringthe
crisis years of the early 1990s. We ndthat, althoughthe endogenous
LGDseems an important improvement to the t relative to the basic
model with a constant LGD, the model still falls short of capturing
the full magnitude of loan losses experienced in Finland during the
crisis. We discuss potential explanations for the remaining gap.
Lastly, we compute the loan-loss distribution both for a pre-1990s
crisis periodandthe most recent periodinour data. This comparison
clearly shows the very signicant impact that the prevailing macro
state has on the conditional loan-loss distribution. It is remarkable
that in 1990 expected aggregate loan losses in Finland were roughly
double what they were in the third quarter of 2008.
186
The rest of the chapter is organised as follows. The next section
describes the industry-specic default-rate model estimation. We
then discuss the estimation results (see page 192 onwards). Next,
we simulate the loan-loss distribution andconsider the effects on the
loss distribution of the various extensions to the basic model. Fur-
thermore, we investigate a scenario of a prolonged deep recession
and a scenario of simultaneous shocks. We also provide the empir-
ical t of the model-based expected loan losses in Finland over the
sample period. Finally, we contrast the pre-1990s crisis loan-loss dis-
tribution with the one simulated with the most recent data before
making some concluding remarks.
ESTIMATION PROCEDURE
The estimated model of industry-specic default rates is in essence
the same as that in Sorge and Virolainen (2006) except that we now
allow industry-specic output (output gap or output growth rate)
to affect the corresponding industry-specic default rate and use
industry-specic real interest rates, which reect industry-specic
ination rates. The level of indebtedness is also industry specic.
Thus, the estimated equation is of the following form
d
it
=
0
+
1
y
it
+
2
r
it
+
3
l
it
+u
it
(8.1)
where d
i
denotes the default rate in industry i (agriculture, con-
struction, manufacturing, trade, transportation and other services),
y denotes private sector output (henceforthoutput, for simplicity),
r denotes the real interest rate and l denotes the indebtedness level.
The estimation period is 1987 Q1 to 2007 Q4. Estimation results for
Equation 8.1 for different output variables are reported in Table 8.1.
In Figure 8.3 we compare the single-equation and the system form
(seeminglyunrelatedregression(SUR)) estimationresults inthe case
of the basic specication where output gap is the output variable. In
Table 8.2 we report diagnostic test results for this basic specication.
The default-rate model (Equation 8.1) provides us with the basis
for evaluating the impact of macroeconomic shocks on corporate
defaults and further on banks loan losses. To obtain empirical coun-
terparts for the macro shocks we use alternative specications. To
begin with, we follow Sorge and Virolainen (2006) and estimate an
AR(2) process for y, r and l to lter out the shocks as the residuals
187
M
O
D
E
L
R
I
S
K
Table 8.1 Estimation results of the basic default-rate model for the various industries
Agr Man Con Trd Trns Oth Tot
Constant 0.002 0.000 0.003 0.001 0.002 0.001 0.000
t-value (4.05) (0.31) (16.87) (4.78) (11.59) (0.90) (0.94)
total output gap 0.008 0.043 0.015 0.034 0.011 0.004 0.032
t-value (2.99) (6.91) (3.61) (6.05) (3.41) (0.55) (6.07)
Interest rate (real) 0.000 0.000 0.000 0.000 0.000 0.001 0.000
t-value (0.16) (0.20) (1.16) (1.41) (0.26) (10.71) (0.47)
Debt 0.025 0.210 0.099 0.141 0.005 0.096 0.235
t-value (1.99) (19.916) (13.710) (18.577) (0.320) (0.973) (9.668)
R
2
0.121 0.891 0.856 0.929 0.175 0.809 0.873
SEE 0.425 1.082 0.670 0.781 0.533 1.092 0.789
DW 2.020 1.308 1.451 1.141 1.497 0.982 0.495
Agr, agriculture; Man, manufacturing; Con, construction; Trd, Trade; Trns, transportation; Oth, other; Tot, all industries.
When testing the coefcient restriction that the coefcients are equal for all sectors, the F-statistics turns out to be 6,03, which
is signicant at all conventional signicance levels. In comparison, the sum of squares of the latter system turned out to be 16%
larger.
1
8
8
T
R
A
N
S
M
I
S
S
I
O
N
O
F
M
A
C
R
O
S
H
O
C
K
S
T
O
L
O
A
N
L
O
S
S
E
S
I
N
A
D
E
E
P
C
R
I
S
I
S
Table 8.1 (Cont.) Estimation results of the basic default-rate model for the various industries
Constant 0.002 0.000 0.003 0.002 0.002 0.001 .
t-value (2.98) (0.58) (15.84) (5.78) (11.21) (0.87) .
industry-specic output gap 0.000 0.021 0.004 0.027 0.011 0.004 .
t-value (0.01) (5.09) (2.60) (9.42) (3.92) (0.35) .
Interest rate (real) 0.000 0.000 0.000 0.000 0.000 0.001 .
t-value (0.71) (0.48) (2.56) (0.76) (0.15) (10.92) .
Debt 0.014 0.209 0.097 0.119 0.006 0.098 .
t-value (1.05) (18.35) (12.06) (13.31) (0.30) (0.99) .
R
2
0.017 0.878 0.844 0.882 0.185 0.808 .
SEE 0.446 1.152 0.696 0.975 0.562 1.093 .
DW 1.816 0.999 1.315 0.828 1.395 0.997 .
1
8
9
M
O
D
E
L
R
I
S
K
Table 8.1 (Cont.) Estimation results of the basic default-rate model for the various industries
Constant 0.002 0.001 0.003 0.001 0.002 0.002 0.000
t-value (3.11) (1.34) (13.80) (1.79) (6.72) (1.24) (0.32)
growth rate of total output 0.001 0.018 0.009 0.006 0.000 0.005 0.007
t-value (0.45) (2.92) (2.75) (1.33) (0.11) (0.91) (1.30)
Interest rate (real) 0.000 0.000 0.000 0.000 0.000 0.001 0.000
t-value (0.40) (0.16) (0.22) (0.33) (0.25) (11.72) (0.48)
Debt 0.014 0.194 0.100 0.167 0.007 0.040 0.214
t-value (1.11) (12.89) (13.35) (14.82) (0.28) (0.34) (6.38)
R
2
0.020 0.851 0.846 0.893 0.009 0.810 0.719
SEE 0.446 1.273 0.692 0.961 0.620 1.088 1.158
DW 1.821 0.910 1.304 0.819 1.093 1.025 0.222
1
9
0
T
R
A
N
S
M
I
S
S
I
O
N
O
F
M
A
C
R
O
S
H
O
C
K
S
T
O
L
O
A
N
L
O
S
S
E
S
I
N
A
D
E
E
P
C
R
I
S
I
S
Table 8.2 Diagnostic tests
Correlogram Squared Residuals 7.700 2.254 0.914 4.361 2.530 7.710 13.131
Prob (0.103) (0.689) (0.923) (0.359) (0.639) (0.103) (0.011)
Serial Correlation LM Test 0.546 6.045 3.959 5.140 4.041 8.024 25.530
Prob (0.702) (0.000) (0.006) (0.001) (0.006) (0.000) (0.000)
Normality Test (Jarque Bera) 0.491 15.513 1.204 1.834 1.494 3.795 0.546
Prob (0.782) (0.000) (0.548) (0.400) (0.474) (0.150) (0.761)
Heteroscedasticity test (BPG): 1.477 3.724 3.157 7.585 0.925 0.800 3.183
Prob (0.228) (0.015) (0.030) (0.000) (0.433) (0.498) (0.029)
White 1.548 2.130 1.472 4.035 0.509 2.071 2.286
Prob (0.220) (0.039) (0.177) (0.000) (0.862) (0.045) (0.027)
Chow test (1999 Q1) 0.444 2.920 2.159 4.769 6.834 13.344 14.062
Prob (0.643) (0.027) (0.083) (0.002) (0.000) (0.000) (0.000)
QuandtAndrews test 1.007 5.650 3.165 5.548 6.090 9.743 17.140
Prob (0.778) (0.165) (0.613) (0.176) (0.126) (0.011) (0.000)
Chow forecast test (2003 Q1) 1.440 0.468 1.244 0.730 1.024 2.084 2.935
Prob (0.145) (0.967) (0.259) (0.775) (0.454) (0.018) (0.001)
Ramsey RESET test 1.380 2.855 3.335 2.023 5.911 4.667 3.320
Prob (0.258) (0.064) (0.041) (0.141) (0.004) (0.013) (0.042)
CUSUM + + + + +
CUSUM
2
+ + +
Recursive coefcient of GDP + + + + + +
Industry types dened as inTable 8.1. CUSUM(+) means that the assumption of coefcient stability cannot be rejected. Accordingly,
recursive coefcient (+) means that the coefcient of output gaps seems to be stable over time.
1
9
1
MODEL RISK
of the AR(2)s. Alternatively, we use a simple random-walk repre-
sentation for the time series of these variables. On the basis of the
estimated shocks we can compute the variancecovariance matrix
of the macroeconomic shocks that is needed in sampling defaults
and thus loan losses from the rm-level micro data.
Inorder tocarryout the samplingprocedure properly, ie, toensure
that default rates are between 0 and 1, a logistic transformation is
needed for the default rate p = (1/(1 + exp(d)) in estimating and
simulating the model (Sorge and Virolainen 2006). These values can
then be conveniently transformed to original default rates by using
the natural logarithmic transformation d = ln((1 p)/p).
When using Monte Carlo methods in simulating the impact of
macroeconomic shocks on loan losses we use the assumption that
all random elements are normally distributed. The expected values
of the macroeconomic shocks are assumed to be zero in the basic
scenarios. When dealing with the depression scenarios in the sec-
tion on prolonged recession and simultaneous shocks (see page 199
onwards), however, we change this assumption by introducing sys-
tematically negative values for output growth or, correspondingly,
more positive values for real interest rates and aggregate corporate
indebtedness.
In the Monte Carlo simulation, the rst step was a Cholesky
transformation of the variancecovariance matrix of the stochastic
terms. Although correlations between these terms were not overly
high, ordering of variables (shocks) turned out to be important (see
page 203).
ESTIMATION RESULTS
In this section we comment on our estimation results presented in
Table 8.1. Initially, the model ts the data onindustry-specic default
rates reasonablywell. It is onlyinthe case of agriculture, andto some
extent other services, that there are some problems. Thus, in the case
of default rates in agriculture we are not able to obtain coefcients of
the correct sign for the real interest rate and indebtedness. We sus-
pect that this failure reects some data problems: it is quite difcult
to distinguish family (household) farming and industrial farming.
3
Other services is also a bit difcult to quantify because it really
represents a mix of various activities. It was thus no big surprise to
us that the main diagnostic problems, related to stability, appear in
192
Figure 8.3 Comparison of OLS and SUR estimates of output gap for
different sectors
Agriculture
Manufacturing
Construction
Trade
Transportation
Other
SUR
OLS
0.05 0.04 0.03 0.02 0.01 0
this sector. Even so, the coefcient estimates make sense in addition
to the simulation results in the next section.
Interestingly, it may be seen that the data quite strongly favours
the disaggregated model. When we just focus on the aggregate
default-rate model (Tot in the last column of Table 8.1) and use
the corresponding aggregate equation, several diagnostic problems
arise. In particular, the stability properties of the aggregate equation
seem dubious (Table 8.2). Moreover, the explanatory power seems
to suffer from aggregation, though not dramatically.
It is possibletogodeeper intothedisaggregatedstructurebyusing
industry-specic output instead of aggregate output in explaining
each industry-specic default rate (the second part of Table 8.1; see
page 189) but it appears that the gain is not signicant (in either
direction). This notion is conrmed when we carry out the loan-
loss simulations in the next session and see that essentially the same
results are obtained for both aggregate output and industry-specic
output. Obviously, the industry-specic output model is required
when we want to examine the role of sector-specic shocks (see
page 202 onwards). Finally, we may note that the output measure of
the basic specication, the HodrickPrescott measure of output gap,
performs much better than just the output growth. Hence, we retain
this measure.
The models are rst estimated with a basic ordinary-least-squares
(OLS) approachwhichmay be subject to certainwell-knownpitfalls.
First of all, the right-hand side variables may depend on the default
193
MODEL RISK
rates, creating a classical simultaneity problem. The problem is par-
ticularly relevant for output, but not so relevant for interest rates or
indebtedness because these variables have been lagged in the nal
estimation specication. As for output, corporate failures probably
affect output by destroying productive capacity and rm-specic
human capital. The bias is probably smaller when we use aggregate
output instead of industry-specic output, but it is hard to say more
on the magnitude of this bias without proper instruments. In one
of our loan-loss simulations, we allow for a simple feedback mech-
anism, meant to capture the effect of defaults on output. Another
problem with OLS estimation is related to correlation of residuals
of equations for different industries. To see the effect, we have used
the more efcient SUR estimator instead of OLS (see Figure 8.3).
Overall, differences in the estimated coefcients do not appear to be
very large, so we use the OLS estimated coefcients in the simula-
tion exercise in the next section. This choice is also supported by the
notion that the data for agriculture may contain some deciencies
that could contaminate parameter estimates of other sectors equa-
tions in the SUR estimation. The important message of Figure 8.3
is perhaps not the fact that the two estimators produce similar esti-
mates but the fact that the coefcients for different sectors are indeed
very different, and hence it seems necessary to have a disaggregated
model when evaluating loan-loss risks with simulations.
To sum up results from diagnostic tests, it seems that the basic
model for industry-specic default rates works quite well. In partic-
ular, we may notice that the stability properties of the equations are
rather good in spite of the huge changes in the default rates which
took place in the early 1990s. This also shows up in recursive esti-
mates of the coefcient of the output gap, although in the interest of
brevity we do not report the results here.
4
On estimating the endogenous LGD
It is well known that LGD is not a constant across defaulted loans,
and that it tends to increase in economic downturns (see, for exam-
ple, Schuermann (2004) and the literature surveyed therein). We
therefore relax the assumption of the basic model of a constant LGD
and replace it with a time-varying annual LGD that is a function
of the state of the business cycle, measured against output gap. In
other words, we assume that the LGD is the same for all defaulted
194
loans in a given year, but it can vary from year to year according to
the business cycle. If the expected annual LGD equals the constant
LGD of the basic model, then we should expect a fatter tail to the
loan-loss distribution, given that both PDs and the annual LGD are
decreasing functions of the output gap.
Because we do not have data from Finland on individual loans
loss given default, we have estimated the annual average LGDfrom
aggregate data by using a method based on random sampling. We
make use of the following equality, in which we assume that LGD
is a constant in a given year t
total loan losses
t
=
N
t
_
i=1
D
i
l
i,t
LGD
t
= LGD
t
N
t
_
i=1
D
i
l
i,t
(8.2)
where D
i
= 1 if rmi is in default and D
i
= 0 otherwise, and where l
denotes the amount of rm is loans; N
i
is the total number of rms.
In other words, the total loan losses in year t simply equal the sumof
loans of bankrupt rms in that year, multiplied by the common LGD
in that year. We have data on annual total loan losses, the number
of defaulted companies in each year and the loan size distribution
across companies. In order to estimate the LGD
t
for each year, we
draw random samples of size k
t
out of the annual population of N
t
rms such that k
t
is the number of defaulted rms in that year. For
each round of sampling, the LGD
t
which satises Equation 8.2 is
computed. After a sufcient number of random samples, we obtain
a distribution of LGD
t
s. The mean of this distribution is then used
as our nal estimate of the LGD
t
. To our knowledge, this type of
method has not previously been used in empirical studies on LGDs
and it might be interesting to explore it further in future work.
5
Figure 8.4 depicts the estimated output gap annual LGD series
against the actual, and provides in the caption the results of regress-
ing the actual LGD on the output gap. This model is then used to
endogenise the LGD in our loan-loss simulations. We nd that dur-
ing the sample period the annual estimated LGD ranges between
12% (in 2006) and 73% (in 1991), the average being 47%, which is
well in line with, say, the 45% reference point used in Basel II.
195
MODEL RISK
Figure 8.4 The estimated average output gap annual LGD against the
actual LGD
0.2
0.1
0.0
0.4
0.1
0.3
0.2
0.5
0.7
0.6
0.8
1988 1990 1992 1994 1996 1998 2000 2002 2004 2006
Actual LGD
Output gap
The regression of the actual LGD on the output gap obtains the following param-
eter estimates: actual LGD
t
= 0. 43 2. 03 output gap
t
. The signicance level
of the coefcient on the output gap is 0.044% and the R
2
of the regression is 22%.
SIMULATION RESULTS
As Sorge and Virolainen (2006) describe, the macro-based empiri-
cal model for industry-specic default rates can be used to simulate
loan losses. Here we give a brief account of the procedure; the reader
is referred to Sorge and Virolainen (2006) for further details. We take
a representative bank loan portfolio of the Finnish corporate sec-
tor and group the included companies according to their industry.
In the absence of rm-specic balance-sheet data, each company
in each period of the simulation is then assigned the default rate
of its industry, obtained from the default-rate model. An indepen-
dent binary random draw is then carried out for each company in
order to determine whether it survives or defaults in a given simula-
tion period. If a company defaults, a share of its outstanding credit,
determinedby the LGD(either constant or endogenous) is taken as a
loss. The multi-periodsimulationprocedure keeps trackof defaulted
rms in the portfolio, so that each company can only default once
during the simulation horizon. Individual rms credit losses are
then summed to obtain the aggregate cumulative bank loan losses
of the corporate sector at the end of the simulation horizon.
As to the representative bank loan portfolio of the Finnish corpo-
rate sector, Sorge andVirolainen(2006) useddata of the 3,000 biggest
196
Finnishcompanies from2002.
6
Together, these companies accounted
for more than 90% of the total loans granted by monetary nancial
institutions (henceforth banks, for brevity) to the corporate sector
that year. Unfortunately, it was not possible, due toits non-public sta-
tus, for us to extend the portfolio data used in Sorge and Virolainen
(2006). However, we have had access to a more limited data source
of the 500 biggest companies in Finland which, on the other hand,
does not provide the division of corporate credit into bank loans and
other credit
7
. Simulation experiments with this alternative data set
indicated that changes in the loan size distribution across compa-
nies have only minor effects on the resulting loan-loss distributions
in our modelling framework. Therefore, our nal choice was to use
the portfolio composition in Sorge and Virolainen (2006) as the rep-
resentative corporate loan portfolio throughout the entire sample
period.
Loan-loss distributions with extensions to the basic model
Our rst set of simulation results compares the basic model of
Sorge and Virolainen (2006) with the various extensions we have
considered:
the endogenous LGD;
industry-specic output shocks;
two alternative shock specications (the standard AR(2) case
used by Sorge and Virolainen (2006) and a random walk).
These extensions are grouped in Table 8.3 in the following way. The
two alternative shock specications (cases (a) and(c)) andthe case of
industry-specic shocks (b) are each run with both a constant andan
endogenous LGD. Hence, a total of six cases are considered. The case
withthe aggregate output shock, modelledwithAR(2), andconstant
LGD, corresponds to the case considered by Sorge and Virolainen
(2006) and is illustrated in Figure 8.5. In each case we computed
three descriptive statistics of the simulated loan-loss distribution
three years ahead, starting at the beginning of 2008: the expectedloss
and the unexpected loss at both 99 and 99.9%condence (or value-
at-risk) level. The expected endogenous LGD is always adjusted
(approximately) to the same level as the constant LGD, so that the
most interestingcaseof howendogenous LGDaffects thetailof the
197
MODEL RISK
Table 8.3 Summary of simulations
Unexpected Unexpected
Expected loss loss
loss (VaR 99%) (VaR 99.9%)
(a) LGD = 0.43 1.72 2.56 3.69
LGD endogenous 1.74 2.69 3.97
LGD min/mean/max
= 0.26/0.43/0.59
(b) LGD = 0.43 1.73 2.56 3.73
LGD min/mean/max
= 0.25/0.43/0.60
(c) LGD = 0.43 1.59 2.51 3.78
LGD min/mean/max
=0.18/0.43/0.70
(a) Aggregate output AR(2); (b) industry-specic output AR(2); (c) aggre-
gate output RW. The (weighted) aggregate PD is 0.0033 in (a), 0.0034
in (b) and 0.0031 in (c). Simulation horizon is three years starting at the
beginning of 2008.
loan-loss distribution can be considered. The number of simulation
rounds in each case is 50,000.
8
When examining the results in Table 8.3 we rst note that, over-
all, differences in magnitude between the various cases are not very
big. However, as expected, the endogenous LGDclearlyhas aneffect
of widening the loan-loss distribution. This is manifested as higher
unexpected losses. In the base case model (a) with aggregate out-
put and AR(2) shocks, the endogenous LGD increases the 99% and
the 99.9% unexpected losses by 58%. Similarly, in cases (b) and (c)
the endogenous LGD increases the unexpected losses. One reason
why the effect of the endogenous LGD is not very large is that the
explanatory power of the LGD-output regression model is only
22% (see Figure 8.4). In actuality, further experiments with the LGD
model revealed that a much higher coefcient on output (and thus
a much higher explanatory power) is achieved if we use the rst
lead of the output as the explanatory variable. Such a lead structure
may be understood as resulting from, say, certain loan-loss account-
ing conventions, but it is problematic to implement in the loan-loss
198
Figure 8.5 Distribution of loan losses (xed LGD)
0 1 2 3 4 5 6 7 8
0.5
1.0
0
1.5
2.5
2.0
Loss in % of total credit exposure
F
r
e
q
u
e
n
c
y

(
%
)
simulation model. Therefore, we have retained the original LGD-
output specication in the simulations. However, should we want
to experiment withloan-loss simulations usingthe higher coefcient
on output in the LGDmodel, the resulting widening of the loan-loss
distribution would naturally be more pronounced.
When we compare cases (a)(c), either with the constant LGD or
with the endogenous LGD, we see that the unexpected losses stay
roughly at the same level for the different model versions. The high-
est 99.9% unexpected loss (the unexpected loss being 3.97% of the
loan stock) is obtained for the base case model (a) with endogenous
LGD. We may conclude from these results that it does not seem to
matter whether we use industry-specic or aggregate output shocks
or how the shock processes are specied. What does matter is that
we replace the constant LGD with the endogenous LGD. We also
investigated the effect on the loan-loss distribution of increasing
the standard deviation of a shock (in this case the output shock).
As we expected, unexpected losses increased correspondingly in an
approximately linear manner.
Prolonged recession and simultaneous shocks
After the analysis of unconditional loan-loss distributions, we wish
to study the impact on loan losses of a prolonged deep recession.
Figure 8.6 illustrates the results of this exercise. The main message is
199
MODEL RISK
Figure 8.6 Expected losses and the length of depression: feedback
from defaults to output
Growing shock
Constant shock
Difference between growing and constant shock
14
12
10
8
6
4
2
0
2
E
x
p
e
c
t
e
d

l
o
s
s
e
s

(
%
)
2 4 6 8 10 12 14 16 18 20
t
Feedback means here that GDP is decreased at a constant quarterly rate that
is thought to correspond to the loss of productive capacity (rms) along with
prolongation of depression.
that, if a constant negative shock to output persists over several peri-
ods (up to ve years), the cumulative impact on expected loan losses
is signicantly convex. Roughly speaking, if cumulative expected
losses are 1% of the loan stock at the end of the rst year, they are
more than10%at the endof the fthyear, while a linear extrapolation
wouldsuggest cumulative expectedlosses of 5%(see Figure 8.6). Fig-
ure 8.6 further compares the base case with a case in which we have
added a simplistic feedback mechanism from defaults to output;
GDP is decreased at a constant quarterly rate which is thought to
correspond to the loss of productive capacity of the corporate sector
as a result of cumulating defaults. Clearly, the convex impact on the
expected losses of the prolonged recession is reinforced. The convex
effect is probably at least partly the result of the fact that the indebt-
edness increases as output decreases, because output directly affects
the denominator in the indebtedness variable. The policy implica-
tion of these simulation results would appear to be that long-lasting
shocks causing an economic downturn should be dealt with at an
early stage before they develop into prolonged recession.
In Figure 8.7 we illustrate another important effect related to
simultaneity of shocks. In other words, does it make a difference
whether shocks that are known to increase loan losses take place
simultaneously or whether they happen one at a time? In order to
study this question in a meaningful manner we have rst chosen
200
Figure 8.7 Comparison of effects of macro shocks
25
20
15
10
5
0
10 50 100
I
n
c
r
e
a
s
e

i
n

e
x
p
e
c
t
e
d

l
o
a
n
l
o
s
s
e
s

(
%
)

d
u
e

t
o

t
h
e

s
h
o
c
k
s
Growth of PD in each individual shock (%)
Sum on individual shocks
Combination of shocks
Sum of individual shocks denotes the sum of differences between the simulated
values and the base in terms of the expected losses due to the three macro shocks
(of equal size in terms of the PD). Combination of shocks denotes the analogous
difference in expected loss due to a simultaneous occurrence of these three macro
shocks (all of equal size).
shocks to output, real interest rate and the corporate sector indebt-
edness in such a way that each individual shock alone would pro-
duce an equal size increase in the weighted average of the industry-
specic PDs. We then make the following comparison. We run our
basic model of loan losses separately with each individual shock
and take the sum over the three model runs of the increase in the
aggregate expected loan loss. This sum of expected losses is then
compared with a single run of the basic model in which all the
three previous shocks take place simultaneously. This comparison is
depicted in Figure 8.7 for three different shock sizes, corresponding
to 10%, 50% and 100% increases in the PD. Clearly the combined
effect is much larger. This result obviously reects the correlations
between individual shocks. This analysis may also provide one way
to better understand what happened in the Finnish crisis of the early
1990s and howthe situation is different fromtodays perspective. In
the early 1990s, clearly, a combination of shocks hit Finland: out-
put dropped as a result of a big export shock, effective indebtedness
increased dramatically as a result of the devaluation of the currency
and interest rates skyrocketed. In contrast, todays conditions seem
essentially less severe, as corporate indebtedness remains moderate
and interest rates are low, making the Finnish banking sector better
prepared to weather the negative export demand shock resulting
from the 20089 global crisis.
201
MODEL RISK
Effect on loan-loss risk of an industrial cluster
When assessing the risk of aggregate loan losses in an economy, we
note that, particularly in many small economies, single industrial
clusters or even individual companies can constitute a sizeable por-
tion of total output. Finland is a good example of such a situation,
as the ICT and forestry-related industrial clusters are central to the
economy. If such clusters, which operate globally, make decisions to
move their production to other countries, it is in principle possible
to have large output effects within a relatively short time period.
Although such moves in themselves would not necessarily induce
any credit risks, the second-roundeffect via overall decline inoutput
could entail increasing loan losses, as our empirical model suggests.
For the purpose of loan-loss-scenario analysis it may be wise to
consider suchshifts inproductionandoutput as separate riskevents,
because the probability of such events may not be properly captured
on the basis of historical output uctuations. In the following we
consider the effect on loan losses that a single industrial cluster (the
ICT cluster in Finland) could have. To this end, the version of our
empirical model with industry-specic output, discussedabove (see
page 197), is quite useful. We rst assess how the ICT cluster affects
output in the respective industry (manufacturing). The shock then
spreads to other industries, having an effect on their respective out-
puts, throughthe models shockcorrelationstructure. For simplicity,
we consider anextreme scenario of the Finnisheconomy withanexit
of the ICT cluster. Because the ICT cluster makes up around 22% of
the manufacturing industry in Finland, we consider a 22% cumula-
tive negative output shocktothe manufacturingindustrythat would
take place over the three-year horizon, our standardsimulationhori-
zoninthis study.
9
Sucha shock turns out to have the largest spillover
effect on output in construction and trade industries, as well as in
agriculture andother service industries, ineachof whichoutput falls
by about 7%. The effect onaggregate loan-loss distributionis that the
expectedloss is 3.22%while the 99%and99.9%unexpectedlosses are
at 4.01%and5.85%, respectively. As we have usedthe model version
with endogenous LGD, these numbers should be compared, respec-
tively, to 1.75%, 2.67% and 3.91% from Table 8.3(b) (second line).
Thus, the ICT cluster shock almost doubles the expected aggregate
loan losses relative to the unconditional expected losses. The size of
the effect on loan losses suggests that the type of approach to stress
202
Figure 8.8 Fit of the constant LGD and the endogenous LGD loan-loss
models
500
0
500
1,000
1,500
2,000
2,500
3,500
3,000
1
9
9
1
1
9
9
2
1
9
9
3
1
9
9
4
1
9
9
5
1
9
9
6
1
9
9
7
1
9
9
9
1
9
9
8
2
0
0
0
2
0
0
1
2
0
0
2
2
0
0
3
2
0
0
4
2
0
0
5
2
0
0
6
2
0
0
7
Actual loan losses
Expected loan losses, endog. LGD
Expected loan losses, LGD = 0.43
testing taken in this subsection may be important. Lastly, as already
discussed (see page 192), when modelling an output shock that orig-
inates from a certain industry and then spills over to the rest of the
economy, this has to be taken into account by setting the original
shock as the rst one in the matrix of the Cholesky decomposition.
Our experiments showed that ignoring the proper order of shocks
may greatly bias the results downwards.
EMPIRICAL FIT OFTHE LOAN-LOSS MODEL
Because the Finnish loan-loss experience of the rst half of the 1990s
was so extraordinary, it is very tempting to try to get at least a rough
idea of how well the current macro-based model could capture that
episode. As discussed earlier in this chapter, we effectively assume
that the individual loan size distribution of the aggregate bank loan
portfolio has stayed invariant.
The results are depicted in Figure 8.8. Although the model can
follow the overall prole of the aggregate loan losses, it exaggerates
loan losses in normal times and falls greatly short of them in the
crisis years in the 1990s. This was partly to be expected: although
our default rates model ts quite well to actual default rates (cf. the
high R
2
s mostly in the range 8090% in Table 8.1), the uctuations
in the aggregate loan losses are larger than in the aggregate default
rate, as shown in Figure 8.2. The endogenous LGD explains part of
the gap, as expected, but not much in relation to the size of the gap.
One possible explanation for the remaining gap is that the effects
203
MODEL RISK
of the large devaluation of the Finnish currency during the crisis
are not fully controlled for in the calculation of loan losses. Namely,
a number of non-exporting companies had taken foreign currency
denominated loans from Finnish banks (cf. the recent experience in
Iceland). As a result of the devaluation, the nominal value of these
loans rose in terms of the domestic currency. Unfortunately, we do
not have sufciently disaggregated data to control for these effects.
A second potential explanation is that, during the Finnish cri-
sis, many big export-oriented companies went bankrupt. Obviously,
their relative weight is not sufciently reected in the industry-
specic default rates we have used. For instance, given that the
almost overnight collapse of trade with the Soviet Union was a cen-
tral reason for the Finnish crisis (see, for example, Gorodnichenko et
al 2009), we should have data on rm-level exposures to the Soviet
trade to have a more disaggregated model of defaults and hence
to better capture the actual loan-loss behaviour. Clearly, back then
the collapse of the Soviet Union was a big unexpected event and,
with hindsight, a big single risk factor to the Finnish economy. In
the same vein as the arguments in the section on industrial clusters
(see page 202), macro-based credit-risk models might benet from
trying to incorporate single risk factors related to single institutions,
markets or products that forma sizeable part of the economy andare
vulnerable to discrete events that might dramatically change their
role in the economy.
LOAN-LOSS DISTRIBUTION: 2007 VERSUS PRE-1990S CRISIS
Finally, we made a comparison between the loan-loss risk outlook
prior to the Finnish crisis in the early 1990s and 2007. Again, we sim-
ply used the sample portfolio from 2002 as a proxy for the portfolio
prevailing before the 1990s crisis.
In particular, we considered the state of the banking sector at the
endof 1989 andtook the starting values for the macro variables from
the rst andsecondquarter of 1990. Withthe basic constant LGDver-
sion of the model, the expected loss and the 99% and 99.9% unex-
pected loss, respectively, were 3.65, 3.53 and 4.95. These results can
then be compared with the situations at the end of 2007. In Table 8.3,
the corresponding numbers are 1.72, 2.53 and 3.58. In terms of the
expected loss, we can see that the aggregate loan-loss risk just before
the 1990s crisis hit was more than twice the value at the end of 2007.
204
The difference in risks is effectively a result of the different macro-
economic position then and now. In particular, the indebtedness of
the corporate sector in Finland was much higher on the eve of the
1990s crisis than it is now. Total corporate sector indebtedness in the
second quarter of 1990 was almost double and the indebtedness of
the manufacturing industry alone was almost three times the respec-
tive indebtedness in the second quarter of 2008. In general, these
results emphasise the role played by the prevailing macroeconomic
conditions in assessing risks of future loan losses.
It is tempting to speculate what might have been done differently,
had the awareness of the size of the aggregate credit risk been better
at the end of the 1980s. Clearly, knowledge and use of the current
type of credit risk portfolio models started their international pro-
liferation only in the latter half of the 1990s. Of course, the crisis
of 20089 also revealed severe inadequacies in the 2009 credit-risk
models. Nevertheless, the comparison for Finland that we have car-
ried out here by using the current modelling framework is justied
on the basis that, for Finland, the 20089 crisis came almost entirely
as anexternal (export-driven) shock. The Finnishbankingsector was
not greatly affected by the contagion in the global banking sector.
CONCLUDING REMARKS
This study illustrates howmacroeconomic shocks affect banks loan
losses from the corporate sector by revisiting and extending the
model of Sorge and Virolainen (2006). In the base model, the cen-
tral macroeconomic factors that drive industry-specic default rates
and hence loan losses are the output gap, the real interest rate and
the corporate sector indebtedness. The empirical model for default
rates is then used in simulating the aggregate loan-loss distribution.
We considered the following extensions to the base model and
studied their effect on the loan-loss distribution. Firstly, instead of
aggregate output we consideredindustry-specic outputs; secondly,
we relaxed the constant LGD assumption used in the simulations
and made LGD depend on the output gap; nally, we considered
alternative ways of specifying shocks to the explanatory macro vari-
ables. It turned out that, in terms of the loan-loss distribution, it was
mainly the endogenised LGD that had the most obvious material
impact. We also showed that disaggregation signicantly improves
the model properties, including its stability properties, but it does
205
MODEL RISK
not have a signicant quantitative impact on the loan-loss simula-
tion results. Moreover, the model with industry-specic outputs is
useful when we consider the potential second-round effects on the
aggregate loan-loss risk of a single industrial cluster, particularly the
ICT cluster, which is central to the Finnish economy.
We also considered the empirical t of the model-based expected
loan losses with actual loan losses and found that, although the
endogenous LGDimproves the t, the model nevertheless falls short
of explaining the large loan losses experienced in the crisis in the
early 1990s in Finland. Identifying the missing risk factors that could
explain the gap remains an issue for future research.
We also emphasise that the severity of a crisis, in terms of mount-
ing loanlosses, may very muchdependonthe exact nature of the cri-
sis. That is, the combination of simultaneous macroeconomic shocks
as well as the duration of these shocks may be important. We stud-
ied these issues with the help of scenario-based analyses and found
that both prolonged deep recessions and a combination of simul-
taneous shocks seem to have a convex effect on loan losses. This
suggests that policy actions should be designed in a way that pre-
vents the acceleration of a looming crisis. Finally, a comparison of
the loan-loss distribution on the eve of the 1990s crisis with the most
recent distributiondemonstratedthe greatlyelevatedrisklevel prior
to the 1990s crisis. More generally, the comparison emphasises the
effect of prevailing macroeconomic conditions on potential future
loan losses.
The authors thank Kimmo Virolainen, who kindly provided the
original computer modelling code. They also thank Jyrki Ali-
Yrkk, Heikki Hella, JoukoVilmunen, KimmoVirolainenandsem-
inar participants at the Bank of Finland for valuable comments.
Finally, they are indebted to Anni-Mari Karvinen for outstanding
researchassistance, whichhas beenessential totheimplementation
of this project. All errors remain the responsibility of the authors.
The views expressed in this chapter are those of the authors and
do not necessarily reect the views of their employers.
1 For more details of the Finnish banking sector loan losses and comparisons with other Nordic
countries, see Pesola (2001).
2 That loss given default tends to increase in recessions is empirically rather well established;
see, for example, Schuermann (2004) and the references therein.
3 For the simulationanalysis, we changedthe equationfor agriculture by setting the coefcients
of incorrect signs equal to zero and re-estimating the equations with these restrictions.
206
4 This result seems to be in striking contrast with some preliminary estimation results fromthe
model for aggregate loan losses. Thus, if loan losses are explained by the aggregate default
rate (d), the explanatory power is reasonably high (0.93) but the RESET test clearly suggests
that the functional form is misspecied (F(2, 76) = 43. 54). The relationship seems to be very
strongly non-linear, which also shows up in the fact that if the second and third powers of d
are introduced as additional explanatory variables the R
2
goes up to 0.97.
5 Further details of the LGDestimation procedure are available fromthe authors upon request.
6 Data also includes some information from 2003.
7 The data source is Talouselm 500.
8 This may still leave roomfor some inaccuracy stemming fromthe nite number of simulation
rounds, so some caution is in order when interpreting this data, particularly the unexpected
loss gures at the far end of the loan-loss distribution.
9 For an alternative way of quantifying the output contribution of the Finnish ICT cluster, see
Ali-Yrkk et al (2000).
REFERENCES
Ali-Yrkk, J., L. Paija, C. Reilly and P. Yl-Anttila, 2000, Nokia - A Big Company in a
Small Country, Research Institute of the Finnish Economy Series, Volume 162 (Helsinki:
Taloustieto Oy).
Conesa, J., T. Kehoe and K. Ruhl, 2007, Modeling Great Depressions: The Depression in
Finland in the 1990s, in T. Kehoe and E. Prescott (eds) Great Depressions of the Twentieth
Century (Federal Reserve Bank of Minneapolis).
Gorodnichenko, Y., E. G. Mendoza and L. L. Tesar, 2009, The Finnish Great Depression:
From Russia with Love, NBER Working Paper 14874.
Honkapohja, S., E. A. Koskela, W. Leibfritz and R. Uusitalo, 2009, Economic Prosperity
Recaptured: The Finnish Path from Crisis to Rapid Growth (MIT Press).
Laeven, L., and F. Valencia, 2008, Systemic Banking Crises: A New Database, IMF
Working Paper 08/224.
Pesola, J., 2001, The Role of Macroeconomic Shocks in Banking Crises, Bank of Finland
Discussion Paper 6/2001.
Reinhart, C., and K. Rogoff, 2008, Banking Crises: An Equal Opportunity Menace,
NBER Working Paper 14587.
Schuermann, T., 2004, What Do We Know about Loss Given Default?, in D. Shimko
(ed.) Credit Risk Models and Management, Second Edition (London: Risk Books).
Sorge, M., and K. Virolainen, 2006, A Comparative Analysis of Macro Stress-Testing
Methodologies withApplication to Finland, Journal of Financial Stability 2(2), pp. 113216.
207
9
Comparison of Credit-Risk Models
for Portfolios of Retail Loans
Based on Behavioural Scores
Lyn C. Thomas; Madhur Malik
University of Southampton; Lloyds Banking Group
Both the fact that the Basel Accord formula is based on a corpo-
rate credit-risk model and the misrating of mortgage-backed secu-
rities which led to the 20089 credit crunch have highlighted that
the development of credit-risk models for retail loan portfolios is far
less advanced than the equivalent modelling for portfolios of corpo-
rate loans. Yet, for more than three decades, behavioural scoring has
proveda very successful way of estimating the credit risk of individ-
ual consumer loans. Almost all lenders produce a behavioural score
for each of their loans every month. This chapter reviews the differ-
ent models that have beendevelopedinorder touse these individual
behavioural scores to assess the credit risk at a portfolio level. The
models have parallels with the types of corporate credit-risk mod-
els, but differ because of the need to deal with the features specic
to retail loans such as the months-on-books effect. Thus, there are
structural type models, plus those based on hazard rates and those
that use Markov-chain stochastic approaches.
INTRODUCTION
Modelling the credit risk of portfolios of consumer loans has
attracted far less attention than the modelling of their corporate
equivalent. This was rst apparent whenthe Basel II Accordformula
for the minimum capital requirement (Basel Committee on Banking
Supervision 2005), which was based on a version of the Merton
Vasicek model of corporate credit risk, was applied to all types of
loans, including consumer loans. The parameters for the consumer
209
MODEL RISK
loan regulations were chosen empirically to produce appropriate
capital levels.
Another example of the lack of research into modelling the credit
risk of portfolios of consumer loans is the failure of the ratings agen-
cies to accurately rate the credit risk of securities based on US mort-
gages. This was one of the causes of the credit crunch of 20089. It is
clear that the models theyusedwere awed, as the number andscale
of the subsequent downgradings indicate; many of these downgrad-
ings occurredwithin nine months of the original rating. This has had
such a severe impact on the worlds banking system that there have
been several enquiries (Securities and Exchange Commission 2008;
Financial Service Authority 2009) and a number of research papers
(Ashcraft and Schuermann 2008; Crouhy et al 2008; Sy 2008) inves-
tigating what went wrong. Some of the problems identied were to
do with the relationship between the ratings agencies and the origi-
nators of the securitisation, with the data (or lack of it) supplied, but
one of the problems was trying to extend a methodology usually
used for the credit risk of individual companies to portfolios of con-
sumer loans. For example, the only data used on the credit risk of
the individual consumer loans was the initial application score, and
some of the special features of consumer loans, such as the length of
time the loan has been in operation, were ignored.
In this chapter we consider three approaches to modelling the
credit risk of portfolios of consumer loans, all of which are based
on the behavioural scores of the individual borrowers who make
up the portfolio. This information is now calculated on a monthly
basis by almost all lenders and by all credit bureaus and gives an
assessment of the current risk of each borrower defaulting. Using
this information, which has proved so successful for 30 years in
making decisions on individual borrowers, would allow lenders to
develop models that can react quickly to the changes in the credit
environment and the economic and market conditions. The three
models have analogies with the three main approaches to corporate
credit-riskmodelling: astructural approach, areduced-formdefault-
mode approach and a ratings-based reduced-formapproach. Exam-
ples of these approaches can be found elsewhere in this book and in
Saunders and Allen (2002), for example.
Most of the models developed for consumer portfolio credit risk
(or retail portfoliocredit risk as it is oftencalledinthe Basel Accord
210
COMPARISON OF CREDIT-RISK MODELS FOR RETAIL LOAN PORTFOLIOS
terminology) use the data on defaults. Thus, Bucay andRosen (2001)
build a sector-based model of a retail portfolio, where the correla-
tionbetweensectors is obtainedbecause theyall dependoncommon
economic variables. The relationshipbetweenthe default rate for the
sector andthe economic variables is obtainedusing linear regression
to estimate the impact of the economy on the logit or probit trans-
formation of the aggregated default rate. Rosch and Scheule (2004)
split a retail portfolio into residential mortgage, revolving loans and
other loansectors anduse essentially the Basel Accordmodel ineach
sector. The relationship between default and economic variables in
each sector is estimated at the individual loan level using a probit
model, where the economic variables are suitably lagged. Perli and
Nayda (2004) concentrated on revolving consumer credit and apply
the corporate structural model, but with the assets depending on
two systemic factors rather than the one factor of the standard Basel
model.
Musto and Souleles (2006) use behavioural scores in a con-
sumer portfolio credit-risk model, but they take the difference in
behavioural score for an individual between one month andthe next
as a surrogate for the return on assets of that borrower. These sur-
rogate return on assets are used to mimic equity pricing models in
order to obtain a value for the consumer loan portfolio. Musto and
Souleles make the point, however, that behavioural scores are eas-
ily available for each borrower and are updated frequently (at least
monthly) and so have analogous properties to the prices in equity
models.
Onthe other hand, this chapter looks at models where behavioural
scores are used for what they really are: measures of the current
default risk of the individual borrowers who make up the portfolio
of loans. In the next section we highlight how such behavioural
scores are obtained and what they mean. We point out that there
is an underlying assumption that the creditworthiness of customers
is time independent but that a simple adjustment can be found that
relaxes this assumption somewhat. We then (see page 214 onwards)
describe a structural model for the credit risk of consumer loans sug-
gested byAndrade and Thomas (2007), where the behavioural score
is a surrogate for the creditworthiness of the borrower. A default
occurs if the value of this reputation for creditworthiness, in terms
of access to further credit, drops belowthe cost of servicing the debt.
211
MODEL RISK
In the following section we look at the default-mode hazard model
developed by Malik and Thomas (2007), where the risk factors were
the behavioural score, the age of the loan and economic variables.
Such an approach is now being used to develop behavioural scores
for the individual borrower. It has the advantage that it can give esti-
mates of the default risk over any future time horizon (Banasik et al
1999; Stepanova and Thomas 2001), and an extra advantage is that
it can be used to build credit-risk models at the portfolio level by
incorporating economic variables. Next, we describe a model more
akin to the corporate reduced-form mark-to-market model. It uses
a Markov-chain approach, where the states are behavioural score
intervals, accounts defaulted or accounts closed, to model the future
dynamics of retail borrowers. As in the hazard-rate approach, we
nd that the current age of the loan has a much more important role
in the credit risk of consumer loans than it does in that of corporate
loans.
BEHAVIOURAL SCORING
Credit scoring has been used for more than 50 years to support
consumer lending decisions. Initially, application scorecards were
developed to assess the credit risk of potential borrowers if they
were to be given a loan. By the mid 1970s, behavioural scoring that
assessed the credit risk of existing borrowers was being widely used
to assist in credit-limit and cross-selling decisions. Its usage was
further enhanced by the introduction of the Basel Accord in 2007,
since it is the basis for the internal ratings systems for assessing the
credit risk of consumer loans which were permitted to be used for
regulatory capital allocation under that accord.
Behavioural scores estimate the risk that the borrower will default
in the next 12 months. They are obtained by taking a sample of
previous borrowers andrelatingtheir characteristics, includingtheir
repayment, arrears and usage during a performance period, with
their default status 12 months after the end of that performance
period. The other characteristics that may be part of the scorecard
include data fromthe credit bureaus, such as the borrowers overall
debt situation, and some socioeconomic data from the application
form, but rarely anything on the current economic situation. Some
of these characteristics indicate whether the borrower can afford to
repay the loan, but the most important characteristics are usually
212
those from the credit bureau and information on the arrears status
of the borrower. A borrower is usually assumed to have defaulted
if their payments on the loan are more than 90 days overdue. If we
dene those who have defaulted as bad (B) and those who have
not defaultedas good (G), then the behavioural score is essentially
a sufcient statistic of the probability of the borrower being good.
Thus, if x are the characteristics of the borrower, a score s(x) has the
property that P(G x) = P(G s(x)). Scores which are constructed
usinglogistic regression, byfar the most commonwayof developing
behavioural score, are log-odds scores, so
s(x) = log
_
P(G x)
P(B x)
_
p(G x) =
1
1 +e
s(x)
Suchscores decompose into two parts, one depending onthe default
rate of the underlying population and the other on the character-
istics of the individual borrower. If p
G
and p
B
are the proportions
of goods and bads in the underlying population and p(x) is
the proportion of the population with characteristics x, then Bayess
Theorem implies that p(G x) = (p(x G)p
G
)/p(x) and hence that
s(x) = log
_
p(G x)
p(B x)
_
= log
_
p
G
p(x G)
p
B
p(x B)
_
= log
_
p
G
p
B
_
+log
_
p(x G)
p(x B)
_
= s
pop
+woe(x)
where
s
pop
= log
_
p
G
p
B
_
and woe(x) = log
_
p(x G)
p(x B)
_
is the weight of evidence of characteristics x. If nothing is known
about a borrower, then they would be given the score s
pop
, reecting
the overall proportions in the population. When we know the char-
acteristics x of the borrower then the woe(x) term is added in order
to obtain the score for that individual borrower.
The hiddenassumptionbehindbehavioural scores is that the rela-
tionship between the score and the probability of being good, or
of defaulting, is time independent, at least over time periods of a
few years. Hence, the same scorecard is used for a number of years
and then, when it has aged, a completely new scorecard is devel-
oped using a more recent sample of borrowers. This assumption of
213
MODEL RISK
time independence is not borne out by experience, especially in tur-
bulent economic times. The true score at time t of a borrower with
characteristics x, if the scorecard was constructed at time t
0
, would
satisfy
s(t, x) = log
_
P(G x, t)
P(B x, t)
_
= s
pop
(t) +woe(x, t)
It may be defendable to assume that the weight of the evidence
term is time independent and so let woe(x, t) = woe(x), even if
it is not really true, but it cannot be reasonable to assume that the
population odds term is independent of t. However, the score s(x)
being used was constructed at t
0
, so
s(x) = s
pop
(t
0
) +woe(x)
Thus, we should adjust the behavioural score so that
s(t, x) = s(x) +(s
pop
(t) s
pop
(t
0
))
toobtaina score that reects the dynamics of the situation. Todothis,
we need to use the current default rate (or, perhaps more correctly,
the projected future default rate in the next year) of the population
of current borrowers. This would give the s
pop
(t) term, while the
s
pop
(t
0
) term can be obtained from the default rate at the time the
sample on which the scorecard was built was active. The impact of
this adjustment is todecrease the scores intimes of difcult economic
conditions and raise them when default rates are low. The equiva-
lent of this adjustment is frequently made for application scores by
increasing the score at which applicants are accepted in bad times
and lowering them in good times. However, no such adjustments
seem to be made for behavioural scores to allow for the changes in
economic conditions.
REPUTATIONAL STRUCTURAL MODEL
The basic tenet of the structural model approach to corporate credit
risk is that a rm defaults if its assets exceed its debts and that the
rms share price is a useful surrogate for describing its assets. Thus,
the shareholders can be consideredto have a call option on the assets
of the rm, which if the assets drop below the debt level they will
not exercise, and so they will let the rmdefault. Such a model does
not translate directly into the consumer context as most consumers
214
do not knowthe value of their assets andwouldnot be able to realise
them anyway; there is no share price of a consumer and consumers
default more because of cashow problems than total asset difcul-
ties. However, Andrade and Thomas (2007) suggested that a similar
model could be built for individual consumer loans and portfolios
of consumer loans by assuming that a consumer has a call option on
their reputation. In such a model the behavioural score can act as a
surrogate for the creditworthiness of the borrower.
Assume that the creditworthiness Q
i
of borrower i is an unobserv-
able quantity. A lender, though, has information on this creditwor-
thiness from credit bureaus and by checking the performance of the
borrower in the recent past, which allows the lender to construct a
behavioural score s(i), which is a useful proxy for this creditworthi-
ness. The chance, P
i
, that borrower i can get access to further credit
must be an increasing function of Q
i
, P
i
= f (Q
i
). This access to credit
must be of value, V
i
, to a consumer and the greater this value is, the
easier (that is, the more likely) it is for borrower i to get credit. So
V
i
= g(P
i
) = gf (Q
i
) = v(Q
i
)
where g and f and hence v are strictly increasing functions. If a bor-
rower defaults, this information is passed to the credit bureaus and
hence to all the lenders. Thus, the lender will lose their reputation
for creditworthiness and will have no access to credit in the immedi-
ate future. The value of their creditworthiness drops to zero. Thus,
a borrower will only default if the cost of paying back the debt D
i
,
exceeds the value of their reputation V
i
, ie, D
i
> V
i
; otherwise, the
borrower will continue to repay the debt. The borrower has a call
optionontheir reputationwhichtheywill exercise if their reputation
is above the debt level D
i
.
Assuming that the behavioural score s(t, i) of borrower i at time
t is a proxy for the borrowers creditworthiness Q
i
, we have V
i
=
v(s(t, i)) and so default occurs if
V
i
= v(s(t, i)) < D
i
s(t, i) < v
1
(D
i
) = K
i
So, to model when borrower i is likely to default, we need to model
the dynamics of the behavioural score. Andrade and Thomas (2007)
suggest that it should be represented by a continuous-time diffusion
with jumps, similar to the model suggested by Zhou (1997) so that
ds(t, i) = a
i
+b
i
dW+c
t
dY
t
215
MODEL RISK
where a
i
is the drift of the process, b
i
dW is a Brownian motion and
dY
t
is a Poisson jump process. Although the process is written in
continuous time, when it comes to estimating the parameters, we
will need to use a discrete-time equivalent with time intervals of one
month. The idea is that a
i
corresponds to a natural drift in creditwor-
thiness caused in part by the account maturing and so improving.
The Brownian motion describes the natural variation in behavioural
score, while the Poisson jump term is included to model jumps in
behavioural scores due to major changes in the economy. Perhaps
a more interesting model would be to make the jumps different for
different individuals and so it can be related to life-changing events
like job loss or marriage. This would give a model of the form
ds(t, i) = a
i
+b
i
dW+c
i,t
dY
i,t
We can estimate the parameters a
i
, b
i
and c
i,t
for each individual
by looking at the time series of that individuals behavioural scores
to date and using Bayesian Markov-chain Monte Carlo techniques
or maximum likelihood estimators.
Twoissues remain: howtochoose the default values K
i
andhowto
incorporate the population odds adjustment into the model to allow
for future changes in the economic conditions. One simple way to
allow for forecasts for the population odds adjustment is to assume
that the economy can be in a number of different states which are
classied according to the default rate for consumer loans. We can
then calculate for each state of the economy what the suitable s
pop
value should be and use historical data to build a Markov chain of
how the economy moves between these states.
For the calculation of the default levels K
i
Andrade and Thomas
(2007) suggested taking the same value K for all borrowers. The
way they choose K is then to apply Monte Carlo simulations of the
behavioural score paths. For each borrower the historical scores are
available and, having calculated the parameters using the historical
data, they apply simulation to obtain the score paths for the next
fewperiods. This is done a number of times for each individual and
by considering a possible default value K, they calculate the num-
ber of paths that go below that value (see Figure 9.1). This gives the
estimated default probability for that borrower. We can then choose
the value of K to ensure good calibration or good discrimination. In
the former case, we set K so that the simulated default rate in the
216
Figure 9.1 Monte Carlo simulation run to calculate appropriate K value
s
pop
Time
36 28 32 24 20 16 12 8 4 0 4 8 12
Historical
Simulated
Table 9.1 KolmogorovSmirnov (KS) results for alternative models
Model KS Increase in KS
Behavioural score (at last observation time) 41.0 0.0
Diffusion with drift and jump process 44.4 3.4
Diffusion with drift but no jump process 45.9 4.9
Diffusion with no drift but jump process 44.8 3.8
Diffusion without drift and no jump process 46.6 5.6
portfolio is equal to the actual default rate, allowing for the changes
in the underlying population that affect the population odds cor-
rection. In the latter case, K is chosen to maximise a measure of
discrimination such as the KolmogorovSmirnov (KS) statistic or
the Gini coefcient. Thus, they obtain a model that both is good at
correctly discriminating between the default risks of the borrowers
who make up the portfolio and hopefully gives a good estimate of
the total number of defaults in such a future period.
Andrade and Thomas (2007) produced a case example based on
Brazilian data. They split the economy into four states, where the
s
pop
values were 0.307, 0.132, 0.026 and 0.368, respectively. The
corresponding Markov-chain transition matrix is
_
_
_
_
_
_
0. 897 0. 103 0 0
0. 103 0. 685 0. 177 0. 035
0 0. 167 0. 733 0. 10
0 0. 067 0. 1 0. 833
_
_
_
_
_
_
The surprise was when they looked at whether they needed all
the terms in the dynamics of behavioural score. For each case they
217
MODEL RISK
calculated what was the Kolmogorov statistic corresponding to the
optimal K. As Table 9.1 shows, they found that the simplest model,
where no drift and no jump process (a
i
= c
i
= 0) was allowed in the
dynamics of the behavioural score, gave better results than the more
complicated models.
Such a model also has the advantage that the continuous-time
version of the model has an analytical solution for the probability
that default will occur within a time horizon t, P(t). The problem
reduces to the rst hitting time of a Brownian motion against a xed
barrier, namely
P(t) = 2N
_
K s(0, i)
b
t
_
Of course, making the value K independent of the borrower and of
time is a signicant simplication, and more realistic models might
make it a function of the vintage of the borrower (ie, when the loan
was taken out) and the current maturity of the loan (how long the
loan has been in operation at time t).
PROPORTIONAL HAZARD MODELS
One feature of the credit risk of consumer loans that does not appear
in corporate loans is the dependence of the risk on the maturity of
the loan, ie, the length of time since it was taken out. One approach
which deals with this is survival analysis. This also has the advan-
tages of accepting censored data as far as the default is concerned
(for example, loans that are paid off early or which are still repaying)
and also allowing estimates of the default risk over any future time
horizon. Survival analysis models the default risk by estimating the
hazardrate, where if T is the number of months until default, the haz-
ard rate h(t) = P(T = t T t). In Coxs proportional hazard model
(Cox 1972), we can connect the default time to the characteristics x
of an individual borrower by dening
h(t, x) = h
0
(t)e
s(x)
where s(x) = c
1
x
1
+ + c
n
x
n
is essentially a form of behavioural
score. Credit scores developed in this way were rst suggested by
Narain (1992) and the behavioural score versions were developed
by Stepanova and Thomas (2001). Such survival analysis or inten-
sity models have also been suggested in the corporate credit risk
context (Dufe et al 2007) but there the time used is calendar time
218
rather than months-on-book (MoB) time. Both approaches, how-
ever, allow the inclusion in the model of changes in the economic
environment.
Malik and Thomas (2007) suggest using this approach when esti-
mating the credit risk for portfolios of consumer loans, so that the
models may include the individuals behavioural score as well as
economic effects. Their model is as follows. The hazard probability
of default at time t for person i whose current behavioural score is
s(t, i) and who took out the loan at time t
0
, which is in the period
Vintage
i
, given that the current economic conditions are given by
EcoVar
i
(t
0
+t) is
h
i
(t) = h
0
(t) exp(as(t, i) +b EcoVar
i
(t
0
+t) +c Vintage
i
)
In the above equation h
0
(t) is the baseline hazard, which represents
the risk due to the age of the loan. We may think of the idiosyn-
cratic risk, systemic risk and the risk that the lenders marketing and
acceptance policy strategy has on the quality of the loans as being
represented by s(t, i), EcoVar(t) and Vintage, respectively. Such a
model has strongparallels withBreedens (2007) dual time dynamics
model, which directly models default risk (and other related quanti-
ties) at the portfolio level by describing MoB, vintage and calendar-
time effects. The vintage terms are a sum of binary variables includ-
ing the periodinwhichthe loanwas takenout, while the h
0
(t) values
give the MoB effect. The correlation between the default rates of dif-
ferent loans is given by the same economic variable values being
applied to all loans in the same period.
We can extend this model to estimate the hazard rate for k months
ahead, h
i
(t + k), for borrower i, who is currently t months into
repaying a loan taken out at time t
0
by dening
h
i
(t +k) = h
0
(t +k) exp(a
k
s(t, i) +b
k
EcoVar
i
(t
0
+t +k) +c
k
Vintage
i
)
The coefcients of these hazard rates can be estimated using Coxs
partial likelihoods (Cox 1972), andthe baseline hazardrate h
0
(t) then
calculated using NelsonAalen estimators.
How changes in economic variables affect the default risk of con-
sumer loans has attractedless attentionthanthecorrespondingques-
tion for corporate loans (Figlewski et al 2006). Malik and Thomas
(2007) describeda case studyusingUKcredit carddata for the period
20015 and used GDP and interest rates to describe the economic
219
MODEL RISK
Table 9.2 Coefcients in the case study of the proportional hazard
model
1-month 2-month 3-month
, .. , .. , ..
A B A B A B
Behavioural score band
1 2.24671 2.3323 2.6242 2.70791 3.09697 3.18273
2 8.54738 8.66029 6.98778 7.09739 5.62776 5.73832
3 9.60406 9.71268 8.08176 8.19148 6.08841 6.20220
4 10.0367110.14294 9.68832 9.79372 6.96915 7.07702
5 11.5715 11.66931 11.9021 11.99937 7.60843 7.71063
SV 6.01966 6.11543 6.47088 6.56873 5.80204 5.90250
Macroeconomic factors
IR 0.26229 0.25069 0.24416
GDP 0.42539 0.38256 0.31511
CPI
Vintages
Q1 1.21870 1.15786 0.99363
Q2 1.13194 1.07397 0.96448
Q3 1.01896 0.92263 0.77844
Q4 0.75628 0.68562 0.59465
Q5 0.81972 0.77152 0.71633
Q6 0.73978 0.69347 0.62936
Q7 0.51075 0.44722 0.39576
Q8 0.38299 0.31696 0.26262
Q9 0.26098 0.22958 0.18531
Q10 0.13124
Q11
Q12
Q13
Q14
Q15 0.77901 1.00105
LR test:
B v. A 218.5007 213.7275 175.94841
Monthly predictive model parameter values are given for models A and
B. IR, interest rate; LR, likelihood ratio. P-value for all B is <0.0001. SV,
special value.
environment. They also examined unemployment rate, consumer
price changes and stock market returns but in that benign envi-
ronment these did not change much and were not very predictive.
220
Table 9.2 (Cont.) Coefcients in the case study of the proportional
hazard model
4-month 5-month 6-month
, .. , .. , ..
A B A B A B
Behavioural score band
1 2.75823 2.84724 2.48733 2.56478 2.61022 2.67749
2 4.9138 5.02491 4.08235 4.18073 3.90613 3.99306
3 5.57769 5.69440 4.84958 4.95517 4.50302 4.60184
4 5.72644 5.83734 5.03777 5.13704 4.83499 4.92699
5 6.20157 6.30834 5.42494 5.52223 5.22294 5.31600
SV 5.1372 5.24298 4.58765 4.68248 4.48938 4.57994
Macroeconomic factors
IR 0.22181 0.22013 0.23729
GDP 0.24414 0.27600 0.34237
CPI
Vintages
Q1 0.79594 0.74695 0.79301
Q2 0.80352 0.74484 0.79143
Q3 0.65098 0.64334 0.70042
Q4 0.50605 0.46615 0.49181
Q5 0.65304 0.60409 0.66492
Q6 0.62144 0.60719 0.64168
Q7 0.35188 0.34361 0.36271
Q8 0.24582 0.21872 0.23177
Q9 0.21149 0.19159 0.19938
Q10
Q11
Q12
Q13 0.19690
Q14
Q15
LR test:
B v. A 139.9434 122.2438 109.3105
Monthly predictive model parameter values are given for models A and
B. IR, interest rate; LR, likelihood ratio. P-value for all B is <0.0001. SV,
special value.
The coefcients a
k
, b
k
and c
k
they obtained are given in Table 9.2,
where the behavioural score was split into ve bands from 1 (the
lowest behavioural score) to 5 (the highest behavioural score) with a
221
MODEL RISK
Figure 9.2 ROC curve for model A and model B of proportional
hazards example
0.0
0.2
0.2
0.0 0.4
0.4
0.6
0.6
0.8
0.8
1.0
1.0
P(s | Default)
P
(
s

|

N
o
n
-
d
e
f
a
u
l
t
)
Model A
Model B
special value, meaning those with no performance history to score
in the past six months. The vintages were quarter years fromthe rst
quarter 2000 to the third quarter 2004, while the different columns
represent the coefcients for one-month ahead to six-months ahead
hazard rates. Two models were considered: model Adid not include
the economic variables, while model B did include them.
The results are very consistent over the six different forward-
looking hazard rates. Recall that negative coefcients make that
effect more likely to increase the chance of default. So, as would
be expected, the least risky score band (band 5) is the least likely to
default and this risk increases as the behavioural score decreases.
An increase in the interest rate in model B increases the likelihood
of default, while an increase in the GDP decreases the chance of
default. Also, it is clear from the vintage coefcients that the qual-
ity of the borrowers accepted has been getting worse over the time
period under consideration.
Adding economic variables makes very little difference to the dis-
criminationpower of the model as canbe see fromthe ROCcurves in
Figure 9.2. These describe the relative rankingof the default riskover
the next 12 months for a holdout sample, using data as of December
2004, compared with their actual default performance during 2005.
The default risk over the 12-month period for a borrower i, t periods
into the loan, was obtained by using the hazard rate h
i
(t+k) for each
of the k look-ahead models k = 1, . . . , 12. The probability that bor-
rower i would not default in that period is then
12
k=1
(1 h
i
(t +k)).
The economic variables values were taken to be those that actually
occurred in that period so that there is no confusion about errors in
222
Table 9.3 Numbers of predicted and actual defaults in out of time
sample using proportional hazard models
Test Actual no. Expected no.
sample of defaults of defaults
Model size in 2005 in 2005
A 1,4091 959 565
B 1,4091 959 1,022
the economic forecasts. The ROC curves are very similar and sug-
gest that the economic variables make very little improvement to
the relative rankings, which is not surprising since they give the
same values to all borrowers and so the differences in the curves are
because of the subsequent small changes in the coefcients of the
other terms.
On the other hand we can sum the probabilities of default cal-
culated above over all the borrowers in the portfolio to get an esti-
mate of the total default rate in 2005 and compare that with the
actual number of defaults that were recorded. The results are given
in Table 9.3 and show quite clearly how incorporating the economic
variables intothemodel gives amuchbetter forecast of thetotal num-
ber of defaults. Thus, including economics is useful for estimating
portfolio-level default rates but may be less useful for discriminat-
ing who will actually default. In this model, the population odds
adjustment was not applied to the behavioural score because the
economic variables were included separately and they performed a
function somewhat akin to this adjustment.
MARKOV CHAIN MODELS OF BEHAVIOURAL-SCORE
DYNAMICS
Jarrowet al (1997) were one of the rst to use a Markov-chain model
of the dynamics of rating-agency grade as the basis of a reduced-
form approach to corporate credit risk. The idea was to build a
model to forecast multi-period distributions of default rates based
on migration matrixes built on historical data of ratings movements.
Asimilar idea can be used for consumer credit based on behavioural
scores and such models were developed by Malik and Thomas
(2009).
223
MODEL RISK
The rst things to do in constructing the Markov chain that
describes the dynamics of the behavioural scores is to determine
the basic time interval and the state space. The time interval could
be one month, two months, three months and so on, as behavioural
scores are updated monthly. The longer the time interval, the less
important it is that the dynamics are truly Markov. However, there
is also less data available to build the model and thus to estimate
the dependencies between the transitions and other factors like
economic conditions and maturity of the loan.
The next issue is the state space that describes the behavioural
scores. It is easiest to work with nite-state Markov chains and so
we need to split the behavioural score range into a number of inter-
vals together with some special states such as closed account,
inactive account and defaulted. One way of nding such inter-
vals is to rst classify the behavioural scores into a large number
of intervals (say, 10 or 15) based on quantiles of the distribution of
the behavioural scores in the portfolio of borrowers used to develop
the model over the time sample being used. Then coarse classify the
state space by seeing whether adjacent intervals can be combined
into the same state. This is done by looking at the one-step transi-
tions from two adjacent intervals and comparing the distributions
of which nely classied states they go to. This can be checked
using the chi-square statistic, and though almost certainly all pairs
of neighbours will produce statistics which suggest they are differ-
ent at the 95% signicance level the actual value of the statistics
gives a ranking of which pairs of adjacent intervals are closest to
one another in terms of their dynamics. By combining some of the
closest ones we obtain a suitable state space.
The simplest model would be to assume that the Markov chain is
rst order and stationary, so that the transition matrix is the same in
all time intervals, The probability of moving from state i to state j is
p(i, j). This can be estimated using the maximum likelihood estima-
tor n(i, j)/n(i), where n(i, j) is the number of transitions from state
i to state j in the T periods of the sample and n(i) is the number of
times a borrower is in state i in the rst T 1 periods of the sample.
A more complex model would look at whether the current
behavioural score has all the information needed to describe the
future dynamics of a borrower or if some of the borrowers earlier
behavioural scores need to be known. This would mean the Markov
224
chainmight needtobe secondor thirdorder rather thanjust have the
rst-order Markov property. Another possible extension is to allow
the dynamics to be non-stationary and to vary between borrower
segments. Analysis of borrower data suggests that the transitions
can depend both on the age of the loan (the MoB effect) and on eco-
nomic conditions. To allow this, the probabilities of the transition
matrixes can be estimated by using a cumulative or ordered logistic
regression. This is equivalent tosayingthat if aborrower is instatei at
time t, then the dynamics of where the borrower moves to in the next
period is given by an unobservable quantity, the creditworthiness
U
t+1
(i, t) at time t +1 which satises
U
t+1
(i, t) =
K
_
k=2
a
ik
State
t+1k
b
i
EcoVar
t
c
i
MoB
t
d
i
t
State
t+1k
is a vector of indicator variables denoting the borrowers
state at time t + 1 k. EcoVar
t
is a vector of economic variables at
time t. MoB
t
is a vector of indicator variables denoting borrowers
MoB at time t.
t
is a standard logistic distribution and a
i,k
, b
i
, c
i
and d
i
are suitable constants. With such a model the dynamics of
the behavioural score s(t, i) is described by a Kth-order Markov
chain, where the transitions depend on economic variables and on
the length of time the loan has been in operation. We can translate
the distribution of U
t+1
(i, t) into the distribution of which state in the
state space Default, 1, 2, . . . , J the behavioural score moves into by
the relationship
s(t +1) = Default U
t+1
(i, t)
1,i
s(t +1) = 1
1,i
< U
t+1
(i, t)
2,i
.
.
.
s(t +1) = j
j,i
< U
t+1
(i, t)
j+1,i
s(t +1) = J
J,i
< U
t+1
(i, t)
As a case study, Malik and Thomas (2009) applied these ideas to
the UK credit card data that was also used for the proportional haz-
ards approach in the previous section. The time interval chosen was
three months and the state space consisted of ve behavioural-score
intervals [13680], [681700], [70115], [716725] and 726+, together
with states corresponding to defaulted accounts and closed
225
MODEL RISK
Table 9.4 First-order stationary transition matrix
Transition state
, ..
13 681 701 716
Initial to to to to
state 680 700 715 725 726+ Cl. Def.
13680 49.0 22.1 9.6 4.0 4.0 4.7 6.7
(0.2) (0.2) (0.1) (0.1) (0.1) (0.1) (0.1)
681700 15.7 34.7 25.1 9.6 11.2 2.8 0.8
(0.1) (0.2) (0.2) (0.1) (0.1) (0.1) (0.0)
70115 6.0 13.6 35.9 18.12 3.4 2.6 0.5
(0.1) (0.1) (0.2) (0.1) (0.1) (0.1) (0.0)
71625 3.0 6.1 15.7 28.3 44.1 2.5 0.3
(0.1) (0.1) (0.1) (0.2) (0.2) (0.1) (0.0)
726+ 0.7 1.2 2.7 4.3 88.4 2.4 0.2
(0.0) (0.0) (0.0) (0.0) (0.0) (0.0) (0.0)
Cl, closed; Def, default.
accounts. The rst-order stationary model led to a Markov chain
with transition matrix given by Table 9.4.
The results were what you would expect, with the chance of
defaulting in the next time interval being highest for the highest
risk group and lowest for those with the highest behavioural score.
Those with the highest scores were most stable, with 88% staying in
that state in the next period; but, surprisingly, the next most stable
were those in the highest risk group even though just less than half
stayed in that state for another period. The gures in brackets are
the standard sampling errors.
The second model was a second-order Markov chain where the
transitions depended both on the MoB effect and on two lagged
economic variables: interest rate and net consumer lending. The lag
was that suggested by Figlewski et al (2006) and the idea of using
net lending was to give an indication of the state of the consumer
credit market as well as the economic conditions. To describe the
second-order effect, the indicator variables describing the previous
state the borrower was in were includedin the transition matrix. The
coefcients of the model are shown in Table 9.5. Each column gives
the coefcients when the current state is that given as the column
heading. Then the coefcients of the economic variables are given (if
226
Table 9.5 Parameters for second-order Markov chain with age and
economic variables
Initial behavioural score
Parameter
, ..
estimates 13680 Std error 681700 Std error
Interest rate 0.0334 (0.0161) 0.092 (0.0143)
Net lending 0.0129 (0.00489)
MoB
06 0.027 (0.0351) 0.0161 (0.0347)
712 0.2019 (0.0241) 0.1247 (0.0225)
1318 0.2626 (0.0262) 0.2663 (0.0236)
1924 0.07 (0.0275) 0.0796 (0.0251)
2536 0.0015 (0.0244) 0.0521 (0.0223)
3748 0.0703 (0.0262) 0.0519 (0.0243)
49high 0.2957 0.2235 0.13781
Sec. state
13680 0.8372 (0.0165) 0.6762 (0.0168)
681700 0.2365 (0.0201) 0.2847 (0.0139)
701715 0.0111 (0.0249) 0.0491 (0.0168)
716725 0.1647 (0.0345) 0.1764 (0.0239)
726high 0.8979 0.8336
Intercept/Barrier
Default 3.213 (0.0756) 5.4389 (0.0826)
13680 0.2078 (0.0734) 2.179 (0.0657)
681700 1.022 (0.0736) 0.3978 (0.0649)
701715 1.9941 (0.0746) 0.861 (0.065)
716725 2.7666 (0.0764) 1.6267 (0.0656)
LR 3661.078 3379.459
P-value <0.0001 <0.0001
MoB, months on book; LR, likelihood ratio.
no gure is given it means that these were not signicant). The MoB
are the coefcients for the seven indicator variables corresponding
to the seven ranges into which the age of the loan was split. As
we would expect, apart from very new loans, the creditworthiness
improves (ie, is more negative) with increasing age of the loan. The
Sec state refers to the state of the borrower in the previous period.
What happens is that if the borrower had previously had a high
score and then dropped down to a lower one this period, they are
more likely to return to a high score next period than someone who,
227
MODEL RISK
Table 9.5 (Cont.) Parameters for second-order Markov chain with age
and economic variables
Parameter
, ..
estimates 701715 Std error 716-725 Std error
Interest rate 0.0764 (0.0123) 0.0834 (0.0134)
Net lending
MoB
06 0.2182 (0.0368) 0.1637 (0.0448)
712 0.2051 (0.0226) 0.2317 (0.0261)
1318 0.2301 (0.0228) 0.2703 (0.0268)
1924 0.1001 (0.0241) 0.0873 (0.0284)
2536 0.00191 (0.0198) 0.00487 (0.0229)
3748 0.019 (0.0206) 0.0801 (0.0241)
49high 0.13781 0.16603
Sec. state
13680 0.5145 (0.0222) 0.3547 (0.0337)
681700 0.3598 (0.0146) 0.1942 (0.0224)
701715 0.1314 (0.0119) 0.1255 (0.0164)
716725 0.1795 (0.016) 0.0098 (0.0152)
726high 0.8262 0.6842
Intercept/barrier
Default 5.8904 (0.1285) 6.011 (0.0967)
13680 3.2684 (0.1175) 3.6011 (0.0648)
681700 1.9492 (0.1168) 2.461 (0.062)
701715 0.1796 (0.1165) 1.2049 (0.0611)
716725 0.7317 0.171 (0.0609)
LR 4137.587 2838.765
P-value <0.0001
thoughinthe same score bandthis period, was ina lower bandinthe
previous period. Thus, there is no momentum effect. In fact, quite
the opposite is true.
To compare howwell the two models perform, Malik andThomas
(2009) comparedthe predictions of the models withthe actual perfor-
mance on an out-of-time sample. The results are shown in Table 9.6,
where the average matrix is the rst-order stationary model and the
model predicted is the second-order one with economic variables
and age of loan effects. The latter does seem better at predicting the
228
Table 9.5 (Cont.) Parameters for second-order Markov chain with age
and economic variables
Parameter
, ..
estimates 726+ Std error
Interest rate 0.0778 (0.00885)
Net lending
MoB
06 0.0849 (0.0315)
712 0.3482 (0.018)
1318 0.2554 (0.0193)
1924 0.031 (0.0206)
2536 0.0254 (0.0162)
3748 0.00709 (0.0166)
49high 0.51721
Sec. state
13680 0.381 (0.0399)
681700 0.5168 (0.024)
701715 0.2991 (0.0178)
716725 0.0525 (0.0162)
726high 1.2494
Intercept/barrier
Default 5.1834 (0.0506)
13680 3.8213 (0.0436)
681700 2.9445 (0.0421)
701715 2.06 (0.0415)
716725 1.326 (0.0413)
LR 20400.65
P-value <0.0001
default levels, but the former is better at predicting the number who
stay in the highest score band level. However, it is fair to say this test
was done using data from 2005, when the economic situation was
fairly benign and also quite static.
CONCLUSIONS
This chapter has shown how individual borrowers behavioural
scores can be exploited in order to develop a number of modelling
229
MODEL RISK
Table 9.6 Results of default predictions for the two transition matrix
models
1-Period
Behavioural
, ..
score Initial Average Model
segments distribution matrix predicted Observed
13680 1,428 949 1,040 1,199
681700 1,278 1,054 1,117 1,096
701715 1,379 1,291 1,384 1,257
716725 876 1,047 1,178 812
726high 7,514 7,994 7,621 7,968
Default 0 139 134 143
2-Period
Behavioural
, ..
score Average Model
segments matrix predicted Observed
13680 879 983 1,080
681700 978 1,061 1,076
701715 1,262 1,393 1,316
716725 1,051 1,228 774
726high 8,059 7,535 7,943
Default 245 274 286
3-Period
Behavioural
, ..
score Average Model
segments matrix predicted Observed
13680 769 889 1,043
681700 894 996 1,001
701715 1,216 1,363 1,219
716725 1,044 1,234 718
726high 8,208 7,596 8,074
Default 344 397 420
approaches to the credit risk of consumer loan portfolios. Such mod-
els have the advantage that behavioural scores are calculated for
almost all borrowers andare updatedmonthly, soare ubiquitous and
timely. They allow the development of robust models because there
is so much data available (much more than on consumer defaults)
230
and the predictions from these models can be updated every month
when a new behavioural score becomes available.
In all three approaches discussed here the correlation between the
defaults of different borrowers in the same portfolio is obtained by
having the same economic conditions apply to all borrowers. This
is one of the three ways, which is used in corporate credit risk to
model correlations. The others are imposing correlations directly or
using cupolas to connect univariate loss distributions.
The differences in the models are in the way the dynamics are
modelled. In the structural model, behavioural scores are assumed
to be jump-diffusions, while in the transition-matrix approach the
dynamics are given by a Markov chain. The proportional hazard
models concentrate only on the time to default for a two-state
default/not-default model.
There are twoextra features that are important inconsumer credit-
risk models and which are rarely considered in corporate credit-
risk models. The rst is the age of the loan. The default hazard rate
increases substantially in the rst 618 months of the loan and then
gradually drops. This effect is included in the proportional hazards
and the Markov-chain models. It is not included in the structural
model example but could be introduced by making the default cut-
off Kafunctionof thetimeonbooks. Thesecondfeatureis thevintage
effect (ie, when the loan was taken out), which is useful in describing
changes in the lenders lending criteria. Again, this appears directly
inthe proportional-hazards model andthoughit is not inthe Markov
chain or the structural model case studies it could be included: in
the transition matrix in the former case and in the default level K in
the latter case. This effect is also important in Breedens dual time
dynamics (Breeden2007). If it is includedandthe model is to be used
for forecasting future defaults, we need to consider how to forecast
the quality of future vintages when the lenders acceptance policy
is not yet determined.
REFERENCES
Andrade, F. W. M., and L. C. Thomas, 2007, Structural Models in Consumer Credit,
European Journal of Operational Research 183(3), pp. 156981.
Ashcraft, A. B., andT. Schuermann, 2008, Understanding the Securitization of Subprime
Mortgage Credit, Foundations and Trends in Finance 2, pp. 191309.
231
MODEL RISK
Banasik, J., et al, 1999, Not If but When Will Borrowers Default, The Journal of the
Operational Research Society 50(12), pp. 118590.
Measurement and Capital Standards (Basel: Bank of International Settlements).
Breeden, J. L., 2007, Modeling Data with Multiple Time Dimensions, Computational
Statistics and Data Analysis 51(9), pp. 476185.
Bucay, N., and D. Rosen, 2001, Applying Portfolio Credit Risk Models to Retail
Portfolios, Journal of Risk Finance 2, pp. 3561.
Cox, D. R., 1972, Regression Models with Life-Tables (with discussion), Journal of The
Royal Statistical Society Series B 74, pp. 187220.
Crouhy, M. G., et al, 2008, The Subprime Credit Crisis of 2007, The Journal of Derivatives
16, pp. 81110.
Dufe, D., et al, 2007, Multi-Period Corporate Default Prediction with Stochastic
Covariates, Journal of Financial Economics 83(3), pp. 63565.
Figlewski, S., et al, 2006, Modeling the Effect of Macroeconomic Factors on Corporate
Default and Credit Rating Transitions, SSRN eLibrary.
Financial Services Authority, 2009, The Turner Review: A Regulatory Response to the
Global Banking Crisis, Report, Financial Services Authority, London.
Jarrow, R., et al, 1997, AMarkov Model for the Term Structure of Credit Risk Spreads,
Review of Financial Studies 10(2), pp. 481523.
Malik, M., and L. C. Thomas, 2007, Modeling Credit Risk of Portfolio of Consumer
Loans, Journal of the Operational Research Society, DOI: 10.1057/jors.2009.123.
Malik, M., and L. C. Thomas, 2009, Transition Matrix Models for Consumer Credit
Ratings, Credit Research Centre Paper 09-02, University of Southampton.
Musto, D. K., and N. S. Souleles, 2006, A Portfolio View of Consumer Credit, Journal
of Monetary Economics 53, pp. 5984.
Narain, B., 1992, Survival Analysis and the Credit Ganting Decision, in L. C. Thomas,
J. N. Crook and D. B. Edelman (eds), Credit Scoring and Credit Control, pp. 10921 (Oxford
University Press).
Perli, R., and W. I. Nayda, 2004, Economic and Regulatory Capital Allocation for
Revolving Retail Exposures, Journal Of Banking and Finance 28, pp. 789809.
Rosch, D., and H. Scheule, 2004, Forecasting Retail Portfolio Credit Risk, Journal of Risk
Finance 5, pp. 1632.
Saunders, A., and L. Allen, 2002, Credit Risk Measurements: NewApproaches to Value-at-Risk
and Other Paradigms (New York: John Wiley and Sons).
Stepanova, M., and L. C. Thomas, 2001, PHAB Scores: Proportional Hazards Analysis
Behavioural Scores, The Journal of the Operational Research Society 52(9), pp. 100716.
Sy, W., 2008, Credit Risk Models: Why They Failed in the Credit Crisis, Report,
Australian Prudential Regulatory Authority, Sydney.
Zhou, C., 1997, A Jump-Diffusion Approach to Modeling Credit Risk and Valuing
Defaultable Securities, Report, Federal Reserve Board, Washington, DC.
232
10
Validating Structural Credit
Portfolio Models
Michael Kalkbrener, Akwum Onwunta
Deutsche Bank AG
Concentrations in a banks credit portfolio are key drivers of credit
risk. These risk concentrations may be causedby material concentra-
tions of exposure to individual names as well as large exposures to
single sectors (geographic regions or industries) or to several highly
correlated sectors. The most common approach to introduce sec-
tor concentration into a credit-portfolio model is through system-
atic factors affecting multiple borrowers. The specication of the
factor model, which determines the dependence structure of rat-
ing migrations and defaults, is a central problem in credit-portfolio
modelling.
Different data sources can be used as input for the calibration of
factor models, eg, equity data, credit spreads or default and rating
information. The most frequent approach is based on information
from equity markets, mainly due to the good quality and wide cov-
erage of this data source. Since individual rmreturns are available,
methodologies based on equity information allow for rm-specic
pairwise correlations and more granular industry denitions than
historical rating data methods. The main disadvantage of equity
data is the fact that equity prices reect not only the credit quality of
companies but also information that is unrelated to credit risk, eg,
liquidity issues, risk aversion of market participants, etc.
Historical rating and default data, on the other hand, are not dis-
torted by information unrelated to credit quality but require some
aggregation, typically by rating class, country or industry. Hence,
these data sources are natural candidates for validating the depend-
ence structure of credit-portfolio models on an aggregate level. In
this chapter, we review the most frequently used techniques for
233
MODEL RISK
deriving correlations from default and rating data: moment estima-
tors (Lucas 1995; Bahar and Nagpal 2001; de Servigny and Renault
2003; Frey and McNeil 2003; Jobst and de Servigny 2005) and max-
imum likelihood estimators (Gordy and Heiteld 2002; Demey et
al 2004). Our empirical results are based on Standard & Poors
(S&P) rating data from 1981 to 2009. We group the rated companies
into cohorts that correspond to industry segments, rating classes or
combinations of both. Intra-cohort and inter-cohort correlations are
calculated using maximum likelihood estimation. Our results are
broadly in line with previous studies reported in the literature.
Correlations within industry segments are consistently higher
than correlations between industry segments. In particular,
this observation conrms that correlations are usually under-
estimatedif industryinformationis ignoredinthespecication
of the cohorts, ie, if cohorts correspond to rating classes.
Correlations vary strongly across industries and over time.
On average, correlations derived from historical default and
rating data are lower than equity correlations. In addition,
we observed a signicant difference between correlations
obtained from default data and correlations obtained from
rating data.
Most of our calculations are based on the standard assumption that
the risk factors follow a multivariate normal distribution. In addi-
tion, we performcalculations usingt-distributions andskew-normal
distributions. The best calibration results are obtained with a multi-
variate t-distribution with a high degree of freedom, eg, of the order
of 65.
This chapter has the following structure. In the next section we
give a formal specication of the structural credit-portfolio model,
whichis usedfor the analysis of historical default data, together with
a review of the concepts of default and asset correlations. Different
approaches for calibrating andvalidating factor models are thendis-
cussed. The followingsectionis devotedtothe derivationof moment
estimators andmaximumlikelihoodestimators whenappliedto his-
torical default data. The actual calibration results based on S&Pdata
are then presented. We next generalise the credit-portfolio model to
cover rating migration and apply the maximum likelihood estima-
tion to S&P rating data. Calibration results for non-Gaussian risk
234
VALIDATING STRUCTURAL CREDIT PORTFOLIO MODELS
factors are then presented. In addition, we reviewdifferent concepts
for incorporating the credit cycle. We conclude by comparing equity
correlations with correlations derived from historical default and
rating data.
CREDIT PORTFOLIO MODEL FOR DEFAULT RISK
Denition of the model
Credit-portfolio models can be divided into reduced-form models
and structural (or rm-value) models. We refer the reader to Crouhy
et al (2000), Bluhm et al (2002) and McNeil et al (2005) for surveys
on credit-portfolio modelling. In this chapter, we use a structural
credit-portfolio model similar to CreditMetrics (Gupton et al 1997).
The progenitor of all structural models is the model of Merton
(1974), which links the default of a rm to the relationship between
the rms assets andliabilities at the endof a giventime period[0, T].
More precisely, in a structural credit-portfolio model the jth obligor
defaults if at time T its ability-to-pay variable A
j
is below a default
threshold c
j
: the default event is dened as A
j
c
j
, where A
j
is a real-valued random variable on the probability space (, , P)
and c
j
R. The corresponding default indicator is denoted by
I
j
= 1
A
j
c
j
(10.1)
If A
j
is standardised and normally distributed, then the default
threshold c
j
is linked to the default probability p
j
[0, 1] via
c
j
= N
1
(p
j
), where N denotes the distribution function of a stan-
dardised normal distribution.
For each obligor, a loss variable is dened by L
j
= l
j
I
j
, where
l
j
R denotes the loss at default. In order to aggregate the loss
variables L
j
of the individual obligors to a loss variable L of the entire
portfolio, a dependence structure of the A
j
is specied. This is done
via the introduction of a factor model consisting of systematic and
idiosyncratic factors. More precisely, each ability-to-pay variable A
j
is decomposed into a sum of systematic factors
1
, . . . ,
m
and an
idiosyncratic (or specic) factor
j
, that is
A
j
=
_
R
2
j
m
_
i=1
w
ji
i
+
_
1 R
2
j
j
(10.2)
The idiosyncratic factors are independent of each other as well
as independent of the systematic factors. It is usually assumed
235
MODEL RISK
that the idiosyncratic and systematic factors are centred and fol-
low a multivariate Gaussian distribution. The systematic weights
w
j1
, . . . , w
jm
R are scaled such that the systematic component
j
:=
m
_
i=1
w
ji
i
is a standardised normally distributed variable. Each R
2
j
[0, 1]
determines the impact of the systematic component on A
j
and there-
fore the correlation between A
j
and
j
: it immediately follows from
Equation 10.2 that R
2
j
= Corr(A
j
,
j
)
2
. Finally, the portfolio loss vari-
able L : R is the sum of the loss variables of the obligors, that
is
L =
n
_
j=1
L
j
=
n
_
j=1
l
j
I
j
(10.3)
Animportant instanceof this model class is thehomogeneous one-
factor model: assume that all obligors have the same loss-at-default
l R, default probability p with corresponding default threshold
c := N
1
(p) and the same R
2
. In this case, Equations 10.2 and 10.3
specialise, respectively, to
A
j
=
_
R
2
+
_
1 R
2
j
, L =
n
_
j=1
L
j
= l
n
_
j=1
I
j
(10.4)
where is a standardised normally distributed variable.
Default and asset correlations
Via Equation 10.1, the credit-portfolio model species a dependence
structure for the default indicators I
i
. We shall measure depend-
ence by calculating default and asset correlations. This subsection
provides a short review of the denitions and main properties.
The default or event correlation
D
ij
of obligors i and j is dened
as the correlation of the respective default indicators. Because
Var(I
i
) = E(I
2
i
) (E(I
i
))
2
= E(I
i
) (E(I
i
))
2
= p
i
p
2
i
the default correlation equals
D
ij
= Corr(I
i
, I
j
) =
E(I
i
I
j
) p
i
p
j
_
(p
i
p
2
i
)(p
j
p
2
j
)
. (10.5)
236
If the vector of default indicators I = (I
1
, . . . , I
n
) is exchangeable,
1
the default correlation
D
ij
has a particularly simple form(see Equal-
ity10.7). Toshowthis, we rst introduce the notationfor joint default
probabilities
k
:= P(I
i
1
= 1, . . . , I
i
k
= 1), i
1
, . . . , i
k
1, . . . , n, k n
(10.6)
ie,
k
is the probability that an arbitrarily selected subgroup of k
obligors defaults. Note that, for all i, j 1, . . . , n with i }= j
E(I
i
) = P(I
i
= 1) = p
i
=
1
E(I
i
I
j
) = P(I
i
= 1, I
j
= 1) =
2
so that Cov(I
i
, I
j
) =
2
2
1
. This implies that, for all i, j with i }= j,
the default correlation in Equation 10.5 is given by
D
:=
D
ij
=

2
2
1
2
1
(10.7)
which is a simple function of the rst- and second-order default
probabilities.
In the credit-portfolio model introduced in the previous subsec-
tion, the indicator variables I
i
are dened in terms of ability-to-pay
variables A
i
, which are often interpreted as log-returns of asset val-
ues. The correlation Corr(A
i
, A
j
) is therefore called the asset corre-
lation
A
ij
of obligors i }= j. As an immediate consequence of Equa-
tion 10.2, the correlation and the covariance of the ability-to-pay
variables of the counterparties i and j are given by
Corr(A
i
, A
j
) = Cov(A
i
, A
j
) =
_
R
2
i
_
R
2
j
m
_
k,l=1
w
ik
w
jl
Cov(
k
,
l
) (10.8)
Note that to specify the covariance structure of all n ability-to-pay
variables it sufces to dene the covariance matrix of the m system-
atic factors and the systematic weights and R
2
values of all coun-
terparties. Since m is usually much smaller than n, the factor-model
approachleads toa signicant reductioninthe amount of input data,
ie, the number of parameters is of the order of m(m+n) insteadof n
2
.
There exists an obvious link between default and asset correla-
tions. For given default probabilities, the default correlations
D
ij
are
determined by E(I
i
I
j
) according to Equation 10.5. It follows from
Equation 10.1 that
E(I
i
I
j
) = P(A
i
c
i
, A
j
c
j
) =
_
c
i
_
c
j
f
ij
(u, v) dudv
237
MODEL RISK
where f
ij
(u, v) is thejoint densityfunctionof A
i
andA
j
. Hence, default
correlations dependonthe joint distributionof A
i
andA
j
. If (A
i
, A
j
) is
bivariate normal, the correlation of A
i
and A
j
determines the copula
of their joint distribution and hence the default correlation
E(I
i
I
j
) =
1
2
_
1 (
A
ij
)
2
_
c
i
_
c
j
exp
_
u
2
2
A
ij
uv +v
2
2(1 (
A
ij
)
2
)
_
dudv
(10.9)
Note that, for general ability-to-pay variables outside the multivari-
ate normal class, the asset correlations do not fully determine the
default correlations.
Counting defaults
Let D : R specify the number of defaults, ie, D :=

n
j=1
I
j
. We
shall now derive a formula for the distribution of D, which will be
used in the calibration techniques presented later. The derivation is
based on the fact that the vector of default indicators I = (I
1
, . . . , I
n
)
satises the denition of a Bernoulli mixture model if m < n (McNeil
et al 2005). More precisely, there exist functions p
j
: R
m
[0, 1],
1 j n, such that, conditional on = (
1
, . . . ,
m
), the default
indicator I = (I
1
, . . . , I
n
) is a vector of independent Bernoulli random
variables with
P(I
j
= 1 = ) = p
j
() for R
m
For a given value of the systematic factors, the conditional inde-
pendence of the I
j
follows from the independence of the
j
. The
functions p
j
have the form
p
j
() = P(A
j
c
j
= )
= P
_
j

c
j
_
R
2
j
m
i=1
w
ji
i
_
1 R
2
j
_
= N
_
c
j
_
R
2
j
m
i=1
w
ji
i
_
1 R
2
j
_
since the
j
are N(0, 1)-distributed variables.
For the rest of this subsection we make the simplifying assump-
tion that the vector of default indicators I = (I
1
, . . . , I
n
) is exchange-
able. As a consequence, the functions p
1
, . . . , p
n
are identical and
238
we therefore drop the subscript. Conditional on = , the num-
ber of defaults D is the sum of n independent Bernoulli variables
with parameter p(). Hence, D has a binomial distribution with
parameters p() and n, ie
P(D = j = ) =
_
n
j
_
p()
j
(1 p())
nj
(10.10)
The unconditional distribution of D is obtained by integrating over
the distribution F of the systematic factors , that is
P(D = j) =
_
R
m
P(D = j = ) dF()
=
_
n
j
_
_
R
m
p()
j
(1 p())
nj
dF()
In particular, for the homogeneous one-factor model dened in
Equation 10.4, the vector of default indicators I = (I
1
, . . . , I
n
) is
exchangeable and the unconditional distribution of D is
P(D = j) =
_
n
j
_
_
R
p()
j
(1 p())
nj
dN()
=
1
(2)
1/2
_
n
j
_
_
R
p()
j
(1 p())
nj
e
2
/2
d (10.11)
where p is specied by
p() = N
_
c
R
2
1 R
2
_
(10.12)
CALIBRATION ANDVALIDATION OF FACTOR MODELS
The calibration problem
The central problem in credit-portfolio modelling is the specica-
tion of the dependence structure of defaults. In the quantitative
frameworkintroducedinthe previous sectionthe probabilityof joint
defaults is specied through a factor model consisting of systematic
andidiosyncratic factors. The calibrationof the factor model consists
of the following steps:
(i) identication of appropriate systematic factors
1
, . . . ,
m
;
(ii) estimation of the correlations of
1
, . . . ,
m
;
(iii) estimation of the systematic weights w
ji
and R
2
j
, where j =
1, . . . , n and i = 1, . . . , m.
239
MODEL RISK
Historical time series of default indicators or of ability-to-pay vari-
ables are typically used as inputs for the calibration procedure. For
example, observations (a
1
(1), . . . , a
n
(1)), . . . , (a
1
(s), . . . , a
n
(s)) R
n
of equityor asset returns at different time points are ofteninterpreted
as realisations of the ability-to-pay variables A = (A
1
, . . . , A
n
).
Several statistical techniques can be applied to the calibration of
the factor model. These techniques estimate the parameters of the
factor model in order to replicate the dependence of the input time
series as closely as possible. Additional requirements arise from the
application of credit-portfolio models in risk management. In par-
ticular, the systematic factors should have a clear economic interpre-
tation, such as representing individual countries or industries. Fur-
thermore, all obligors should have plausible factor weights. These
features are important for the economic interpretation of risk con-
centrations identiedbythe portfoliomodel. Theyare alsonecessary
for the implementation of economic stress scenarios in the model.
These non-statistical model requirements are one reason why stat-
istical techniques for factor-model calibration are usually comple-
mentedbyqualitative assessments. For example, countryandindus-
tryweights maybe derivedfrombalance sheet andlegal information
to avoid counterintuitive results from regression on market data.
Note, however, that there is typically a trade-off between the accu-
rate replication of the dependence structure of the input data and
the economic interpretation of the model.
Calibration with equity data
The calibration of factor models can be classied according to the
data source used as input, ie, equity data, credit spreads or default
and rating information. The most frequent approach is based on
information fromequity markets. The main advantage is that equity
information has wider coverage in terms of both industry and geog-
raphy. The informationcomes fromliquidmarkets, is of goodquality
andis readily available. Thus, unlike the ratings-basedapproach, we
can empirically observe nuances between rms on a global basis.
Since individual rm returns are available, methodologies based
on equity information allow for rm-specic pairwise correlations
and more granular industry denitions than historical rating data
methods. Withthese methodologies, however, there is a concernthat
240
equity prices reect (in addition to credit-related information) infor-
mation that is unrelated to credit risk, such as liquidity issues, risk
aversion of market participants, etc.
Another important issue is the question of whether equity or asset
returns canbe usedfor the calibrationof ability-to-payvariables. Fol-
lowing Merton (1974), a company defaults if the value of its assets
is no longer sufcient to cover its liabilities. Default correlations are
therefore determined by the corresponding asset correlations and
default probabilities. Since asset processes cannot be observed in the
market and are difcult to determine, an alternative approach is to
derive default correlations from equity correlations instead of asset
correlations. However, this raises the questionof whether equitycor-
relations are reliable proxies for asset correlations. The relationship
between asset and equity correlations is still a controversial issue
in the literature, since authors come to rather different conclusions
(see, for example, Zeng and Zhang 2002; de Servigny and Renault
2003). In our concluding remarks (see page 258), we will compare
equity correlations to correlations derived from historical default
and rating data.
Validation with historical default information
The main objective of the factor model in Equation 10.2 is the speci-
cationof the dependence structure of default indicators I
1
, . . . , I
n
. It is
therefore natural to use historical default data for model calibration.
Compared with other data sources such as equity data, the use of
historical defaults has the important advantage that the calibration
results are not distorted by information unrelated to credit quality.
Time series of historical defaults are provided by rating agencies for
a large pool of companies with a particularly good coverage in the
US. Despite these advantages, it is uncommon in the industry to rely
on historical default data for formal statistical estimation of model
parameters. There are good reasons for this, the main one being that,
particularly for higher-rated companies, defaults are rare events. In
other words, there is simply not enough relevant data on historical
defaults to obtain reliable parameter estimates for individual com-
panies by formal inference alone. As a consequence, methodologies
based on historical default data require some aggregation, typically
by rating class, country or industry. Hence, the main application of
default data is the validation of portfolio models calibrated to equity
data or credit spreads.
241
MODEL RISK
ESTIMATORS APPLIEDTO DEFAULT DATA
Homogeneous cohorts
In the following, we analyse different techniques for estimating
dependence in time series of default data. The methods we describe
aremotivatedbytheformat of thedataweconsider, whichcomprises
observations of the default or non-default of groups of monitored
companies in a number of time periods t = 1, . . . , s. This kind of
data is readily available from rating agencies. In each time period
t = 1, . . . , s we group the monitored companies into cohorts. Let
n(t) denote the number of obligors in a given cohort at the start of
period t and let d(t) be the number of observed defaults during that
period. The companies within a cohort are assumed to be homo-
geneous in the following sense: conditionally on systematic factors
(t) = (
1
(t), . . . ,
m
(t)) their default indicators I
1
(t), . . . , I
n(t)
(t)
are independent and identically distributed. Hence, in each time
period the defaults of a cohort are generated by an exchangeable
Bernoulli mixture model. Furthermore, the random variable D(t)
that counts the number of defaults is conditionally binomially dis-
tributed and its unconditional distribution has the form given in
Equation 10.11. Note that d(1), . . . , d(s) are observations of the ran-
dom variables D(1), . . . , D(s). Finally, we assume that the vectors of
systematic factors (1), . . . , (s), as well as the random variables
D(1), . . . , D(s), are independent.
We consider two methods for estimating the dependence struc-
ture of default indicators I
1
(t), . . . , I
n(t)
(t), t = 1, . . . , s, from the time
series d(1), . . . , d(s) of default counts. The method of moments cal-
culates the joint default probabilities
k
specied in Equation 10.6
within a cohort (intra-correlations) and between different cohorts
(inter-correlations). The maximumlikelihood method uses the para-
metric form of the conditional probability function p specied in
Equation 10.12 to calibrate the parameters of the factor model.
Moment estimator
The application of moment estimators is basedon the following gen-
eralisation of the default count D(t) (McNeil et al 2005). Dene the
random variable
D
(k)
(t) :=
_
i
1
,...,i
k
1,...,n(t)
I
i
1
(t) I
i
k
(t) =
_
D(t)
k
_
(10.13)
242
representing the number of possible subgroups of k obligors among
the defaulting obligors in period t. Obviously, D
(1)
(t) equals D(t).
By taking expectations in Equation 10.13 we get
E(D
(k)
(t)) =
_
n(t)
k
_
k
where
k
species the joint default probability of an arbitrarily
selected subgroup of k exchangeable obligors (see Equation 10.6).
Hence, the unknown default probability
k
can be easily estimated
from the rst moment of the generalised default count D
(k)
(t) by
taking the empirical average of s years of data. More precisely, the
estimator for
k
is dened by

k
:=
1
s
s
_
t=1
D
(k)
(t)
_
_
n(t)
k
_
=
1
s
s
_
t=1
_
D(t)
k
_
_
_
n(t)
k
_
For a given time series of default counts d(1), . . . , d(s) we obtain

k
=
1
s
s
_
t=1
d(t)(d(t) 1) (d(t) k +1)
n(t)(n(t) 1) (n(t) k +1)
For k = 1, 2 we get the estimators

1
=
1
s
s
_
t=1
d(t)
n(t)
,
2
=
1
s
s
_
t=1
d(t)(d(t) 1)
n(t)(n(t) 1)
(10.14)
The default correlation
D
can be estimated by Equation 10.7. The
estimator
k
is unbiased and consistent (Frey and McNeil 2001).
However, de Servigny and Renault (2003) have argued that for low-
default cohorts the estimator
2
might lead to spurious negative
correlations and proposed the following (biased) estimator instead

2
=
1
s
s
_
t=1
d(t)
2
n(t)
2
(10.15)
The estimators
2
and
2
are used for calculating the default cor-
relation of two obligors in a homogeneous cohort. The methodology
can easily be extended to default correlations of obligors in different
cohorts. More precisely, let n
1
(t) andn
2
(t) be the number of observed
companies at the start of period t in cohort 1 and in cohort 2, respec-
tively. The corresponding numbers of defaults are denoted by d
1
(t)
andd
2
(t). Assume that obligor i is inthe rst cohort andobligor j is in
243
MODEL RISK
the second cohort. Their joint default probability E(I
i
I
j
) is estimated
by
1
s
s
_
t=1
d
1
(t)d
2
(t)
n
1
(t)n
2
(t)
(10.16)
and the corresponding default correlation is specied by Equa-
tion 10.5.
Maximum likelihood estimator
To implement a maximum likelihood procedure we assume a para-
metric form for the function p
b
: R
m
[0, 1] with parameters
b = (b
1
, . . . , b
j
), where p
b
() determines the binomial distribution
(Equation10.10) of the default count D(t) conditionallyon(t) = .
The likelihood function L(b d) is the product of the uncondi-
tional probabilities of the default counts D(1), . . . , D(s) evaluated
at observed defaults d = (d(1), . . . , d(s)), that is
L(b d) =
s
t=1
P(D(t) = d(t))
=
s
t=1
_
n(t)
d(t)
_
_
R
m
p
b
()
d(t)
(1 p
b
())
n(t)d(t)
dF()
where F is the distribution function of the systematic factors . The
parameters b = (b
1
, . . . , b
j
) can be calculated by maximising the like-
lihood function L(b d) or the log-likelihood function log(L(b d))
with respect to b. The log-likelihood function
log(L(b d)) =
s
_
t=1
log(P(D(t) = d(t))) (10.17)
has the same maxima as L(b d) but its maximisation is usually
easier from a numerical point of view.
We shall now apply the maximum likelihood estimator to the
homogeneous one-factor model (Equation 10.4). More precisely, we
assume that the homogeneous default probabilityp has alreadybeen
specied. Hence, the default threshold is determined by c = N
1
(p).
We will apply the maximum likelihood estimation to the remaining
model parameter, the R
2
. It follows from the unconditional distri-
bution (Equation 10.11) of D(t) that the likelihood function has the
244
following form
L(R
2
d) =
s
t=1
_
n(t)
d(t)
_
_
R
p
R
2 ()
d(t)
(1 p
R
2 ())
n(t)d(t)
dN()
(10.18)
where
p
R
2 = N
_
c
R
2
1 R
2
_
The integrals in Equation 10.18 are easily evaluated numerically.
Again, the methodology can be extended to default correlations
of obligors in different cohorts. For example, consider two homo-
geneous groups of obligors, eg, cohorts 1 and 2. We assume that all
obligors in cohort i dependonly on the systematic factor
i
andhave
the same default threshold c
i
and R
2
i
. The systematic factors
1
and
2
are standardised and follow a bivariate normal distribution N
with correlation .
We assume that observations for s time periods are given: the
cohort size and the number of defaults in the rst and second cohort
are denoted by n
1
(t), n
2
(t) and d
1
(t), d
2
(t), respectively; the corre-
sponding count variables are specied by D
1
(t) and D
2
(t). We apply
the maximum likelihood estimation to the parameters R
2
1
, R
2
2
, and
obtain the likelihood function
L(R
2
1
, R
2
2
, d
1
, d
2
)
=
s
t=1
_
R
2
2
i=1
P(D
i
(t) = d
i
(t)
i
=
i
) dN
(
1
,
2
)
=
s
t=1
_
R
2
2
i=1
_
n
i
(t)
d
i
(t)
_
p
R
2
i
(
i
)
d
i
(t)
(1 p
R
2
i
(
i
))
n
i
(t)d
i
(t)
dN
(
1
,
2
)
(10.19)
MODEL VALIDATION BASED ON DEFAULT DATA
Correlations of rating cohorts
We now apply the estimators presented in the previous section to
Standard & Poors data set CreditPro consisting of data from July
1981 to June 2009. Each of the companies included in this data set
is rated according to the eight rating categories AAA, AA, A, BBB,
BB, B, CCC and Default and classied with respect to 13 industry
segments.
245
MODEL RISK
Table 10.1 Asset correlations of rating cohorts
intra-corr. (%) inter-corr.
, ..
(%)
A BBB BB B CCC Avg Avg
Moment estimator 7.8 2.8 7.8 8.9 7.2 6.9
(10.14)
Moment estimator 13.6 8.2 10.5 9.9 12.3 10.9
(10.15)
Moment estimator 4.7
(10.16)
ML estimator 11.8 6.2 9.5 11.1 7.9 9.3
(10.18)
ML estimator 5.9
(10.19)
Numbers in parentheses refer to equation numbers.
In this subsection, we ignore industry information and partition
the set of rated companies into seven cohorts that correspond to
rating classes AAA to CCC. We model each cohort as a homoge-
neous one-factor model: the ability-to-pay variable of each obligor j
in rating class k = AAA, . . . , CCC has the form
A
j
=
_
R
2
k
k
+
_
1 R
2
k
j
(10.20)
where
AAA
, . . . ,
CCC
are the seven systematic factors of the model.
The default threshold c
k
is derived from the average default rate of
the companies in rating class k.
Twelve time series with non-overlapping one-year (1y) observa-
tion periods are constructed for each rating class, say BB. The time
series cover 27 years; their starting dates are shifted by one month:
the rst time series covers July 1981 to July 2008, the second starts in
August 1981 and the last time series covers June 1982 to June 2009.
Each of these BB time series d(1), . . . , d(27) counts the number of
defaults d(t) of those companies that had a BB rating at the begin-
ningof periodt anddefaultedwithinthe next 12 months. Correlation
estimates are derived separately fromevery BB time series and then
averaged. This procedure is motivated by the observation that the
correlation estimates are rather sensitive to the precise bucketing of
defaults. For example, correlation estimates derived from data sets
246
Table 10.2 intra-correlations (%) of industry cohorts
Industry Ind/Rat Ind/Rat Demey Demey
1y 1y 3y 1F 2F
Aerospace/ 8.2 8.9 11.5 11.6 11.2
automotive
Consumer/ 6.4 9.7 7.5 7.5 8.7
service sector
Energy and 16.5 26.2 23.9 11.5 21.3
natural resources
Financial 12.6 15.8 19.9 12.2 15.7
institutions
Forest and 10.9 17.1 10.5 14.5 6.8
building products
Health care/ 8.5 11.3 12.3 9.2 8.3
chemicals
High 14.5 17.7 16.6 4.7 6.8
technology
Insurance 9.2 10.3 11.4 7.6 12.2
Leisure time/ 10.8 18.0 20.6 7.0 7.0
media
Real estate 44.7 51.0 59.1 27.7 35.9
Telecommunications 34.9 31.1 27.5 34.3 27.1
Transportation 7.8 11.2 9.5 8.3 6.8
Utility 27.1 29.8 26.8 21.2 18.3
Average 16.3 19.8 19.8 13.6 14.3
July 1981 to July 2008 and January 1982 to January 2009 might differ
signicantly.
Since the default indicators of each rating cohort form an ex-
changeable Bernoulli mixture model, we can use the moment
estimators
2
and
2
dened in Equations 10.14 and 10.15 for cal-
culating the joint default probabilities within each rating cohort.
The corresponding asset correlations are derived fromequality 10.9.
Asset correlations between different cohorts are calculated with esti-
mator 10.16 andequality10.9. intra-correlations for the ratingclasses
A, BBB, BB, B and CCC as well as the average inter-correlation are
given in the rst three rows of Table 10.1.
2
We now turn to maximum likelihood (ML) estimation. Note that
each rating cohort species a likelihood function given by Equation
10.18. The maximumlikelihoodestimate of eachR
2
k
inEquation10.20
247
MODEL RISK
is calculated by maximising this likelihood function.
3
In a second
step, the correlation between two systematic factors is estimated
by maximising Equation 10.19. The results obtained with both ML
estimators are shown in rows 4 and 5 of Table 10.1.
In the Basel II Accord (Basel Committee on Banking Supervision
2006) asset correlations for corporates, sovereigns and banks are
modelled as a decreasing function of default probabilities. The cor-
relation estimates in Table 10.1 do not support a systematic depend-
ence between asset correlations and credit quality. Furthermore, the
estimates for intra-correlations are smaller thanthe Basel II asset cor-
relations for large corporates, which range between 12% and 24%.
The difference is even more pronounced for inter-correlations. In
the following subsection we will show that these differences can be
partly reconciled by taking industry information into account.
Note that the estimates calculatedwiththe unbiasedmoment esti-
mator (Equation 10.14) are always lower than the MLestimates. The
biasedmoment estimator (Equation10.15) produces the highest esti-
mates for all but one rating class. These results seem to conrm that
the rst moment estimator tends to underestimate correlations in
small or low-default cohorts, whereas the second moment estima-
tor overestimates correlations for these cohorts.
4
These observations
are in line with a Monte Carlo study in Gordy and Heiteld (2002),
where MLestimators outperform moment estimators. Hence, in the
rest of the chapter we will focus on maximumlikelihood estimation.
Correlations of industry cohorts
The correlation estimates in Table 10.1 were calculated without
using the S&P industry classication. We now apply the same MLE
methodology to 13 cohorts that correspond to the S&P industry seg-
ments. The estimates for the 13 intra-correlations are presentedinthe
rst column of Table 10.2. Demey et al (2004) and Demey and Ron-
calli (2004) apply the same set-up to S&P data covering the 22 years
from1981 to 2002. Inadditionto the one-factor MLEmodel, theyalso
present an ML estimator based on two systematic factors for each
cohort. Both sets of results are given in columns 4 and 5 of Table 10.2
for comparison.
Note that each of these industry cohorts consists of companies
with different ratings. In order to ensure that the rating proles of
the cohorts are homogeneous, we partitioneachindustry cohort into
248
seven subcohorts corresponding to the different rating classes. More
precisely, the improved model has the following form.
(i) Ninety-one cohorts are specied by taking all combinations of
seven rating classes and 13 industries.
(ii) The model has 13 systematic factors
i
that correspond to the
13 industries.
(iii) All companies in an industry depend only on the systematic
industry factor and have the same R
2
. More formally, if a com-
pany j is in industry i, then the ability-to-pay variable A
j
has
the form
A
j
=
_
R
2
i
i
+
_
1 R
2
i
j
(iv) The default threshold of a cohort is specied by the average
default rate of the companies in the cohort.
Conditions (ii) and (iii) ensure that only 13 R
2
values and 78 corre-
lation parameters have to be estimated. The ML estimate of each R
2
i
is derived from the seven time series corresponding to the different
rating cohorts of the ith industry. In order to apply the MLestimator
simultaneously to seven cohorts, the maximum likelihood function
(Equation 10.18) is generalised to
L(R
2
i
d
i1
, . . . , d
i7
)
=
s
t=1
_
R
7
k=1
_
n
ik
(t)
d
ik
(t)
_
p
R
2
i
()
d
ik
(t)
(1 p
R
2
i
())
n
ik
(t)d
ik
(t)
dN()
(10.21)
where d
ik
and n
ik
are the time series of observed defaults and cohorts
size for i = 1, . . . , 13 and k = 1, . . . , 7. The maximumlikelihood func-
tion(Equation10.19) for estimatingthe correlations of the systematic
factors is generalised in the same way. We will use this MLE model
in all subsequent calculations.
The second column in Table 10.2 displays the R
2
estimates
obtainedby maximising the likelihoodfunction (Equation 10.21). As
in all previous calculations, the underlying time series consist of 27
default counts for one-year observationperiods. It is obvious that the
defaults are concentrated in the low rating classes since each obser-
vationperiodcovers onlyone year. Incontrast, the results inthe third
column are obtained by applying the MLestimator (Equation 10.21)
to time series with three-year observation periods.
249
MODEL RISK
Acomparison of the results in Tables 10.1 and 10.2 clearly shows
that the average intra-correlationof the industrycohorts inTable 10.2
is signicantly higher than the average intra-correlation of the rating
cohorts inTable 10.1. The increasedcorrelations seemtobe causedby
the higher homogeneity of the industry cohorts. This explanation is
also supported by a comparison of the rst and the second columns
in Table 10.2: for all but one industry segment, intra-correlations
increasedif the estimates are derivedfromcohorts that correspondto
combinations of industries andratings. The thirdcolumnshows that
the intra-correlations are relatively stable if the same ML estimator
is applied to three-year observation periods.
On average, the intra-correlations calculated in Demey et al (2004)
andDemeyandRoncalli (2004) are lower (see alsoJobst andde Servi-
gny 2005). The main reason is that correlations increased for most
industries inthe period20039, whichis not coveredinthese papers.
In fact, correlations vary signicantly over time, as illustrated in
Table 10.3. The correlation estimates in Table 10.3 were computed
for time series of non-overlapping one-year observationperiods that
cover 18 years. The 18-year periods are shifted by one year, ie, the
rst period covers 198299, the last period covers 19912008. The
calculations are performed for all industry segments. For exam-
ple, Table 10.3 shows that, for high technology, the initial intra-
correlation of 6.3% in the period 198299 increases to 17.7% in the
last period (19912008).
In addition to intra-correlations, we calculated inter-correlations
for each pair of the 13 industries using the same model as in the
second column of Table 10.2. The average inter-correlation is 10.1%,
which is approximately half the average intra-correlation of 19.8%.
VALIDATION BASED ON RATING DATA
Credit-portfolio model for default and migration risk
The portfolio model denedonpage 235 distinguishes only between
two states: default and non-default. We shall now extend the model
to cover rating migrations, which provides the appropriate frame-
work for deriving correlation estimates from time series of rating
data. More precisely, instead of focusing on defaults within a given
cohort, we shall now consider time series that specify the entire
rating distribution of a cohort at the end of each observation period.
250
V
A
L
I
D
A
T
I
N
G
S
T
R
U
C
T
U
R
A
L
C
R
E
D
I
T
P
O
R
T
F
O
L
I
O
M
O
D
E
L
S
Table 10.3 Variation (%) of intra-correlations over time
Time period
, ..
8299 8300 8401 8502 8603 8704 8805 8906 9007 9108
Aerospace/automotive 4.5 5.3 10.0 9.3 9.0 8.2 8.2 7.8 9.2 10.2
Consumer/service sector 5.2 6.1 7.9 7.5 6.9 6.1 6.8 9.0 10.3 10.2
Energy and natural resources 21.2 21.4 20.7 19.0 18.2 6.4 7.2 10.3 14.1 14.4
Financial institutions 14.1 17.7 14.9 14.3 16.1 19.6 20.8 21.8 17.6 14.2
Forest and building products 16.6 15.1 14.5 15.0 14.5 15.1 18.3 19.7 19.4 17.4
Health care/chemicals 6.4 9.2 7.8 5.7 4.8 7.4 8.8 12.5 14.7 13.5
High technology 6.3 6.6 8.5 8.7 9.8 13.9 14.7 13.6 16.4 17.7
Insurance 14.0 13.3 11.3 9.8 8.8 7.2 9.3 9.3 11.5 10.4
Leisure time/media 13.4 12.6 10.9 8.9 10.3 14.0 14.9 17.4 17.8 13.5
Real estate 44.9 52.6 54.0 54.1 57.4 60.1 62.3 66.4 66.3 51.5
Telecommunications 11.7 7.6 19.7 22.9 23.7 25.6 29.4 30.0 32.0 38.3
Transportation 9.4 9.6 8.5 10.4 10.6 12.2 10.6 10.2 13.2 12.1
Utility 42.8 30.2 18.9 25.6 22.3 25.0 24.6 27.0 28.8 29.8
2
5
1
MODEL RISK
Let r be the number of rating classes including default and dene
the thresholds
= c
r,j
c
r1,j
c
1,j
c
0,j
= , j = 1, . . . , n
for each of the n obligors. The event c
k,j
< A
j
c
k1,j
is interpreted
as the jth obligor being in rating class k at the end of the time period
[0, T]. The corresponding migration probabilities
p
k,j
:= P(c
k,j
< A
j
c
k1,j
) (10.22)
are usually taken from a rating migration matrix. Note that the
default probability is dened by p
r,j
= P(A
j
c
r1,j
). The vector
(l
1,j
, . . . , l
r,j
) R
r
with l
1,j
l
r,j
species the loss l
k,j
if the jth
obligor migrates toratingclass k.
5
Hence, the loss variable L
j
: R
of the jth obligor is generalised to
L
j
:=
r
_
k=1
l
k,j
1
c
k,j
<A
j
c
k1,j
The denition of portfolio loss remains unchanged, ie, L =

n
j=1
L
j
.
Note that if r = 2, then the multi-state model is just the two-state
model.
Counting rating migrations
The objective of this subsection is the generalisation of the one-
dimensional random variable D that counts defaults to an r-
dimensional variable M that counts rating migrations. First of all,
we dene for k = 1, . . . , r and j = 1, . . . , n a rating indicator
I
k,j
:= 1
c
k,j
<A
j
c
k1,j
(10.23)
Note that I
r,j
is the default indicator. Let
M
k
:=
n
_
j=1
I
k,j
specify the number of obligors in rating class k at time T.
For the rest of this subsection, we assume that the n-dimensional
vector
((I
1,1
, . . . , I
r,1
), . . . , (I
1,n
, . . . , I
r,n
)) (10.24)
is exchangeable. As a consequence, each obligor has the same vec-
tor of migration thresholds (c
0
, . . . , c
r
) and there exist functions
252
p
k
: R
m
[0, 1], 1 k r, which specify the conditional migration
probabilities for all obligors, ie
p
k
() = P(c
k,j
< A
j
c
k1,j
= )
for all R
m
and j = 1, . . . , n.
Wewill nowcomputeconditional andunconditional distributions
of M = (M
1
, . . . , M
r
). Conditional on = , the probability that M
equals j = (j
1
, . . . , j
r
) with
r
k=1
j
k
= n is given by
P(M = j = ) =
n!
r
k=1
j
k
!
r
k=1
p
k
()
j
k
The unconditional distribution of M is obtained by integrating over
the distribution F of the systematic factors , that is
P(M = j) =
_
R
m
P(M = j = ) dF()
=
n!
r
k=1
j
k
!
_
R
m
r
k=1
p
k
()
j
k
dF()
As in the two-state case, we consider the example of a homo-
geneous one-factor model: assume that all obligors have the same
migration probabilities p
1
, . . . , p
r
with corresponding thresholds
c
0
, . . . , c
r
speciedby Equation 10.22 andthe same R
2
. For this model
class the vector of rating indicators (Equation 10.24) is exchangeable
and the unconditional distribution of M equals
P(M = j) =
n!
r
k=1
j
k
!
_
R
r
k=1
p
k
()
j
k
dN()
=
1
(2)
1/2
n!
r
k=1
j
k
!
_
R
r
k=1
p
k
()
j
k
e
2
/2
d (10.25)
where p
k
is specied by
p
k
() = N
_
c
k1
R
2
1 R
2
_
N
_
c
k
R
2
1 R
2
_
(10.26)
Maximum likelihood estimator for rating data
In the previous sections, moment estimators and MLE techniques
were presented in order to derive parameters of the factor model
from default data. We shall now generalise maximum likelihood
estimation to rating data. The replacement of default data by rating
data signicantly increases the amount of information available for
253
MODEL RISK
calibration of the model: we now infer dependence not only from
joint defaults but also from rating upgrades and downgrades. This
is an important improvement, particularly for good rating classes,
which are characterised by a low number of defaults.
The set-up is similar to that above (see page 242). In each time
period t = 1, . . . , s we group the monitored companies into cohorts.
Let n(t) denote the number of obligors ina givencohort at the start of
periodt andm(t) = (m
1
(t), . . . , m
r
(t)) the rating vector of the cohort.
More precisely, m
k
(t) species the number of obligors inratingclass k
at the end of period t. The companies within a cohort are assumed to
be homogeneous, ie, conditionallyonsystematic factors their vectors
of rating indicators (I
1,1
, . . . , I
r,1
), . . . , (I
1,n
, . . . , I
r,n
) are independent
and identically distributed.
The specication of the likelihood function follows a similar ap-
proach to the case of MLE applied to default data (see page 244). The
main difference is that the variables D(t), which count defaults, are
replaced by the more general variables M(t) = (M
1
(t), . . . , M
r
(t)),
whichspecifythe number of companies inthe different ratingclasses
at the end of period t.
We shall now apply the maximum likelihood estimator to the
homogeneous one-factor model. Assuming that the homogeneous
migration probabilities p
1
, . . . , p
r
with corresponding thresholds
c
0
, . . . , c
r
have already been specied, it remains to apply the maxi-
mum likelihood estimation to the homogeneous R
2
. It follows from
the unconditional distribution (Equation 10.25) of M(t) that the
likelihood function has the form
L(R
2
m) =
s
t=1
n(t)!
r
k=1
m
k
(t)!
_
R
r
k=1
p
k,R
2 ()
m
k
(t)
dN() (10.27)
where p
k,R
2 is specied by
p
k,R
2 () = N
_
c
k1
R
2
1 R
2
_
N
_
c
k
R
2
1 R
2
_
(10.28)
Finally, we derive the likelihood function for the model consist-
ing of two homogeneous cohorts that was specied on page 245.
We assume that we have observations for s time periods: the rating
vector for cohort i = 1, 2 and time period t = 1, . . . , s is denoted
by m
i
(t) = (m
1i
(t), . . . , m
ri
(t)). The likelihood function L(R
2
1
, R
2
2
,
254
Table 10.4 intra-correlations (%) using rating data
Rating Default
Aerospace/automotive 5.5 8.9
Consumer/service sector 3.0 9.7
Energy & natural resources 11.4 26.2
Financial institutions 13.4 15.8
Forest & building products 5.9 17.1
Health care/chemicals 5.0 11.3
High technology 5.4 17.7
Insurance 13.9 10.3
Leisure time/media 9.8 18.0
Real estate 29.0 51.0
Telecommunications 10.3 31.1
Transportation 7.6 11.2
Utility 6.1 29.8
Average 9.7 19.8
m
1
, m
2
) for the parameters R
2
1
, R
2
2
, equals
s
t=1
_
R
2
2
i=1
n
i
(t)!
r
k=1
m
k,i
(t)!
r
k=1
p
k,R
2
i
(
i
)
m
k,i
(t)
dN
(
1
,
2
),
where p
k,R
2
i
is specied by Equation 10.28.
Model validation with rating data
We repeat the calculations presented in the second column of
Table 10.2 but replace the time series of defaults by time series of
rating migration vectors. The ML estimation is based on the likeli-
hoodfunctioninEquation10.27. The results are displayedinthe rst
column of Table 10.4. The second column is given for comparison
and presents the corresponding estimates derived fromdefault data
(column 2, Table 10.2).
The intra-correlations derived from the rating time series are
lower for all but one industry. The average intra-correlation is
reduced by more than 50%, from 19.8% to 9.7%. inter-correlations
show the same behaviour: the average inter-correlation is reduced
from 10.1% to 4.5%. These estimates are in line with the results
in the literature.
6
We shall discuss potential explanations for the
striking difference between correlations obtained from rating data
255
MODEL RISK
Table 10.5 Variation of intra-correlation (%) and log-likelihood function
with degrees of freedom
Degrees
of freedom intra-correlation Log-likelihood
4 4.8 847.71
10 5.2 733.35
30 6.1 712.80
50 6.7 710.89
60 6.9 710.73
65 7.0 710.72
70 7.0 710.76
90 7.3 711.11
9.7 726.15
and correlations obtained from default data in the section on rating
philosophies and the impact of the credit cycle (see page 257).
MODEL EXTENSIONS
Relaxing the normal distribution assumption
In all the previous calculations we have assumedthat the risk factors
and, consequently, the ability-to-pay variables followa multivariate
normal distribution. In the following, we relax the normal distribu-
tion assumption to allow for fat tails and skews. More precisely, we
calculate estimates for intra-correlations under three different distri-
bution assumptions for each of the 13 industry cohorts (with seven
rating subcohorts):
(i) the systematic factor follows a t-distribution, the idiosyncratic
factors are normally distributed;
(ii) the systematic factor and the idiosyncratic factors follow a
multivariate t-distribution;
7
(iii) The systematic factor follows a skew-normal distribution, the
idiosyncratic factors are normally distributed.
8
Compared with the Gaussian model, the additional parameter in
each of these models improves the t to the data. Overall, the best
calibration results are obtained in model (ii) or, more precisely, in a
model that follows a multivariate t-distribution with a high degree
of freedom. Table 10.5 illustrates this result for rating time series.
256
The rst column shows the average intra-correlations (R
2
values) for
selecteddegrees of freedom. The last row, labelledby, corresponds
to the normally distributed model. The average intra-correlation of
9.7% is consistent with the average correlation shown in Table 10.4.
Note that intra-correlations decrease for lower degrees of freedom,
which is counterbalanced by higher tail dependencies in these mod-
els. The second column in Table 10.5 displays the average value
of the log-likelihood function at the 13 R
2
values for each of the
selected degrees of freedom. The maximum is obtained for a degree
of freedom that is close to 65. Asimilar result holds for default data.
Further analysis is required to decide whether it is justied to
replace the Gaussian distribution by a more complex distribution
class.
9
Modelling the credit cycle
Rating systems are usually characterised as point-in-time (PIT) or
through-the-cycle (TTC) (Basel Committee on Banking Supervision
2000). In a PIT process, a rating reects an assessment of the bor-
rowers current condition and most likely future condition over the
course of the chosen time horizon, typically one year. As a conse-
quence, the PIT rating changes as the borrowers condition changes
over the course of the credit cycle. Incontrast, a TTCprocess requires
assessment of the borrowers riskiness based on a bottom-of-the-
cycle scenario. In this case, a borrowers rating would tend to stay
the same over the course of the credit cycle. Ratings from exter-
nal rating agencies are typically characterised as TTC. It is obvious,
however, that the S&P rating system is a combination of both rating
philosophies (Cornaglia and Morone 2009). This is highlighted by
the fact that both the proportion of downgrades to upgrades and
the actual default ratios of the different rating classes are strongly
correlated with the credit cycle.
In our analysis we have ignored these characteristics of S&P
data. In fact, we made the assumptions that risk factors do not
exhibit serial dependence over time and default probabilities of rat-
ing classes are constant over time. The application of this oversim-
plifying estimationmodel to rating data withrather complex charac-
teristics might explain the signicant differences between estimates
derived from default data and estimates derived from rating data.
The challenge of having to cope with different rating philoso-
phies has been addressed in Hamerle et al (2003) and Rsch (2005).
257
MODEL RISK
They also use S&P default data, but mimic a PIT rating by includ-
ing information about the credit cycle. As a consequence, correlation
estimates are signicantly reduced.
McNeil andWendin(2006, 2007) propose generalisedlinear mixed
models (GLMMs) as a framework for modelling dependent defaults
and rating migration. The GLMM setting allows for the inclusion of
observable factors (eg, macroeconomic indicators as proxies for the
credit cycle) and serial dependence of latent risk factors. A draw-
back of this technique is that serially correlated latent risk factors
yieldjoint migrationdistributions interms of high-dimensional inte-
grals, which are difcult to calculate with standard ML techniques.
Alternative calibration methodologies are simulated ML (Koopman
et al 2005) and Bayesian techniques. The latter approach is used
by McNeil and Wendin (2006, 2007), who report intra-correlations
(inter-correlations) of 10.9% (6.5%) for S&P default data and 8.7%
(2.6%) for rating data.
CONCLUSION
Let us return to our initial question and compare the correlation
estimates derived from historical default and rating data to correla-
tions of equity time series. Akhaveinet al (2005) apply Fitchs Vector
Model 2.0, which is an equity-based approach, and obtain an aver-
age intra-correlation of 24.1% and an inter-correlation of 20.9% for
the 25 Fitch industry segments. We used a different set of equity
time series covering a 10-year period from 1998 to 2007. Based on
the S&P industry classication we computed rather homogeneous
intra-correlations for the 13industrysegments withanaverage value
of 24.7%. The average inter-correlation is 20.4%. These equity cor-
relations are higher than all the estimates derived from historical
default and rating data that were presented in this chapter. The dif-
ference is particularlypronouncedfor correlations betweenindustry
segments. Hence, we do not see any indication that credit-portfolio
models underestimate joint default probabilities if they are cali-
bratedwithequitydata. Onthecontrary, thereseems tobethedanger
that equity correlations might underestimate diversication effects
between different industries.
The difference between intra-correlations and inter-correlations
alsohas animportant consequence for the designof acredit-portfolio
model. Portfolio models that use a single systematic factor across all
258
industries donot provide sufcient exibilitytocapture the complex
dependence structure exhibitedin rating data. This fact clearly high-
lights an important weakness of the Basel II framework for credit
risk.
Historical ratings are an indispensable data source for calibrating
and validating credit-portfolio models. However, our analysis pro-
vides another conrmation that it is a notoriously difcult task to
calibrate a realistic dependence structure for joint defaults andrating
migrations. The impact of the credit cycle together with the complex
characteristics of the historical rating data make it difcult to specify
an appropriate estimation model. The signicant differences in the
correlation estimates obtained from default and rating data further
highlight this difculty. Another problem is the volatility observed
in correlations: the estimates vary greatly across industries as well as
over time, whichmakes it difcult todevelopreliable estimationpro-
cedures. More work is needed to better understand the dynamics of
correlation changes, particularly under stressed market conditions.
Financial support from the EU Commission through MRTN-CT-
2006-034270 COMISEF is gratefully acknowledged. The views
expressed in this chapter are those of the authors and do not
necessarily reect the position of Deutsche Bank AG.
1 Avector of randomvariables X = (X
1
, . . . , X
n
) is called exchangeable if, for any permutation
: 1, . . . , n 1, . . . , n, the vectors (X
1
, . . . , X
n
) and (X
(1)
, . . . , X
(n)
) have the same
distributions.
2 Rating classes AAA and AA are not displayed since no reliable estimates of their intra-
correlations could be calculated due to the low number of defaults.
3 As mentioned above, the correlation estimates are very sensitive to the precise bucketing of
defaults. For example, the intra-correlation of 9.5% for BB has been obtained as the average
of ML estimates that vary between 7.5% (November 1981 to November 2008) and 11.9%
(September 1981 to September 2008).
4 Acomprehensive theoretical analysis of the different estimators is presented in the PhDthesis
of the second author.
5 Note that l
k,j
can also be negative, ie, l
k,j
represents a gain.
6 See, for example, the correlation estimates obtained with the directional rating transition
matrix in Akhavein et al (2005).
7 Note that in this set-up the idiosyncratic factors are no longer independent (McNeil et al 2005).
8 The skew-normal distribution is an extension of the normal distribution allowing for the
presence of skewness (Arellano-Valle and Azzalini 2006).
9 Hamerle and Rsch (2005) analyse the impact of different distribution assumptions for the
risk factors on the loss distribution of the credit-portfolio model. They come to the conclusion
that even if the correct distribution is specied by a t copula, a misspecied Gaussian model
does not necessarily underestimate risk.
259
MODEL RISK
REFERENCES
Akhavein, J. D., A. E. Kocagil andM. Neugebauer, 2005, AComparative Empirical Study
of Asset Correlations, Quantitative Financial Research Special Report, Fitch Ratings.
Arellano-Valle, R. B., and A. Azzalini, 2006, On the Unication of Families of Skew-
Normal Distributions, Scandinavian Journal of Statistics 33, pp. 56174.
Bahar, R., and K. Nagpal, 2001, Measuring Default Correlation, Risk 14(3), pp. 12932.
Basel Committee on Banking Supervision, 2000, Range of Practice on Banks Internal
Ratings Systems, Bank for International Settlements, January.
Measurement and Capital Standards: ARevised Framework Comprehensive Version,
Bank for International Settlements, June.
Bluhm, C., L. Overbeck and C. Wagner, 2002, An Introduction to Credit Risk Modeling (Boca
Raton, FL: CRC Press/Chapman & Hall).
Cornaglia, A., andM. Morone, 2009, RatingPhilosophyandDynamic Properties of Inter-
nal Rating Systems: A General Framework and an Application to Backtesting, Working
Paper, Intesa Sanpaolo.
Crouhy, M., D. Galai and R. Mark, 2000, AComparative Analysis of Current Credit Risk
Models, Journal of Banking and Finance 24, pp. 59117.
Demey, P., J. F. Jouanin, C. Roget and T. Roncalli, 2004, Maximum Likelihood Estimate
of Default Correlations, Risk 17(11), pp. 1048.
Demey, P., and T. Roncalli, 2004, ACorrection Note on Maximum Likelihood Estimate
of Default Correlations, Working Paper, Credit Agricole.
De Servigny, A., and O. Renault, 2003, Correlation Evidence, Risk 16(7), pp. 904.
Frey, R., and A. J. McNeil, 2001, Modelling Dependent Defaults, Working Paper,
Department of Mathematics, ETH Zrich.
Frey, R., and A. J. McNeil, 2003, Dependence Modeling, Model Risk and Model
Calibration in Models of Portfolio Credit Risk, The Journal of Risk 6(1), pp. 5992.
Gordy, M., and E. Heiteld, 2002, Estimating Default Correlations from Short Panels of
Credit Rating Performance Data, Technical Report, Federal Reserve Board.
Gupton, G., C. Finger and M. Bhatia, 1997, CreditMetrics Technical Document,
JP Morgan, New York.
Hamerle, A., T. Liebig and D. Rsch, 2003, Benchmarking Asset Correlations, Risk
16(11), pp. 7781.
Hamerle, A., and D. Rsch, 2005, Misspecied Copulas in Credit Risk Models: How
Good is Gaussian?, The Journal of Risk 8(1), pp. 4158.
Jobst, N. J., and A. De Servigny, 2005, An Empirical Analysis of Equity Default Swaps
II: Multivariate Insights, Working Paper, S&Ps Risk Solutions.
Koopman, S. J., A. Lucas and R. J. Daniels, 2005, A Non-Gaussian Panel Time
Series Model for Estimating and Decomposing Default Risk, Working Paper, Tinbergen
Institute.
Lucas, D., 1995, Default Correlations and Credit Analysis, Journal of Fixed Income 9(4),
pp. 7687.
McNeil, A. J., R. Frey and P. Embrechts, 2005, Quantitative Risk Management: Concepts,
Techniques, and Tools (Princeton University Press).
260
McNeil, A. J., and J. P. Wendin, 2006, Dependent Credit Migrations, The Journal of Credit
Risk 2(3), pp. 87114.
McNeil, A. J., and J. P. Wendin, 2007, Bayesian Inference for Generalized Linear Mixed
Models of Portfolio Credit Risk, Journal of Empirical Finance 14, pp. 13149.
Merton, R., 1974, On the Pricing of Corporate Debt: The Risk Structure of Interest Rates,
Rsch, D., 2005, An Empirical Comparison of Default Risk Forecasts from Alternative
Credit Rating Philosophies, International Journal of Forecasting 21(1), pp. 3751.
Zeng, B., and J. Zhang, 2002, Measuring Credit Correlations: Equity Correlations Are
Not Enough!, Working Paper, KMV Corporation.
261
11
Asymmetric Asset Correlation:
Some Implications for the
Estimation of Probability of Default
Peter Miu; Bogie Ozdemir
McMaster University; BMO Financial Group
The assumption of the time-invariant asset correlation could be too
restrictiveinexplainingthedynamics of default rates, especiallydur-
ing stressed periods. Empirical studies in international equity mar-
kets document an asymmetric correlation effect in which the return
correlation is signicantly higher during a downturn than during
a booming state of the economy. The default correlation implicit in
the market prices of tranched index products also tends to increase
during a market downturn. In this chapter, we consider a single-
factor credit risk model which exhibits such an asymmetric correla-
tioneffect. Througha number of simulationexercises, we thenassess
and compare the performances of two different estimators of uncon-
ditional probability of default under such an asymmetric correlation
effect: the simple time-series average estimator and the estimator
under a constant asset correlation. In a number of small sample set-
tings, the latter is found to be more efcient than the former, result-
ing in smaller root-mean-square errors. Finally, we document the
difculty of estimating the downturn correlation with an estimator
derived from a constant asset correlation, which tends to under-
state the true value of asset correlation. Economic capital require-
ment may therefore be underestimated and thus capital adequacy
overstated.
INTRODUCTION
Default correlation is one of the most important risk parameters in
governing the risk of a portfolio of credit instruments (eg, in terms of
263
MODEL RISK
value-at-risk (VaR)). Tarashev and Zhu (2007) demonstrate that the
misspecication and incorrect calibration of cross-sectional correla-
tions could result in signicant inaccuracies in assessing portfolio
credit risk. In the structural model of Merton, the default correlation
is a function of the cross-sectional correlation of the returns of the
underlying asset values of the obligors. Asset correlations (and thus
default correlations) could be estimated by observing the realised
default rates (and/or migration rates) over time (Gordy and Heit-
eld 2002; Demey et al 2004). As demonstrated by Castro (2009), the
values of these implied asset correlations are, however, sensitive to
the specication (eg, number of systematic factors) of the model.
Under a typical length of the time span of historical data, Gordy
and Heiteld show that there could be a substantial downward
bias in the correlation estimators resulting in an understatement of
value-at-risk.
In the internal-rating-based (IRB) approach of Basel II, asset cor-
relation is considered to be a function of the probability of default of
the obligors, while disregarding the possibility that it might be gov-
erned by other obligor-specic factors (eg, the size of the company,
thecharacteristic of theindustrysectors towhichtheobligors belong,
etc). McNeil and Wendin (2006) and Castro (2009) consider credit
risk models which cater for the heterogeneity of correlation among
obligors. Except for in structure nance, the time-series properties
of default correlation and their implications are, however, less stud-
ied in the literature. Given the fact that constant correlation models
do not t the tranche prices of structured products well, a number
of studies (eg, Burtschell et al 2005, 2007; Turc et al 2005; Andersen
and Sidenius 2005) examine the implications of using time-varying
correlation models.
1
In this chapter, we study the implications of the asymmetric asset
correlation in the management of credit risks. The asymmetric cor-
relation behaviour among international equity returns has been well
documented in, for example, Ang and Bekaert (2002), King et al
(1994) and Longin and Solnik (1995, 2001). This is the phenomenon
where the correlationof equity returns is higher during bear markets
(bad states of the economy) than during bull markets (good states
of the economy). Essentially, the diversication benet disappears
when we need it the most. Ignoring this effect in portfolio manage-
ment therefore leads to an understatement of the critical loss value
264
ASYMMETRIC ASSET CORRELATION
(eg, VaR). Given the fact that equity return is closely related to asset
return underlying the credit risk of an obligor, it is not surprising
that a couple of studies in the credit markets also document the
same asymmetric correlation behaviour. For example, the empirical
study conducted by Gasha and Morales (2004), based on data from
a group of Latin American countries, suggests that the growth rate
of GDP affects non-performing loans only when it is below a cer-
tain threshold. The research of Marcucci and Quagliariello (2009),
using the historical default rates of borrowers of a sample of Italian
banks, shows that credit risks are more sensitive to (ie, more cor-
related with) the business cycle during downturns. Moreover, the
default correlation implied by tranched index products (eg, iTraxx)
is typically higher during market downturns.
The focus of this chapter is on the effect of the asymmetric asset
correlation on the measurement of the long-run (ie, unconditional)
probability of default (LRPD). The determination of LRPD is an
important building block in the measurement of capital require-
ment. In the IRB approach of Basel II, it is one of the inputs in the
computations of the amount of risk-weighted assets, which in turn
determine the amount of required regulatory capital at the portfolio
level. LRPD can be estimated jointly with the implicit asset correla-
tion by observing historical default rates from a uniform portfolio.
Gordy and Heiteld (2002) consider different estimators and their
performances over various time spans. Their results suggest that
all of their estimators generally produce very accurate estimates of
(unconditional) probability of defaults even when the sample size is
relatively small (ie, time span is relatively short). Their conclusions
are, however, drawn based on the assumption that asset correlations
are time invariant. In this chapter, we want to examine the perfor-
mances of a couple of LRPD estimators based on stylised credit risk
models of stochastic asset correlations. Througha number of simula-
tion exercises, we wouldlike to measure andcompare the accuracies
of these estimators given the existence of the asymmetric correlation
behaviour.
Burd (2009) considers an asymmetric correlation model under the
asymptotic single-risk factor framework of Basel II Pillar I. LRPD is
assumed to be a known constant in the demonstration of the impli-
cations of the asymmetric correlation effect on the estimated VaRs
and expected shortfalls of credit portfolios. In this chapter, we are,
265
MODEL RISK
however, interestedinexamining the potential bias inthe estimation
of LRPD that might result from the asymmetric correlation effect.
The rest of this chapter is organised as follows. In the rst section,
we review the asymptotic single-risk factor model of Basel II Pillar I
and some of its characteristics. Then, we introduce an asymmetric
correlation model. Simulation exercises are then conducted in the
third section in order to gauge the impact of the asymmetric corre-
lation effect on the estimation of LRPD. Finally, we conclude with a
few remarks.
OVERVIEW OFTHE ASYMPTOTIC SINGLE-RISK FACTOR
MODEL AND LRPD
We start with a single-factor model of the variations of asset val-
ues under Mertons structural model. This is the model used by
Vasicek (1987) in generating the loss distribution of a credit port-
folio. Gordy (2003) derives the conditions that ensure the portfolio-
invariant propertyof the resultingVaR, whichunderpins the validity
of the Basel II risk-weight function.
Suppose obligors are uniform in terms of their credit risks within
a certain segment of the portfolio. Under a single-factor model, the
return on assets of an individual obligor a
t
at time t is related to both
a systematic risk S
t
and the obligor-specic risk e
t
. For example, for
obligor i
a
i
t
= S
t
+
_
1
2
e
i
t
(11.1)
We assume that a
t
, S
t
and e
t
follow the standard normal distribu-
tion, where S
t
ande
t
are independent. The obligor defaults whenits a
t
becomes less than a certain default point c. The unconditional prob-
ability of default, which is equivalent to LRPDunder the framework
of Basel II Pillar Is risk-weight function (see, for example, Miu and
Ozdemir 2008), is therefore equal to Pr[a
i
t
< c]. Since a
t
is assumed
to be normally distributed with mean zero and unit variance, the
unconditional probabilityof default is simply(c), where () is the
cumulative standard normal distribution function. In Equation 11.1,
the coefcient measures the sensitivity of an individual obligors
asset return to the systematic risk. The parameter
2
is therefore the
pairwise correlation in asset returns among obligors as a result of
the systematic risk factor.
It can be shown (Vasicek 1987) that the probability of default
of obligor i conditional on observing the systematic risk S
t
can be
266
expressed as
Pr[a
i
t
< c S
t
] = (z(S
t
, , c)) (11.2)
where
z(S
t
, , c) =
1
_
1
2
(c S
t
)
Since e
i
t
is assumedto be independent of e
j
t
for i j, the probability
of observing k
t
defaults out of n
t
initial number of obligors within a
uniform portfolio given the realisation of S
t
is equal to
(k
t
, n
t
; S
t
, , c) =
_
n
t
k
t
_
((z(S
t
, , c))
k
t
(1 (z(S
t
, , c)))
n
t
k
t
)
(11.3)
and thus the unconditional cumulative probability that the number
of defaults does not exceed k
t
can be obtained by integrating over S
t
F
n
t
_
k
t
n
t
_
=
k
t
_
w=0
_
(w, n
t
; S
t
, , c) dS
t
(11.4)
Let us use
t
to denote the observed default rate at time t, that
is
t
= k
t
/n
t
. As n
t
approaches innity (ie, in a portfolio of innite
granularity), it can be shown that F
(
t
) and its density function
f
(
t
) can be respectively expressed as
2
F
(
t
) =
_
1
_
_
1
2
1
(
t
) c
_
_
(11.5)
where
1
() is the inverse of the cumulative standard normal
distribution function. Equation 11.5 shows that
1
(
t
) is in fact
normally distributed with mean and standard deviation equal to
c/
_
1
2
and /
_
1
2
, respectively. We can therefore estimate
c and by computing the mean and variance of the time series of
1
(
t
) observedfromt = 1 to T. Specically, basedonthe unbiased
estimator of mean and variance, we have
E(
1
(
t
)) =
T
t=1
(
1
(
t
))
T
=
c
_
1
2
(11.6)
var(
1
(
t
)) =
T
t=1
(
1
(
t
) E(
1
(
t
)))
2
T 1
=

2
1
2
(11.7)
By solving Equations 11.6 and 11.7, the asymptotically unbiased
estimators of c and are therefore equal to
c =
E(
1
(
t
))
_
1 +var(
1
(
t
))
(11.8)
267
MODEL RISK
and
=
_
var(
1
(
t
))
_
1 +var(
1
(
t
))
(11.9)
From Equation 11.8, we obtain an asymptotically unbiased esti-
mator of LRPD
LRPD =
_
E(
1
(
t
))
_
1 +var(
1
(
t
))
_
(11.10)
As demonstrated by Miu and Ozdemir (2008), this LRPD esti-
mator outperforms a couple of alternative estimators (including the
simple average of default rates) in a number of small sample settings
considered in their study. We will examine the performance of this
LRPD estimator later (see page 269) under asymmetric correlation
effects.
ASYMMETRIC CORRELATION MODELS OF CREDIT RISK
In this section, we examine a possible formulation of asymmet-
ric correlation models in extending the constant correlation single-
factor model considered in the previous section. Let us consider a
mixed model with two regimes of high versus lowasset correlation.
Specically, the return on asset value of obligor i is given by
a
i
t
=
_
_
_
H
S
t
+
_
1 (
H
)
2
e
i
t
if S
t
<
L
S
t
+
_
1 (
L
)
2
e
i
t
if S
t

(11.11)
where
H
>
L
ensures the asymmetric correlation effect, in which
the asset return correlation is higher during the downturn of the
business cycle. Systematic and obligor-specic risk factors S
t
and
e
t
are again assumed to be independent and follow the standard
normal distribution. The threshold denes the high versus low
correlation regimes.
3
Again, an obligor defaults when its a
t
becomes
less than a certain default point c.
In this set-up, rather than following the standard normal distri-
bution of Equation 11.5, the distribution of the inverse cumulative
normal functionof default rate
1
(
t
) is composedof the following
two truncated normal distributions
f (
1
(
t
))
_
_
c
_
1 (
H
)
2
,

H
_
1 (
H
)
2
_
if S
t
<
_
c
_
1 (
L
)
2
,

L
_
1 (
L
)
2
_
if S
t

(11.12)
where () is the normal distribution function.
268
ASYMMETRIC CORRELATION EFFECT ANDTHE ESTIMATION
OF LRPD
In this section, we conduct a number of simulation exercises in order
to compare the performances of two estimators of LRPD under the
asymmetric correlation effect. The rst estimator (LRPD
ave
) is the
simple average of time-series default rates, which is still quite com-
monly used in practice. The second estimator (LRPD
cc
) is the LRPD
estimator derived fromEquation 11.8 under the constant correlation
single-factor model. Specically
LRPD
ave
=
T
t=1
t
T
(11.13)
and
LRPD
cc
= ( c)
=
_
T
t=1
(
1
(
t
))
T
__
1 +
_
1
T 1
T
_
t=1
_
1
(
t
)
1
T
T
_
t=1
1
(
t
)
_
2
__
1/2
_
1
_
(11.14)
Even though both estimators are asymptotically unbiased under
the asymptotic single-risk factor model with constant asset return
correlation, the simulationexercises conductedbyMiuandOzdemir
(2008) show that LRPD
cc
outperforms LRPD
ave
, resulting in a more
accurate estimate of LRPD in a number of small sample settings.
In the study, we want to assess their performances under the
asymmetric correlation model examined in the previous section.
We consider the 12 models with the governing parameters given
in Table 11.1. Models 14 represent obligors of relatively high credit
quality, whereas models 912 are of relatively low credit quality.
Obligors in models 58 are of intermediate credit quality. For each
of these three groups of models, we consider two different lengths of
the time series of observations, namely 10 and 25 periods. From our
experience, default rate data internal to nancial institutions rarely
date back more than 10 years, whereas external data, for example,
compiled by the rating agencies, could date back to the early 1980s.
The models are all asymmetric correlation models, in which the
asset return correlation becomes higher when the systematic factor
269
MODEL RISK
Table 11.1 True parameters of the asymmetric correlation model
True
Model (c) LRPD
no. (%) (
H
)
222
(
L
)
222
T (%)
1 0.50 0.30 0.20 10 0.0 0.52
2 0.50 0.35 0.15 10 0.0 0.54
3 0.50 0.30 0.20 25 0.0 0.52
4 0.50 0.35 0.15 25 0.0 0.54
5 1.00 0.30 0.20 10 0.0 1.05
6 1.00 0.35 0.15 10 0.0 1.10
7 1.00 0.30 0.20 25 0.0 1.05
8 1.00 0.35 0.15 25 0.0 1.10
9 2.00 0.30 0.20 10 0.0 2.11
10 2.00 0.35 0.15 10 0.0 2.23
11 2.00 0.30 0.20 25 0.0 2.11
12 2.00 0.35 0.15 25 0.0 2.23
LRPDs (ie, unconditional probabilities of default) are numerically evalu-

ated via simulations.
becomes lower than certain threshold value . The asymmetric cor-
relation effects of models 1, 3, 5, 7, 9 and 11 are weaker than those of
models 2, 4, 6, 8, 10 and 12. For all models, the value of is assumed
to be equal to 0. In the last column of Table 11.1, we also present
the corresponding true values of the unconditional probability of
default (ie, LRPD). They are numerically evaluated via simulations.
As expected, the larger the asymmetric correlation effect, the higher
the unconditional probability of default.
For eachof these models, we simulate 50,000 time-series of default
rates according to the distribution of Equation 11.12 by assuming the
corresponding parametric values listed in Table 11.1 are in fact the
true parametric values of the processes. We thencompute LRPD
ave
andLRPD
cc
basedonEquations 11.13 and11.14 for eachof the 50,000
simulations. Finally, we compute the following statistics for each
estimator across the 50,000 simulations:
the mean value of the estimator;
the median value of the estimator;
the standard deviation of the estimator;
the 97.5 and 2.5 percentiles of the estimator;
270
the root-mean-square error (RMSE) in estimating the true
LRPD.
The results are tabulated in Table 11.2.
Let us start with the simulation results of model 1 in Table 11.2.
Even though LRPD
cc
suffers from a misspecication error under
the asymmetric correlation model considered here, it is still more
efcient than LRPD
ave
in estimating LRPD, resulting in a smaller
RMSE and smaller dispersion (ie, the difference between 97.5 and
2.5 percentiles) around the true LRPD. Moreover, given the fact that
LRPD
ave
has a higher 97.5 percentile than that of LRPD
cc
, while both
have very similar 2.5 percentile, the inefciency of LRPD
ave
can be
attributed to the chance of signicantly overestimating LRPD.
Comparing the results of models 1 and 2, we notice that the ef-
ciency of LRPD
ave
reduces (larger RMSE) when the asymmetric cor-
relationeffect increases, whereas that of LRPD
cc
is more or less main-
tained. The same conclusions could be drawn when the length of the
observed time-series of default rates increases from 10 to 25 periods
(eg, comparing models 1 and 3 or models 2 and 4). As expected,
RMSE becomes smaller when we have longer data series. Judging
from the median values of the estimators, the chance of LRPD
ave
understating the true LRPD is smaller than that of LRPD
cc
when
T = 25, whereas the chances of understatement are quite similar
when T = 10.
Similar conclusions couldbe drawnfor obligors withhigher prob-
abilities of default (eg, models 58 and models 912). Judging from
the relative reduction of RMSE from LRPD
ave
to LRPD
cc
of each
model, the gain in efciency by using LRPD
cc
tends to be smaller
for lower credit quality obligors. For example, the percentages of
reductions of RMSEs for models 1 and 9 are 7.5% (= 1 0. 37/0. 40)
and 2.7% (= 1 1. 08/1. 11), respectively.
We also compute some statistics for the asset return correlation
estimator of Equation 11.9. In Table 11.3, we give the mean, median,
standard deviation and 97.5 and 2.5 percentiles of the estimator
across the 50,000 simulations conducted for each of the 12 models.
We alsotabulate the proportionof simulations inwhichthe estimator
is higher than (
H
)
2
and lower than (
L
)
2
, respectively.
The results reported in Table 11.3 are consistent with the nd-
ings in Gordy and Heiteld (2002) regarding the downward bias of
the asset correlation estimator. The mean and median of the asset
271
MODEL RISK
Table 11.2 Performance of LRPD
ave
and LRPD
cc
in estimating
unconditional probability of default
Percentiles
True
, ..
Model LRPD Mean Median SD 97.5 2.5 RMSE
no. (%) Estim. (%) (%) (%) (%) (%) (%)
1 0.52 LRPD
ave
0.52 0.41 0.40 1.60 0.12 0.40
LRPD
cc
0.53 0.43 0.37 1.51 0.13 0.37
2 0.54 LRPD
ave
0.55 0.40 0.46 1.82 0.13 0.46
LRPD
cc
0.51 0.40 0.38 1.55 0.13 0.38
3 0.52 LRPD
ave
0.52 0.47 0.25 1.17 0.20 0.25
LRPD
cc
0.49 0.45 0.21 0.99 0.20 0.21
4 0.54 LRPD
ave
0.55 0.48 0.29 1.31 0.20 0.29
LRPD
cc
0.46 0.42 0.21 0.99 0.20 0.22
5 1.05 LRPD
ave
1.05 0.87 0.67 2.81 0.28 0.67
LRPD
cc
1.06 0.91 0.63 2.73 0.31 0.63
6 1.10 LRPD
ave
1.10 0.89 0.76 3.17 0.31 0.76
LRPD
cc
1.05 0.87 0.67 2.86 0.32 0.67
7 1.05 LRPD
ave
1.05 0.97 0.43 2.09 0.45 0.43
LRPD
cc
1.00 0.94 0.37 1.89 0.47 0.37
8 1.10 LRPD
ave
1.10 1.00 0.48 2.32 0.46 0.48
LRPD
cc
0.99 0.92 0.38 1.93 0.46 0.40
9 2.11 LRPD
ave
2.10 1.86 1.11 4.93 0.69 1.11
LRPD
cc
2.14 1.91 1.08 4.89 0.74 1.08
10 2.23 LRPD
ave
2.22 1.91 1.24 5.47 0.75 1.24
LRPD
cc
2.18 1.90 1.16 5.19 0.78 1.16
11 2.11 LRPD
ave
2.11 2.00 0.70 3.76 1.04 0.70
LRPD
cc
2.06 1.96 0.65 3.57 1.06 0.65
12 2.23 LRPD
ave
2.23 2.09 0.79 4.12 1.08 0.79
LRPD
cc
2.09 1.98 0.69 3.73 1.08 0.70
SD, standard deviation.
correlation estimates are close to the respective value of the low-
correlationregime(ie, (
L
)
2
). Thechanceof havinganestimatelower
than (
L
)
2
is much higher than having it higher than (
H
)
2
. The
former could be as high as 56% (for model 1). Based on the 97.5
percentile values, the chances of obtaining estimates higher than or
equal to (
H
)
2
are lower than 2.5% for ve out of these 12 models.
The possibility of capturing the high-correlation regime is therefore
quite slim. Finally, as expected, the dispersions (ie, standard devi-
ation and the difference between 97.5 and 2.5 percentiles) of the
272
Table 11.3 Summary statistics of the asset return correlation estimator
Percentile
Model
, ..
no. (
H
)
222
(
L
)
222
Mean Median SD 97.5 2.5 PEH PEL
1 0.30 0.20 0.19 0.19 0.08 0.37 0.06 0.11 0.56
2 0.35 0.15 0.16 0.15 0.08 0.34 0.03 0.02 0.52
3 0.30 0.20 0.20 0.20 0.05 0.31 0.10 0.03 0.53
4 0.35 0.15 0.16 0.16 0.05 0.28 0.07 0.00 0.45
5 0.30 0.20 0.20 0.19 0.08 0.37 0.06 0.12 0.54
6 0.35 0.15 0.16 0.15 0.08 0.35 0.03 0.02 0.49
7 0.30 0.20 0.20 0.20 0.05 0.31 0.11 0.04 0.49
8 0.35 0.15 0.17 0.16 0.05 0.28 0.07 0.00 0.40
9 0.30 0.20 0.20 0.20 0.08 0.38 0.06 0.13 0.52
10 0.35 0.15 0.17 0.16 0.08 0.36 0.04 0.03 0.45
11 0.30 0.20 0.21 0.21 0.05 0.32 0.11 0.05 0.46
12 0.35 0.15 0.18 0.17 0.05 0.29 0.08 0.00 0.34
SD, standard deviation; PEH, proportion of estimator greater than (
H
)
2
;
PEL, proportion of estimator less than (
L
)
2
.
estimates are smaller for those models with T = 25 than for those
with T = 10.
This downward bias of the asset correlation estimator has impor-
tant practical implications. Manynancial institutions estimate asset
correlations from historical default rates for their retail and SME
portfolios and use them in their economic capital (EC) estima-
tions. As EC is a very important input in a nancial institutions
Internal Capital Adequacy Assessment Process (ICAAP), potential
underestimation of the correlations, and thus EC, may result in the
mis-statement of its capital adequacy.
CONCLUSIONS
Inthis chapter, we have examinedthe implications of the asymmetric
correlation effect on the estimation of the unconditional probability
of default. In simulating such an effect, we considered an extension
of the single-factor credit risk model by incorporating two distinct
correlationregimes, whichare contingent onthe value of the system-
atic factor. Specically, when the value of the systematic factor is low
(ie, probability of default is high), a higher asset return correlation
is realised, and vice versa.
273
MODEL RISK
By conducting a number of simulation exercises, we compared
the performances of two estimators of the unconditional probability
of default given the existence of the asymmetric correlation effect.
The two estimators were the simple time-series average estimator
and the estimator under constant asset correlation. In a number of
small sample settings, the latter was found to be more efcient than
the former, resultinginsmaller root-mean-square errors. The relative
inefciency of the former could be attributed to its higher chance of
overstating the true unconditional probability of default.
Finally, again given the existence of the asymmetric correlation
effect, we examined the performance of an estimator of asset return
correlation derived under the assumption of constant asset correla-
tion. Our simulation results show that it is difcult to estimate the
correlation of the high-correlation regime given that the estimator
tends to understate the true value of asset correlation.
Opinions expressed in this chapter are those of the authors and are
not necessarily endorsed by the authors employers.
1 Under a constant correlation model, the implied correlations for different tranches exhibit a
correlation smile.
2 Please refer to Vasicek (1987) for details.
3 Note that the higher the value of S
t
, lower the probability of default.
REFERENCES
Andersen, L., and J. Sidenius, 2005, Extensions to the Gaussian Copula: Random
Recovery and Random Factor Loadings, The Journal of Credit Risk 1(1).
Ang, A., andG. Bekaert, 2002, International Asset AllocationwithRegime Shifts, Review
of Financial Studies 15, pp. 113787.
Burd, O., 2009, Breaking Correlation Breakdowns: Non-Parametric Estimation of Down-
turn Correlations and Their Application in Credit Risk Models, The Journal of Risk Model
Validation 2(4; Special Issue), pp. 1326.
Burtschell, X., J. GregoryandJ.-P. Laurent, 2005, AComparativeAnalysis of CDOPricing
Models, Working Paper, ISFAActuarial School, University of Lyon and BNP-Paribas.
Burtschell, X., J. GregoryandJ.-P. Laurent, 2007, Beyondthe GaussianCopula: Stochastic
and Local Correlation, The Journal of Credit Risk 3(1), pp. 3162.
Castro, C., 2009, Uncertainty In Asset Correlation For Portfolio Credit Risk: The Short-
comings of the Basel II Framework, Working Paper, ECARES, Universite Libre de
Bruxelles, Belgium.
Demey, P., J.-F. Jouanin, C. Roget and T. Roncalli, 2004, Maximum Likelihood Estimate
of Default Correlations, Risk (November), pp. 1048.
274
Gasha, J. G., and R. A. Morales, 2004, Identifying Threshold Effects in Credit Risk Stress
Testing, Working Paper, IMF.
Gordy, M. B., 2003, A Risk-Factor Model Foundation for Ratings-Based Bank Capital
Rules, Journal of Financial Intermediation 12(3), pp. 199232.
Gordy, M., and E. Heiteld, 2002, Estimating Default Correlations from Short Panels of
Credit Rating Performance Data, Working Paper, Federal Reserve Board.
King, M., E. Sentana and S. Wadhwani, 1994, Volatility and Links between National
Stock Markets, Econometrica 62, pp. 90134.
Constant: 19601990, Journal of International Money and Finance 14, pp. 326.
Longin, F., and B. Solnik, 2001, Extreme Correlation of International Equity Markets,
Marcucci, J., and M. Quagliariello, 2009, Asymmetric Effects of the Business Cycle on
Bank Credit Risk, Journal of Banking & Finance 33, pp. 162435.
McNeil, A., and J. Wendin, 2006, Dependent Credit Migrations, The Journal of Credit
Risk 2(3).
Miu, P., and B. Ozdemir, 2008, Estimating and Validating Long-Run Probability of
Default with Respect to Basel II Requirements, The Journal of Risk Model Validation 2(2),
pp. 139
Tarashev, N., andH. Zhu, 2007, Modeling andCalibrationErrors inMeasures of Portfolio
Credit Risk, Report, Bank of International Settlements.
Turc, J., P. Very and D. Benhamou, 2005, Pricing CDOs with a Smile, Report, SG Credit
Research.
Vasicek, O., 1987, Probability of Loss on Loan Portfolio, Working Paper, KMV
Corporation.
275
12
A Latent Variable Approach to
Validate Credit Rating Systems
Kurt Hornik, Rainer Jankowitsch,
Christoph Leitner, Stefan Pichler;
Manuel Lingo, Gerhard Winkler
Wirtschaftsuniversitt Wien; Oesterreichische Nationalbank
The validation of credit rating systems became an important eld
of research over the past few years (see, for example, Krahnen and
Weber (2001) and Crouhy et al (2001) for a discussion of the impor-
tant properties of credit rating systems). Recent advances in credit
risk management and banking regulation called forth the need for
accurate estimates of model parameters. According to the regulatory
Basel II framework, which has been implemented in many national
legislations, banks have several incentives to make use of internal
rating systems to estimate risk parameters which are the essen-
tial input to calculate their regulatory capital requirements (Bank
for International Settlements 2004). One of the most important risk
parameters is the probability of default (PD), which is dened to
measure the likelihood of the occurrence of a default event for a
certain obligor over a one-year horizon. Modern credit risk man-
agement is crucially based on the risk-adjusted pricing of loans and
other credit-risk contingent claims, which again heavily relies on a
valid and accurate PD estimation methodology. The accuracy of PD
estimates is of particular importance for virtually all pricing models
for structured credit derivatives. While some pricing models need
accurate measures of the average PD of a bond or loan portfolio
as an input, some more advanced models require the distribution
of individual PDs of such a portfolio, which imposes even greater
challenges for the validity of the estimation models.
In this chapter we introduce a new framework to assess the accu-
racy of PD estimates. In contrast to backtesting methods, where ex
277
MODEL RISK
ante PD estimates are compared with ex post realisations of actual
defaults, our model is based on the existence of contemporane-
ous PD estimates for the same obligor provided by different rat-
ing sources. Such a benchmarking approach is particularly helpful
when sufcient default observations are not available or competing
rating systems are to be compared(for a discussionof benchmarking
approaches see Bank for International Settlements (2005) or Hornik
et al (2007)).
Traditional methods of validating credit rating systems focus pri-
marily on their discriminatory power, ie, the ability to ex ante distin-
guish between defaulting and non-defaulting obligors. Among the
best-knownmethods of this type are analyses basedonthe accuracy
ratio and the receiver operating characteristic (see, for example,
Bank for International Settlements 2005). Unfortunately, these meth-
ods do not provide any conclusive information about the accuracy
of the PD estimates. Consider a hypothetical rating system which
systematically underestimates the true PDs by one half. Obviously,
such a rating system, though remarkably inaccurate, has maximal
theoretical discriminatory power. Onthe other hand, ina population
of obligors with identical PDs, even a perfectly accurate PD estima-
tion will show zero discriminatory power. Recent studies also deal
with the shortcomings of the use of these concepts for rating vali-
dation (see, for example, Lingo and Winkler 2008). As pointed out
by the regulatory bodies (Bank for International Settlements 2005),
validation methods have to aim at directly assessing the calibra-
tion quality, ie, the accuracy and reliability of PD estimates (see, for
example, Stein 2002).
In our model the true PD is taken as a latent variable. Rating
outcomes from different sources (eg, banks, rating tools or rating
agencies) are treated as noisy observations of this latent variable. A
parametric specicationof sucha model has to include the following
components:
the distribution of the latent PDs;
the formal relation by which the noise or error terms are linked
to the latent PDs;
the distribution of the error terms.
In general, provided with sufciently many observations of PDesti-
mates for a set of obligors stemming from different sources, such a
278
A LATENT VARIABLE APPROACHTOVALIDATE CREDIT RATING SYSTEMS
model can be estimated via maximum likelihood. The key assump-
tion of our model is that the error terms are related additively to
the latent PDs transformed from the unit interval to the real axis via
a suitable link function (eg, the probit function which is discussed
in this chapter and used in the empirical example). The means and
(co)variances of these error distributions are the key outcome of the
model.
The meanparameters indicate the ratingbias, ie, the expectedshift
of PD estimates of a certain rating system compared to the average
PDacross all rating sources. The variance parameters reect the gen-
eral size of undirected estimation errors, ie, they reect the precision
of the rating system. Finally, the covariances convey information
about potential error dependencies across rating systems.
The proposed model can also be used to compute consensus PDs
for all obligors, which, particularly in the case of only partially avail-
able rating information (ie, not all obligors rating observations by
eachrater or ratingsource are available), is non-trivial (see, for exam-
ple, Cook et al 1986, 2007). To the best of our knowledge there is no
viable methodology available to obtain consensus ratings for these
kinds of data sets. The estimation of consensus PDs is thus one of the
major contributions of this chapter. Deviations betweenobservedPD
estimates and consensus PDs can be analysed on an atomistic case-
by-case basis. Furthermore, these differences can be aggregated on
rating source or obligor levels, which can aid in detecting potential
systematic patterns or anomalies in rating behaviour.
Our framework is applicable in the context of banking supervi-
sion and development of rating models. Banking supervisors are
interested in the parameters of the error terms in order to assess
the calibration quality of the internal rating system of a supervised
bank. Supervisors might also be interested in the consensus PDesti-
mates for analysing nancial stability of a banking system (Elsinger
et al 2006). Finally, developers of rating systems such as banks and
rating agencies have a natural interest in comparing their outcomes
to peers at different stages of the development. Note, however, that,
as for any benchmarking method, our model provides information
about the relative rather than the absolute rating errors, since actual
default information is not incorporated.
We estimate the proposed model for a multi-rater panel provided
by the Austrian central bank, where we observe PD estimates for
279
MODEL RISK
2,090 obligors provided by 13 banks observed in September 2007.
We employ standard maximum likelihood techniques to estimate
the parameters of the distributions of the errors and latent PDs. Our
empirical example shows that our framework is suitable to identify
bank-specic regularities with respect to grouping variables, like
industry afliation, legal formand exposure size, and to conduct an
outlier analysis of the estimated rating errors of individual banks.
The model presented in this chapter is related to other studies on
benchmarking credit rating systems (see, for example, Carey 2001;
Hornik et al 2007; Stein 2002). In contrast to these contributions, our
model explicitly proposes a probabilistic framework which reects
the stochastic nature of the true PDs and the rating errors incurred
in the process of PD estimation. This allows us to directly estimate
parameters reecting the calibration quality of rating systems and
to derive consensus PD estimates consistent with the suggested
framework.
This chapter is organised as follows. The next section establishes
and formally describes the proposed latent variable model for PD
estimation. The following section contains a description of the data
set andthe estimationmethodology, anda presentationof the empir-
ical results. The nal section summarises the results and indicates
some potential applications for practitioners and supervisors.
A LATENT VARIABLE MODEL FOR PD ESTIMATION
The basic assumption of our model is that raters cannot observe the
true PD of the obligor. This assumption is justied by the general
informational asymmetry between rm owners and debt holders
which constitutes the cornerstone of modern corporate nance (see,
for example, Berk and DeMarzo 2007; Leland and Pyle 1977). This
asymmetry can be due to limited access to the existing information,
suchas incomplete accountinginformation(Dufe andLando2001),
or delayed observations of the driving risk factors (Guo et al 2009).
Generally, the modeller is entirely free in the functional specica-
tion of the relationship between the estimation error and the latent
PD.
The motivation for the specication used in this chapter builds
on the main concept of structural or rm value models (see, for
example, Lando 2004; Merton 1974). The standard models assume
280
the rm value to be the only driving factor of credit risk and to fol-
low a geometric Brownian motion. As a consequence, an important
stylised property of these models is that the probit of the PD is lin-
ear in the natural logarithm of the rm value. Let V
i
be the log asset
value of rm i and let PD
i
be its probability of default. The basic
model of Merton (1974) can be written as
PD
i
= (DD
i
), DD
i
= a
i
+b
i
V
i
where is the distribution function of the standard normal distri-
bution, a
i
and b
i
are constants independent of V
i
, and DD
i
is the so-
calleddistance todefault of rmi (BharathandShumway2008; Cros-
bie and Bohn 2003). Along the lines of Dufe and Lando (2001), we
assume that the error in the observation of the rm value is normal
and additive to the logarithmof the rmvalue. This assumption can
also be justied by the widespread use of structural models for PD
estimation in the banking industry; these gained additional impor-
tance by the introduction of the Basel II supervisory framework.
The most prominent industry model was developed by Moodys
KMV(Crosbie and Bohn 2003) and is used in several extensions and
modications by many nancial institutions. In this class of model
the distance to default is derived from stock market and accounting
data and is used as the key input to the PDestimation. It thus seems
natural to assume that erroneous observations andincomplete infor-
mation lead to normally distributed errors which are additive to the
distance to default.
The estimate PD
ij
as derived by bank j for the true PD of rm i
is thus of the form
PD
ij
= (DD
i
+
ij
)
where
ij
is the correspondingerror (whichdepends onj inparticular
as raters might have access to different information sets for rm i).
Equivalently
1
(PD
ij
) = DD
i
+
ij
=
1
(PD
i
) +
ij
Asecond important class of industry models estimate PDs based
on a probit (or logit) regression of observed default indicators on a
set of risk characteristics of the rm (see, for example, Blume et al
1998; Nickell et al 2000). Alinear combination of the risk characteris-
tics of the rm, where the resulting regression coefcients are used
as weights, constitutes the rating score of the rm. The PD estimate
281
MODEL RISK
is then obtained by transforming the score to the unit interval by the
corresponding transformation. In the case of the probit transforma-
tion this approach is fully consistent with our framework, because
the error term in the regression is normal and additive to the score.
Consolidating the Merton-type and the regression approaches,
we are led to consider a general class of models relating the raters
estimated (observed) PDs to the (unobservable) true PDs in the form
(PD
ij
) = (PD
i
) +
ij
for a suitable (strictlymonotonicallyincreasing) linkfunction map-
ping the (0, 1) PD scale to (, +). Via , the PDs are mapped
to corresponding scores. On the score scale, the rating errors are
modelled additively. Let S
ij
= (PD
ij
) and S
i
= (PD
i
) denote the
observed and latent scores, respectively. For Merton-type models,
=
1
and S
i
= DD
i
. Writing
ij
and
ij
to denote the mean and
standard deviation of
ij
, respectively, the above latent trait model
can be written as
S
ij
= S
i
+
ij
+
ij
Z
ij
(12.1)
where the standardised rating errors Z
ij
= (
ij
ij
)/
ij
have mean
zero and unit variance. We prefer to think of PD
i
, and hence the
corresponding scores S
i
, as drawn randomly from an underlying
obligor population, and assume that rating errors are independent
of the true PDs and that true PDs and rating errors are indepen-
dent and identically distributed across obligors. (Of course, these
assumptions could be relaxed if more involved probabilistic speci-
cations are desired.) We thus obtain a mixed-effects model (see, for
example, Pinheiro and Bates 2000) for the observed S
ij
, where the
latent true PD scores enter as random effects. We refer to this model
as the latent trait model for the multi-rater panel of PD estimates.
Givenparametric models for the
ij
and
ij
andthedistributions of
true PD scores and standardised rating errors, the latent trait model
can be estimated using, for example, marginal maximumlikelihood
(provided that the marginal distributions of the S
ij
, ie, the convo-
lutions of the true PD score and rating error distributions, can be
computed well enough) or Bayesian techniques.
Avery simple parametric specication of the bias/variance struc-
ture of the rating errors is
ij
=
j
and
ij
=
j
, in which case the
rating errors would be independent of the obligors and their charac-
teristics (in particular, their creditworthiness itself). We suggest the
282
employment of exible models of the form
ij
=
g(i),j
,
ij
=
g(i),j
where g(i) 1, . . . , G is the groupof obligor i, for a suitable group-
ing of obligors relative to which raters exhibit homogeneous rating
error characteristics. Note that we use for the means of the rat-
ing errors and the respective model parameters, with
ij
= E(
ij
)
and
g,j
the parameter for group g and rater j. The notation is
analogous. Such groups can, for example, be dened by industry,
obligor type, size and legal form, or combinations thereof. The
importance of accounting for industry group effects in the analysis
of creditworthiness patterns has been emphasised in Crouhy et al
(2001).
As
E(S
ij
) = E(S
i
) +
g(i),j
conditions relating the rating biases to the mean observed PDscores
are required to ensure identiability. More generally, note that com-
monrandomeffects inthe rating errors cannot be separatedfromthe
true PDscores. Anatural conditionconsistent withthe interpretation
as rating bias relative to an underlying unbiased truth is
_
j
g,j
= 0
ie, that the raters average rating bias within each obligor group is
zero. With this identiability constraint, the marginal group effects
in the means are absorbed into the means of the latent scores, for
which we shall employ the basic model E(S
i
) =
g(i)
consistent with
the identiability constraint. Possible models for the variances of
the PD scores include var(S
i
) =
2
(constant for all obligors) or
var(S
i
) =
2
g(i)
(constant for all obligors in the same group).
With these specications, the consensus score for obligor i is given
by

S
i
, the estimated random effect in the tted mixed-effects model.
Consensus PD estimates are readily obtained by transforming the
consensus scores back to the PD scale using the inverse
1
of the
linkfunction, ie, as
PD
i
=
1
(
S
i
). Finally, residuals arethepart of the
observations unexplained by the model and given by S
ij

S
i

ij
,
the observed scores minus the estimated random and xed effects.
In what follows, we assume that true PD scores and rating errors
have a (multivariate) normal distribution; possible extensions are
283
MODEL RISK
Table 12.1 Descriptive statistics of the characteristics of the rating
information and the 13 Austrian banks in the data set
Min. Median Max. Mean
Number of obligors per bank 70 182 1,700 420
Number of ratings per obligor 2 2 11 2.5
Size of banks measured by 1.0 8.7 128.5 11.8
their total assets (in bn)
discussed later (see page 293). With these assumptions, the latent
trait model can be estimated using standard software for mixed-
effects models, provided that these allow for sufciently exible
specications of the error covariance structure.
EMPIRICAL ANALYSIS
Data description
For the empirical example we employ a data set on rating infor-
mation provided by Oesterreichische Nationalbank, the Austrian
central bank. The data contain rating information (one-year PDesti-
mates) from13 major Austrian banks on 2,090 obligors in September
2007 and cover a signicant share of the Austrian credit market. For
each obligor, at least two PD estimates are available. The number of
co-ratings (occurrences of ratings of a single obligor by two differ-
ent banks) is 5,460. In addition to the PD estimates, we have cross-
sectional information about the obligors, like legal form, industry
afliation and outstanding exposure. Table 12.1 reports descriptive
statistics of the data set.
Note that even the smallest bank has at least 70 obligors in com-
mon with one or more of the other institutions. Apart from looking
at the number of co-ratings on a bank level, we also compute the
number of co-ratings on an obligor level. The median number of
these co-ratings is 2, suggesting that most obligors have business
relations to only a small number of banks.
For a deeper analysis we group all obligors by their industry
afliation and their legal form. Based on the NACE codes (Euro-
pean Commission 2008) we classify obligors according to nine main
industries. Table 12.2 shows the distribution of the obligors across
the industries.
284
Table 12.2 Distribution of the co-ratings of the 13 Austrian banks
across industries
No. of No. of
Label Industry co-ratings co-ratings (%)
Manufac Manufacturing 938 17.2
Energy Energy and environment 180 3.3
Constr Construction 184 3.4
Trading Trading 641 11.7
Finance Financial intermediation 1,737 31.8
RealEst Real estate and renting 754 13.8
Public Public sector 344 6.3
Service Service 435 8.0
Private Private individuals 247 4.5
Total 5,460 100.0
Table 12.2 shows that the total numbers of co-ratings (5,460) is not
uniformly distributed across the nine industries, ranging from 180
co-ratings in energy and environment (energy) to 1737 co-ratings in
nancial intermediation (nance). With 13 banks and 9 industries
there are 117 possible sub-portfolios to be analysed. However, in 17
of these there are no observations.
In addition, the obligors can be grouped with aspect to their legal
form: 79.6% of the obligors are limited companies, 12.2% unlimited
companies and 8.2% are private individuals.
Finally, we use information on the banks relative exposures
against each obligor. Relative exposure is measured as the outstand-
ing amount against the obligor expressed as a fraction of the total
volume of outstanding loans of a specic bank and serves as a rough
indicator for the size of the obligor.
Results
This section describes the model-selection process and the empirical
results. The general model class allows for many competing model
specications based on the data structure (see page 280). Our selec-
tion process yields a model using the industry as grouping variable.
We present bank- and industry-specic analyses based on rating
biases and error variances derived from this specication. Further-
more, we analyse the errors from the consensus rating of this model
on the obligor level based on exposure and legal form.
285
MODEL RISK
Table 12.3 Industry-specic means
g
and PD intervals measured in
basis points (10
4
)
Industry
g
(
g
) (
g
) (
g
+)
Manufac 2.542 55.1 17.7 151.1
Energy 2.993 13.8 3.8 44.2
Constr 2.448 71.8 23.8 190.9
Trading 2.375 87.7 29.8 227.5
Finance 3.256 5.6 1.5 19.2
RealEst 2.474 66.8 22.6 174.7
Public 3.330 4.3 1.1 15.1
Service 2.517 59.2 19.8 157.0
Private 2.296 108.4 40.0 267.4
Intervals are obtained by applying the standard normal distribution.
Model selection and parameter estimates
Equation 12.1 describes a very general model class allowing for a
variety of specications for the means and variances of the nor-
mal distributions of the latent PD scores and rating errors. We use
parametric models that group obligors based on the available co-
variates (for the data set at hand, industry afliation, legal formand
exposure). For the error distributions, bank effects as well as group
bank interaction terms are considered. We also investigate models
allowing for general correlation patterns between rating errors. For
eachttedmodel, we compute theAkaike InformationCriterionand
Bayesian Information Criterion (also known as Schwarzs Bayesian
Criterion). The best model is then selected based on these criteria.
For all calculations we use the R system (version 2.8.1) for statis-
tical computing (R Development Core Team 2008). The parameters
of the mixed-effects models are estimated via maximum likelihood
(see, for example, Pinheiro and Bates 2000) using the R package
nlme (Pinheiro et al 2007). The best model found by the model-
selection procedure uses industry afliation as the sole grouping
variable and a single variance parameter PD score, and is given by
S
ij
= S
i
+
g(i),j
+
g(i),j
Z
ij
, S
i
N(
g(i)
,
2
) (12.2)
where
g,j
is the ratingbias tothe meanPDscore of bankj for obligors
inindustryg,
g,j
is thestandarddeviationof theratingerror of bankj
for obligors in industry g and
g
is the mean PD score in industry g.
This model forms the basis for further analysis. Note that the and
286
parameters are inestimable for industrybank combinations with
no observations.
We begin our analysis of the estimation results by showing the
parameters describingthedistributionof thetruelatent scores. These
are the industry-specic means
g
and the standard deviation . For
ease of interpretation we additionally show the images under the
inverse link function of the mean PD scores for each industry and
the respective one standard deviation intervals (see Table 12.3).
We infer from Table 12.3 that on an aggregate level the portfolio
of our sample banks might exhibit important differences in average
credit quality across industries ranging from 4.3 basis points (1bp
corresponds to 10
4
) measured in terms of PDs for public obligors
to 108.4bp for private obligors. Furthermore, a standard deviation
of 0.357 on the score level yields intervals of different width on
the PD scale depending on the industry-specic
g
, eg, for public
obligors we obtainaninterval rangingfrom1.1bpto 15.1bp, whereas
it is much wider among private obligors, spanning from 40.0bp to
267.4bp.
Analysing the rating biases, Table 12.4 shows the bank-specic
bias estimates (
g,j
). Anumber of conclusions canbe drawnfromthis
analysis. First, we might want to interpret the results froman aggre-
gate bank-specic perspective. Evidently, in columns 1 and 3 most
parameter estimates show a negative sign, whereas the opposite
holds for column 13, with all estimates being positive. Banks 1 and 3
thus seem to be too optimistic in their credit assessment, whereas
bank 13 might exhibit an extremely conservative rating behaviour.
We also employ so-calledrelationshipplots to visualise the esti-
mated rating bias and error variance parameters. These plots use
the strucplot framework according to Meyer et al (2006) and the
corresponding strucplot function of the R package vcd (Meyer
et al 2007) to visualise the relationship between measurements of
a quantitative variable (here, estimated model parameters) and the
interaction of two qualitative factors (here, industrybank combi-
nations). Each combination of factor levels is represented in Fig-
ure 12.1 by a rectangular cell shaded by grey values representing
the corresponding measurement values.
Wecanalsousetheindustries tobetter understandthegeneral ten-
dencyof identiedpotential outlier banks. Bank1 rather generally
overestimates credit quality relative to the other banks (particularly
287
M
O
D
E
L
R
I
S
K
Table 12.4 Rating bias
g,j
for bank/industry combinations of the 13 Austrian banks
Bank
, ..
1 2 3 4 5 6 7 8 9 10 11 12 13
Manufac 0.214 NA 0.313 0.002 NA 0.120 0.097 0.053 0.042 0.138 0.011 NA 0.472
Energy 0.301 0.148 0.191 0.166 NA NA 0.168 0.282 0.063 0.113 0.200 NA NA
Constr 0.211 0.026 0.123 0.053 NA NA NA 0.113 0.058 0.012 0.117 NA 0.607
Trading 0.253 NA 0.192 0.071 0.156 0.129 0.122 0.176 0.173 0.082 0.048 0.163 0.403
Finance 0.029 0.068 0.434 0.259 0.267 0.154 0.259 0.259 0.171 0.304 0.259 0.259 0.402
RealEst 0.249 0.145 0.337 0.004 0.234 0.008 0.080 0.094 0.008 0.251 0.013 0.090 0.474
Public 0.084 0.148 0.297 0.219 NA 0.130 0.224 0.216 0.162 0.048 NA 0.236 0.194
Service 0.214 0.129 0.316 0.019 0.390 0.156 0.068 0.221 0.099 0.132 0.023 NA 0.361
Private 0.197 0.085 0.617 0.240 0.032 0.029 0.285 NA 0.195 0.169 0.193 0.202 NA
NA denotes not available: in the case of no observations the bias cannot be estimated.
2
8
8
Figure 12.1 Rating bias for bank/industry combinations
g,j
of the
13 Austrian banks
Bank
I
n
d
u
s
t
r
y
0.75
0.58
0.42
0.25
0.08
0.08
0.25
0.42
0.58
0.75
Private
Service
Public
Real estate
Finance
Trading
Construction
Energy
Manufacturing
13 1 2 3 4 5 6 7 8 9 10 11 12
A dark cell represents a high absolute value, whereas a light cell represents a
very low absolute value. If the underlying value is negative, the border of the cell
is dotted. In the case of no observations the bias cannot be estimated (hence the
missing cells).
Figure 12.2 Standard deviations
g,j
bank/industry combinations of the 13 Austrian banks
0.00
0.15
0.30
0.45
0.60
0.75
Bank
I
n
d
u
s
t
r
y
Private
Service
Public
Real estate
Finance
Trading
Construction
Energy
Manufacturing
13 1 2 3 4 5 6 7 8 9 10 11 12
A dark cell represents a high absolute value, whereas a light cell represents a very
low absolute value. In the case of no observations the standard deviation cannot
be estimated (hence the missing cells).
in the energy, trading and real estate industries). Bank 3 possibly
overestimates the credit quality for private individuals. Conversely,
for bank 13 we note that it potentially acts over-cautiously in the
construction industry.
One of the key strengths of our framework is its ability to estimate
the standarddeviations of the bank-specic errors andthus measure
289
M
O
D
E
L
R
I
S
K
Table 12.5 Standard deviations
g,j
of the rating errors for bank/industry combinations of the 13 Austrian banks
Bank
, ..
1 2 3 4 5 6 7 8 9 10 11 12 13
Manufac 0.447 NA 0.394 0.270 NA 0.407 0.098 0.033 0.140 0.249 0.136 NA 0.383
Energy 0.152 0.095 0.401 0.096 NA NA 0.248 0.355 0.038 0.355 0.029 NA NA
Constr 0.039 0.149 0.306 0.199 NA NA NA 0.252 0.284 0.256 0.155 NA 0.343
Trading 0.206 NA 0.521 0.175 0.411 0.054 0.212 0.109 0.160 0.296 0.221 0.028 0.215
Finance 0.328 0.166 0.412 0.012 0.041 0.446 0.018 0.020 0.337 0.380 0.020 0.026 0.360
RealEst 0.503 0.530 0.462 0.181 0.319 0.245 0.183 0.282 0.344 0.329 0.163 0.282 0.414
Public 0.221 0.204 0.331 0.036 NA 0.271 0.035 0.006 0.014 0.297 NA 0.048 0.242
Service 0.349 0.346 0.397 0.238 0.051 0.247 0.169 0.522 0.210 0.336 0.233 NA 0.149
Private 0.041 0.287 0.160 0.386 0.386 0.259 0.332 NA 0.133 0.428 0.283 0.612 NA
NA denotes not available: in the case of no observations the standard deviation cannot be
estimated.
2
9
0
the precision of the respective rating systems. Table 12.5 contains
the results and Figure 12.2 shows the corresponding relationship
plot. The values range from 0.006 for bank 8 for public obligors to
a maximum of 0.612 observed for bank 12 for private individuals
indicating that the rating tool employed by this bank might be inap-
propriate for this industry. Froma bank-wide perspective Figure 12.2
furthermore suggests that the level of precision is particularly low
for bank 3. We observe the highest standard deviation levels for the
real estate industry, indicating that the banks have particular dif-
culties in accurately assessing the credit quality and suggesting that
informational asymmetries are rather pronounced in this industry.
Consensus Rating and Residual Analysis
The calibration results of the model presented in the previous sec-
tion allows us to estimate a consensus rating for each obligor (see
page 280). The consensus rating itself can be used for many applica-
tions where deviations of individual raters or ratings froman aggre-
gated rating is of interest, eg, in banking supervision. In this section,
the consensus ratings are used to calculate residuals for each rat-
ing, which allows for a deeper economic analysis. The two variables
not employed for grouping in the model specication, ie, legal form
and exposure, are used in this section to illustrate possible further
analysis.
Figure 12.3 presents the residuals for each bank for corporate
obligors with limited and unlimited liability. Based on this repre-
sentation, we analyse the locations and dispersions of the residuals
on the bank level. In general, the medians of the residuals are small
in absolute terms and we nd no structural effect over all banks,
ie, there is no general difference in terms of the median residuals
between limited and unlimited liability obligors. On a bank-specic
level we nd residual medians that differ markedly between the
two legal forms, eg, for banks 2 and 13. Whereas bank 2 assigns
favourable ratings for limited liability obligors, bank 13 assigns
favourable ratings for unlimited liability obligors. We also see no
systematic difference in residual dispersion between limited and
unlimited corporates, but note that individual banks exhibit rather
marked levels of residual dispersion for a specic legal form. Such
results could be particularly interesting for supervisors, as it might
allow identication of problem areas of individual banks.
291
MODEL RISK
Figure 12.3 Residual analysis for all 13 banks across the legal forms:
limited and unlimited companies
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Limited Unlimited
1.0
0.5
0
0.5
1.0
Bank 4
Bank 8
Bank 12
Bank 1
Bank 5
Bank 9
Bank 13
Bank 2
Bank 6
Bank 10
Bank 3
Bank 7
Bank 11
Figure 12.4 Residual analysis for two banks (bank 13 and bank 8)
across the relative exposure
0.00 0.01 0.02 0.03 0.04 0.05 0.06
Bank 13
Relative exposure
R
e
s
i
d
u
a
l
s
0.00 0.01 0.02 0.03 0.04 0.05 0.06
Bank 8
Relative exposure
R
e
s
i
d
u
a
l
s
1.0
0.5
0
0.5
1.0
1.5
1.5
1.0
0.5
0
0.5
1.0
1.5
1.5
292
As a second illustrative example, we analyse the residuals with
respect to the relative exposure size of the obligors. The relative
exposure shows the importance of an obligor for the bank. Thus
we are particularly interested in detecting obligors with large rela-
tive exposures andexcessivelyfavourable ratings. Figure 12.4 shows
residuals against relative exposures for two selected banks (bank 13
and bank 8). Bank 13 rates two obligors (marked in Figure 12.4) with
high relative exposures (more than 2.5%of the total exposure) rather
too favourable relative to the market consensus. For bank 8, how-
ever, we cannot nd comparable outliers. Such an outlier analysis
is very important, as it might help supervisors to identify problem
loans which have a signicant size within a banks credit portfolio.
DISCUSSION
In this chapter we proposed a new probabilistic framework for
credit-ratingmodel validationinamulti-rater set-up, ie, insituations
where PD estimates from different sources for the same obligors are
available. In our model the unobservable true PD of all obligors is
treated as a latent variable and raters obtain only noisy observations
of the latent true PD. In the general framework three ingredients are
to be specied: the distribution of the latent PD, the distribution of
the error terms and a suitable link function which transforms the
PD to the real axis such that the error terms are additive. Building
on the theory of structural credit risk models, we proposed a spec-
ication with normal error terms which used the probit as a link
function. In an empirical example we estimated suitable parameter-
isations of this model and presented results on parameter estimates
and possible economic interpretation. Our framework has a variety
of potential applications. Banking supervisors might be interested
in parameters of the error terms to assess the calibration quality of
bank internal rating systems. Developers of rating systems such as
banks and rating agencies have a natural interest in benchmarking
their rating outcomes to competing models.
Results of benchmarking analyses always need to be assessed rel-
ative to the representativity the available data panel. In our case, it
obviously needs to be ensured that raters in the panel follow mutu-
ally distinct rating procedures. Data in the panel must be measure-
ments of the same underlying entity. In the application framework
293
MODEL RISK
of this chapter, banks must supply ratings for the same underly-
ing notion of obligor creditworthiness, which in our case are one-
year issuer-specic point-in-time probabilities of default. If different
notions are used, eg, when trying to incorporate ordinal ratings of
creditworthiness such as those provided by the major rating agen-
cies, we couldattempt tomapthese totheir one-year PDequivalents.
In this case, rating errors will include the corresponding mapping
errors.
The suggested framework for modelling multi-rater panels of
PD estimates is very general and allows for a variety of possible
enhancements. We have already indicated the possibility of includ-
ing additional terms in the parametric specications of the means
and variances of the PD scores and rating errors, or allowing for
correlations of rating errors across raters (again, note that a com-
mon error factor is indistinguishable from the latent PD score).
In addition, we could aim at employing more exible models for
the distributions of the PD scores or rating errors, eg, via suitable
mixtures of normals. We could also try to model potential censoring
effects mandated by regulatory frameworks, eg, Bank for Interna-
tional Settlements (2004, Paragraph 331) states that the PDfor retail
exposures is the greater of the one-year PDassociated with the inter-
nal borrower grade to which the pool of retail exposures is assigned
or 0.03%, suggesting enhancement of Equation 12.1 along the lines
of S
ij
= max(S
i
+
ij
, c
i
) with known obligor-specic cut-offs c
i
. It
should be noted, however, that in many applications co-rating pat-
terns are rather sparse, limiting the exibility of statistical models
that can be inferred from available data. Finally, we could think of
extending the cross-sectional set-up to a dynamic framework where
the PD estimates are also observed at different points in time, and
hence the latent PDs and the error terms have to be modelled by
suitable stochastic processes. Such a framework would allow fore-
casting of future PDs as well as a lead-lag analysis across differ-
ent raters. We intend to explore these possible enhancements in our
future research.
Any views expressed in this chapter represent those of the authors
and not necessarily those of Oesterreichische Nationalbank.
294
REFERENCES
Bank for International Settlements, 2004, International Convergence of Capital Mea-
surement and Capital Standards: ARevised Framework, Report.
Bank for International Settlements, 2005, Studies on the Validation of Internal Rating
Systems, Revised Report.
Berk J., and P. DeMarzo, 2007, Corporate Finance, Statistics and Computing (Boston, MA:
Pearson International).
Bharath, S. T., and T. Shumway, 2008,Forecasting Default with the Merton Distance-to-
Default Model, Review of Financial Studies 21(3), pp. 133969.
Blume, M., F. LimandA. MacKinlay, 1998, The DecliningCredit Qualityof US Corporate
Debt: Myth or Reality, Journal of Finance 53, pp. 13891413.
Carey, M., 2001,Some Evidence on the Consistency of Banks Internal Credit Ratings,
Technical Report, Federal Reserve Board.
Cook, W. D., M. Kress and L. M. Seiford, 1986, Information and Preference in Partial
Orders: ABimatrix Representation, Psychometrika 51, pp. 197207.
Cook, W. D., B. Golany, M. Penn and T. Raviv, 2007, Creating a Consensus Ranking of
Proposals from Reviewers Partial Ordinal Rankings, Computers & Operations Research,
34, pp. 95465.
Crosbie, P., and J. Bohn, 2003,Modeling Default Risk, Technical report, Moodys KMV,
December.
Crouhy, M., D. Galai and R. Mark, 2001, Prototype Risk-Rating System, Journal of
Banking and Finance 25, pp. 4795.
Dufe D., and D. Lando, 2001, Term Structures of Credit Spreads with Incomplete
Information, Econometrica 69, pp. 63364.
Elsinger, H., A. Lehar and M. Summer, 2006, Risk Assessment for Banking Systems,
Management Science 52(9), pp. 130114.
European Commission, 2008, Statistical Classication of Economic Activities in the
European Community, Report.
Guo, X., R. A. Jarrow and Y. Zeng, 2009, Credit Risk Models with Incomplete
Information, Mathematics of Operations Research 34(2), pp. 32032.
Hornik, K., R. Jankowitsch, M. Lingo, S. Pichler and G. Winkler, 2007, Validation of
Credit Rating Systems Using Multi-Rater Information, The Journal of Credit Risk 3(4),
pp. 329.
Krahnen, J. P., and M. Weber, 2001, Generally Accepted Rating Principles: A Primer,
Journal of Banking and Finance 25, pp. 323.
Lando, D., 2004, Credit Risk Modeling, First Edition (Princeton University Press).
Leland, H. E., and D. H. Pyle, 1977, Informational Asymmetries, Financial Structure, and
Financial Intermediation, Journal of Finance 32, pp. 37187.
Lingo, M., and G. Winkler, 2008, Discriminatory Power: An Obsolete Validation
Criterion?, The Journal of Risk Model Validation 2(1), pp. 127.
Merton, R. C., 1974, On the Pricing of Corporate Debt: The Risk Structure of Interest
Rates, Journal of Finance 29, pp. 44970.
Meyer, D., A. Zeileis and K. Hornik, 2006, The Strucplot Framework: Visualizing Multi-
Way Contingency Tables with vcd, Journal of Statistical Software 17(3), pp. 148.
295
MODEL RISK
Meyer, D., A. Zeileis andK. Hornik, 2007, vcd: Visualizing Categorical Data, Rpackage
version 1.0-6, URL: http://CRAN.R-project.org/package=vcd.
Nickell, P., W. Perraudin and S. Varotto, 2000, Stability of Rating Transitions, Journal of
Pinheiro, J., and D. Bates, 2000, Mixed-Effects Models in S and S-PLUS, Statistics and
Computing (New York: Springer).
Pinheiro, J., D. Bates, S. DebRoyandD. Sarkar, 2007, nlme: Linear andNonlinear Mixed-
Effects Models, R package version 3.1-86, URL: http://CRAN.R-project.org/package
=nlme.
R Development Core Team, 2008, R: ALanguage and Environment for Statistical Computing
(Vienna: R Foundation for Statistical Computing), URL: http://www.R-project.org.
Stein, R., 2002, BenchmarkingDefault PredictionModels: Pitfalls andRemedies inModel
Validation, Technical Report 020305, Moodys KMV.
296
Part IV
Liquidity, Market and
Operational Risk Models
13
Modelling Derivatives Cashows
in Liquidity Risk Models
Stefan Reitz
Hochschule fr Technik Stuttgart
Managing liquidity is the most important activity performed within
banks, since the ability to meet funding requirements and payment
obligations when and where due is essential for the going concern of
any banking organisation. In order to achieve a sound liquidity risk
management, banks, among others, must provide projections and
reporting of all deterministic and stochastic, on- and off-balance-
sheet cashows. The result is a liquidity gap report showing cumu-
lative cashows over time. In this chapter we investigate the estima-
tion of expected cashows resulting fromderivatives positions. The
problem is that those cashows are stochastic with respect to their
amount or timing in many cases. We demonstrate here how pricing
models for derivatives can be used in order to calculate expected
values of future cashows. Various kinds of products and possible
approaches are analysed.
EXPECTED CASHFLOWS FOR NON-PATH-DEPENDENT
DERIVATIVES
The calculation of expected cashows from derivatives can be
done using derivatives-pricing theory. In the rst section, we will
investigate some types of non-path-dependent derivatives.
Notation and basics from derivatives-pricing theory
In arbitrage-free markets the price of a derivative is the expected
value of future cashows, where the expected value is calculated
using an appropriate probability measure, the so-called martingale
measure or risk-free measure. More precisely, for every numraire
(ie, the price process (N
t
) of a traded asset with positive value and
without any coupon or dividend payments) there exists a (unique)
299
MODEL RISK
probability measure, the martingale measure Q
N
, such that the fair
value (price) of a Europeanderivative withpayment Xat future time
T > 0 can be calculated as
V
0
= N
0
E
Q
N
_
X
N
T
_
where E
Q
N
() denotes the expected value operator.
In the case when we are able to write the last expression in the
form
V
0
=
N
0
N
T
E
Q
N
(X) (13.1)
we cangive anexpressionfor the expectedpayment of the derivative
under the measure Q
N
as
E
Q
N
(X) = V
0
N
T
N
0
.
We want to use this concept for calculating expected cashows for
various derivative instruments. As the value V
0
of a derivative con-
tract is provided by the front ofce system within a bank, it is, in
principle, an easy exercise to calculate the above expression.
Which numraire has to be chosen in order make Equation 13.1 a
valid formula?
The most popular numraire is the savings account with deter-
ministic risk-free interest rate r: N
t
= e
rt
. As this expression is
deterministic, we can conclude
V
0
= 1E
Q
N
_
X
e
rT
_
= e
rT
E
Q
N
(X)
Another important numraire is N
t
= B(t, T), ie, the price at time t of
a risk-free zero bondwithface value 1 andexpirydate T. The martin-
galemeasurebelongingtothis numraireis calledthetimeT forward
measure and denoted by Q
T
. It is easy to see that under the mea-
sure Q
T
we can calculate the time T forward value of a traded asset
with price S
t
(without coupon- or dividend-payments) as E
Q
T
(S
T
):
the forward value is dened as the strike-price K, which makes the
price of a T maturing forward-contract with payout S
T
K to zero
0 = N
0
E
Q
T
_
S
T
K
N
T
_
= B(0, T)E
Q
T
_
S
T
K
1
_
K = E
Q
T
(S
T
).
300
MODELLING DERIVATIVES CASHFLOWS IN LIQUIDITY RISK MODELS
How can we calculate E
Q
T
(S
T
)? For this purpose it is important to
understand that in an arbitrage-free market the discounted price
processes of tradeable assets (without coupons or dividends) are
martingales, ie, their expected values as seen fromtime t equal their
value at time t. This implies the relation
E
Q
N
_
S
t
N
t
_
=
S
0
N
0
(13.2)
for every chosen numraire N
t
. If we apply this statement to the time
T forward measure, we can write
E
Q
T
(S
T
) = E
Q
T
_
S
T
B(T, T)
_
=
S
0
B(0, T)
Hence, we have found an expression for the time T forward value
S
fwd
0
:=
S
0
B(0, T)
The time T
1
forward value of a zero bond (with maturity in T > T
1
)
is calculated by setting S
0
:= B(0, T). We then have
B(0, T)
fwd
=
B(0, T)
B(0, T
1
)
If we write B(0, T) = e
L(0,T)T
with the zero rate L(0, T) and
B(0, T)
fwd
= e
L(0,T
1
,T)(TT
1
)
we immediately get the continuously compounded forward rate
L(0, T
1
, T) =
L(0, T)T L(0, T
1
)T
1
T T
1
Animportant property of the time T forwardmeasure is, that evenin
the case of stochastic interest rates, we can separate the expression
1/N
t
= 1/B(t, T) from the expected value operator: the price of a
European derivative with payout X at T under the measure Q
T
is
V
0
= N
0
E
Q
T
_
X
N
T
_
= B(0, T)E
Q
T
_
X
1
_
= B(0, T)E
Q
T
(X)
so we are again in the situation of Equation 13.1, and we conclude
that
E
Q
T
(X) =
V
0
B(0, T)
(13.3)
301
MODEL RISK
Symmetric derivatives
Expectedcashows for theses types of instruments (swaps, forward-
rate agreements, forwards) can be calculated by using forward rates
or forward prices.
As an example we consider a plain vanilla interest rate swap,
where a bank receives annual xed payments and pays an annual
oating amount. The expected cashow at the end of year t
i+1
will
then be modelled as
(C L(0, t
i
, t
i+1
)) NPA
where C is the xed swap rate and NPA is the notional amount
of the swap. We can model the uncertainty of the oating amount
by assuming a lognormal distribution of the future spot rate
L(t
i
, t
i
, t
i+1
), for example, in a Libor market model (LMM)
dL
i
(t) =
i
(t)L
i
(t) dW
i
t
(13.4)
if and only if
L
i
(t) = L
i
(0) exp
__
t
0
i
(u) dW
i
u
1
2
_
t
0
(
i
(u))
2
du
_
(13.5)
with L
i
(t) := L(t, t
i1
, t
i
) the forward Libor rate for [t
i1
; t
i
] as seen
fromt, andwhere (W
i
t
) is a Wiener process under the time t
i
forward
measure Q
t
i
.
To give an example, we examine the payment at the end of year
t
4
= 4. If we assume L
4
(0) = 3%,
4
(t) = 15% for all t < t
3
, under
Q
t
4
we have
ln
_
L
4
(t
4
)
L
4
(0)
_
N
_
1
2
_
3
0
(
4
(u))
2
du,
_
3
0
(
4
(u))
2
du
_
= N(0. 033750. 0675)
For NPA = 100 million and C = 5%, we conclude that the expected
cashow under Q
t
4
is
(0. 05 L
4
(0)) 100,000,000 = 2,000,000
and the cashows have a standard deviation of
exp
__
4
0
(
4
(u))
2
du
_
1 =
_
e
0.0675
1 L
4
(0) 100,000,000
792,762
302
We can quantify the impact of the model risk due to a misspeci-
cation of the volatility
t
i
(t) =
i
(0) (in this case we work with
a constant volatility): the expected cashow is not inuenced by
this quantity, but the standard deviation depends linearly on
t
i
(0),
because
exp
__
t
i
0
(
t
i
(0))
2
du
_
1
_
t
i
(
t
i
(0))
2
=
_
t
i
t
i
(0)
Calculating expected value for all future oating interest pay-
ments requires a formulation of the LMM under a common proba-
bility measure for all maturities. Note that Equations 13.4 and 13.5
are validonlyunder the respective t
i
forwardmeasure Q
t
i
. If we want
to work under a single probability measure Q, we need to express
the dynamics of the Libor rate under this measure Q. We decide
to select Q = Q
t
n
. It is well known that the equations for L
i
(t) then
appear as
dL
i
(t) = L
i
(t)
_
i
(t) dW
i,n
t

i
(t)
n
_
k=i+1
i,k
(t
k
t
k1
)
k
(t)L
k
(t)
1 +(t
k
t
k1
)L
k
(t)
dt
_
Here (W
i,n
t
) is a Wiener process under the measure Q
n
t
and
i,k
is the
correlation of dW
i,n
t
and dW
k,n
t
. The structure of the matrix (
i,k
)
i,k
determines how the random factors dW
i,n
t
in our model are corre-
lated to each other. If we decide to work with a one-factor model
with random factor dW
i,1
t
we can choose
1,k
=
k,1
:= for k > 1,
1,1
:= 1 and
i,k
:= 1 for i, k > 1
An application of the Ito lemma yields the following expression for
L
i
(t) under Q
t
n
L
i
(t) = L
i
(0) exp
__
t
0
i
(u) dW
i
u
1
2
_
t
0
(
i
(u))
2
du
_
exp
__
t
0
i
(u)
n
_
k=i+1
i,k
(t
k
t
k1
)
k
(u)L
k
(u)
1 +(t
k
t
k1
)L
k
(u)
du
_
Obviously in this general case the expected value of the future cash-
ow L(t
i1
, t
i1
, t
i
) under Q
t
n
is not the forward rate L(0, t
i1
, t
i
).
Instead, it is the forward rate multiplied by a factor
(1 +convexity adjustment)
where the convexityadjustment comes fromthe secondfactor exp()
in the above calculation.
303
MODEL RISK
Usually the impact of the convexity adjustment is not large so that
we could argue that
E
Q
tn
L
i
(0)
is a reasonably good approximation for the expected cashow in
this case too. An estimation of the standard deviation of the esti-
mated future cashows can be obtained by performing a Monte
Carlo simulation of the future rates using the above representation
of L
i
(t).
Asymmetric derivatives
Future cashows from option positions are uncertain with respect
to their amount and sometimes the payment time. It is necessary to
consider the kind of settlement agreement for an option: in the case
of physical settlement, the cashows of the resulting position in the
underlying instrument also have to be accounted for in the future
cashow projection, whereas for cash settlement only the intrinsic
option value at expiry is relevant.
As future cashows from option positions are uncertain, we will
calculate exercise probabilities for options. If this probability is
greater then 0.5, we assume that the option will be exercised and the
relevant cashows at expiry will be considered. If the probability is
less than or equal to 0.5, no exercise is assumed.
Exerciseprobabilities for European-styleoptions canbecalculated
in many cases from closed-form option-pricing formulas. A useful
result that can be used in this context is the general option-pricing
formula: in an arbitrage-free, complete market, the pricing formulas
for a European call option and put option with maturity at T, strike
K and a non-dividend-paying underlying S
t
are
c
0
= S
0
Q
S
(S
T
> K) B(0, T)KQ
T
(S
T
> K)
p
0
= B(0, T)KQ
T
(S
T
< K) S
0
Q
S
(S
T
< K)
Here Q
T
denotes the martingale measure with respect to the nu-
mraire N
t
= S
t
and Q
T
is the time T forward measure. It is impor-
tant to notice that no further assumptions are made about how the
price process of the underlying is modelled, ie, this formula is valid
for the BlackScholes model with a geometric Brownian motion for
the underlying as well as other possible price processes including
stochastic interest rates or jumps.
304
The derivation of the above formulas is as follows (we consider
the call price formula)
D
0
= B(0, T)E
Q
T
(maxS
T
K, 0)
= B(0, T)E
Q
T
(1
S
T
>K
(S
T
K))
= B(0, T)E
Q
T
(1
S
T
>K
S
T
) B(0, T)E
Q
T
(1
S
T
>K
K)
= B(0, T)E
Q
T
(1
S
T
>K
S
T
) B(0, T)KQ
T
(S
T
> K)
Now a change of numraire is necessary in order to deal with the
rst expectation value. We will use the numraire N
t
= S
t
and apply
the well-known change-of-numraire theorem
D
0
= B(0, T)E
Q
S
_
1
S
T
>K
S
T
S
0
B(T, T)
B(0, T)S
T
_
B(0, T)KQ
T
(S
T
> K)
= E
Q
S
(1
S
T
>K
S
0
) B(0, T)KQ
T
(S
T
> K)
= S
0
Q
S
(S
T
> K) B(0, T)KQ
T
(S
T
> K)
We can see from the above formula that both terms Q
S
(S
T
> K)
(corresponding to the (d
1
) in the BlackScholes formula) and
Q
T
(S
T
> K) (corresponding to (d
2
)) can be seen as exercise prob-
ability: the rst with respect to the measure Q
S
and the second with
respect to Q
T
.
Within the BlackScholes model we have
Q
S
(S
T
> K) = (d
1
), Q
T
(S
T
> K) = (d
2
) for calls
Q
S
(S
T
< K) = (d
1
), Q
T
(S
T
< K) = (d
2
) for puts
with
d
1,2
=
ln(S
0
/K) +(r :
1
2
2
)T
T
and d
2
= d
1
T
For more complex models, such as stochastic volatility models
(eg, the Heston model), the formulas for the probabilities are much
more complex and need to be approximated by numerical methods.
As the probabilities depend on the implied volatility, model risk
arises here. Figure 13.1 shows the dependence of (d
1
) and (d
2
)
on the parameter and the moneyness S
0
/K for a European call
option.
In the following, we denote the exercise probability by EP. The
expected future cashows will be
cashow from underlying and payment of strike price if EP > 0. 5
intrinsic value if EP 0. 5
305
MODEL RISK
Figure 13.1 Exercise probabilities for a one-year call option
0.1 0.2
0.4 0.5
0.90
0.95
1.00
1.05
1.10
0.2
0.4
0.6
0.8
0.3
d
1,2
S
0
/ K
For foreign exchange options with physical settlement the cashow
from the underlying will be the notional amount in foreign cur-
rency which is bought or sold in the option, whereas for options on
securities the cashow from the underlying is the expected value
of the underlying at expiry (assuming that the underlying will be
sold/bought in the market at expiry).
We need to calculate expected values for the underlying an the
intrinsic value of the optionat expiry. Under the time T forwardmea-
sure Q
T
, the expected value for a non-dividend-paying underlying
can be calculated as
E
Q
T
(S
T
) = E
Q
T
_
S
T
B(T, T)
_
=
S
0
B(0, T)
= S
fwd
0
(13.6)
which follows from Equation 13.2.
The expected intrinsic value for a European option with payout
X is given by Equation 13.3
E
Q
T
(X) = V
0
/B(0, T)
As anexample, we investigate a cap(Europeanstyle), whichconsists
of n caplets paying
X = N(t
i
t
i1
) maxL(t
i1
, t
i1
, t
i
) K, 0
at t
2
< t
3
< < t
n+1
. Here, L(t
i1
, t
i1
, t
i
) denotes the Libor spot
rate for the period [t
i1
; t
i
] xed at t
i
and K is the cap rate. From
Equation 13.5 we knowthat this quantity is lognormally distributed
306
Figure 13.2 Simulated paths of L(t, 5, 6)
0.1
0
0
0.2
0.3
0.4
1 2 3 4 5 6
under the measure Q
t
i
, hence
PV
Caplet
= N(t
i
t
i1
)B(0, t
i
)E
Q
t
i
(maxL(t
i1
, t
i1
, t
i
) K, 0)
= N(t
i
t
i1
)B(0, t
i
)(L(0, t
i1
, t
i
)(
d
1
) K(
d
2
))
where
d
1,2
=
ln(L(0, t
i1
, t
i
)/K) :
1
2
2
t
i1
t
i1
and the parameter denotes the at volatility.
For each caplet we can estimate the level of for which the
exercise probability (
d
2
) is above 0.5
(
d
2
) > 0. 5 ln(L(0, t
i1
, t
i
)/K)
1
2
2
t
i1
> 0
<
2 ln(L(0, t
i1
, t
i
)/K)
t
i1
The expected t
i
-intrinsic value is
N(t
i
t
i1
)(L(0, t
i1
, t
i
)(
d
1
) K(
d
2
))
Until now, expectedfuture cashows andexercise probabilities have
beeninvestigatedbasedontodaysmarket data. Inorder toreceive
a distribution of future cashows (and exercise probabilities) we
could simulate future scenarios for the relevant parameters and
study the cashows under these scenarios.
We proceed with the cap example and perform an LMM-based
simulation for the spot rates L(t
i1
, t
i1
, t
i
) under Q
t
i
, leaving the
volatility parameter unchanged. Figure 13.2 shows some simulated
paths L(t, 5, 6), where the underlying process is dened by
dL(t, 5, 6) = L(t, 5, 6) 0. 15dW
t
, L(0, 5, 6) = 3%
307
MODEL RISK
under the measure Q
6
.
The distribution of L(5, 5, 6) is given by
ln(L(5, 5, 6)) = N(ln(L(0, 5, 6)) 0. 5 5 0. 15
2
, 5 0. 15
2
)
= N(3. 56, 0. 1125)
For a caplet with expiry at t
i
= 5, strike K = 2. 5%, N = 1 million
and expiry at t
i+1
= 6, the exercise probability is
EP = (
d
2
) = 0. 64
Furthermore, the intrinsic value at expiry has an expected value of
1,000,000(0. 03(
d
1
) 0. 025(
d
2
)) = 6. 684
andits standarddeviation can be estimated(basedon 10,000 simula-
tions) to 8,700. Consequently the one sigma range of possible future
cashows is [2,016; 15,384].
EXPECTED CASHFLOWS FOR PATH-DEPENDENT
DERIVATIVES
Path-dependent derivatives are characterised by the fact that future
expected cashows occur before the nal maturity of the derivative
or theynever occur, dependingoncertainevents happeningbetween
the start and end dates of the transaction. Examples of this kind of
derivatives are
barrier options,
American options,
credit derivatives.
Barrier options
For barrier options, the expected future cashows are determined
by both the exercise probability at expiry and the probability of a
knock-in/knock-out event during the lifetime of the option. Here,
we will examine a specic kind of currency barrier option, namely
the Europeandown-and-out call option. The payout at expiry equals
X = maxS
T
K, 01
min
0tT
S
t
H
ie, the payout is identical to a European call option but the option
becomes worthless if, at any time t prior to T, the currency exchange
308
rate S
t
falls belowa predeterminedlevel H. It is evident that the prob-
abilities for future expected cashows differ from the probabilities
in the case of plain vanilla European options.
First we consider the case H < K and H < S
0
. It is clear that the
option will be knocked out if it is out-of-the-money. We write
S
0
= S
0
exp(r
d
r
f
1
2
2
)t +W
t
(using the numraire N

t
= e
r
d
t
), where r
d
is the domestic risk-free
rate and r
f
is the foreign risk-free rate. If Q
N
denotes the martingale
measure, the current price D
0
of the option is
D
0
= e
r
d
T
E
Q
N
(X) = e
r
d
T
E
Q
N
((S
T
K)1
S
T
K, min
0tT
S
t
H
)
= e
r
d
T
(E
Q
N
(S
D
1
E
) K1
E
) (13.7)
where E denotes the event E = S
T
K, min
0tT
S
t
H.
We conclude that for a long position in this option the expected
future cashow, e
r
d
T
D
0
, will be realised if the probability Q
N
(E) =
E
Q
N
(1
E
) is greater than 0.5. Now the probability is known to be
(Musiela and Rutkowski 1998)
Q
N
(E) =
_
ln(S
0
/K) +(r
d
r
f
1
2
2
)T
T
_
_
H
S
0
_
2r
d
/
2
1
_
ln(H
2
/S
0
K) +(r
d
r
f
1
2
2
)T
T
_
In the case when K H < S
0
, Equation 13.7 remains valid, but
in this case we have E := min
0tT
S
t
H, since min
0tT
S
t
H
implies S
T
> K. The probability of E can now be written as
Q
N
(E) =
_
ln(S
0
/H) +(r
d
r
f
1
2
2
)T
T
_
_
H
S
0
_
2r
d
/
2
1
_
ln(S
0
/H) +(r
d
r
f
1
2
2
)T
T
_
Figure 13.3 shows the probability Q
N
(E) in the case when K H <
S
0
, T = 2, r
d
= r
f
= 3% for some values of S
0
/H and .
Remark 13.1. If the barrier option pays a xed rebate amount at the
knock-out time, we have to calculate the expected value of the rst
passage time, ie, the point in time at which the barrier is hit the rst
time. This can be done numerically by using the following density
309
MODEL RISK
Figure 13.3 Probabilities for a positive cashow at expiry
0.5
0
1.0
Q
N
(E)
0.1
0.1
1.1
1.2
1.3
1.4
1.5
S
0
/ H
function for the rst passage time (Musiela and Rutkowski 1998)
f (x) =
1
2x
3
ln
_
S
0
H
_
exp
_
(ln(S
0
/H) +(r
d
r
f
1
2
2
)x)
2
2
2
x
_
American options
Cashows from American options can occur at any time prior to
expirywhenthe optionis exercised. Therefore, Americanoptions are
harder to evaluate thanEuropeanoptions. The possibility of anearly
exercise leads to a free-boundary problem for the optimal exercise
boundary, which divides the region where exercise is optimal from
the region where it is optimal to hold the option. As there are no
closed-form solutions available for calculating the optimal exercise
boundary, we must use numerical solutions and approximations. A
common approach is to investigate the optimal exercise boundary
close to expiry andtry to ndan asymptotic expansion in terms of ,
the time to expiry. Inthe following, let K denote the strike price of the
option, let S denote the value of the underlying and let x := ln(S/K).
For an American call option, which gives the holder the right to
buy the underlying asset, an asymptotic analysis can be done in
order to nd a series expansion for the optimal exercise boundary
x() (exercise at t = T will happen in the case when x > x()).
The necessary calculations for the call are straightforward and can
be found in Alobiadi (2000) and Wilmott (1998). For American put
options it is quitedifcult tondanexpansionfor x(). This problem
310
has been studied by Barles et al (1995), Kuske and Keller (1998) and
Alobiadi and Mallier (2001). Barles et al (1995) were able to nd
upper and lower bounds for the optimal exercise boundary near
to expiration and then prove that these upper and lower bounds
approached each other as expiration was approached, leading them
to conclude that the value of the option behaved like
K
_
1
_
ln()
_
Kuske and Keller (1998) derived an expansion for x() with the
leading-order term on their solution agreeing with that of Barles et
al (1995).
In Alobiadi and Mallier (2001) an asymptotic analysis of the exer-
cise boundary near to expiration is done which produces similar
results to the two studies mentioned.
We consider here the case of an American put option. It is well
known that the value P(S, t) of a put option on a dividend paying
stock obeys the BlackScholes partial differential equation (PDE)
P
t
+
1
2
2
S
2
2
P
S
2
+(r q)
P
S
rP = 0
where r is the risk-free rate, q is the (constant) dividend yield and
is the volatility. The payout at expiry is given by P(S, T) = maxK
S, 0, subject to the boundary conditions that P(0, T) = Ke
r(Tt)
,
P(S, t) 0 as S . For an American put, the possibility of early
exercise leads to the constraint
P(S, t) maxK S, 0, t T
One approach commonly taken with these equations is to apply
certain transformations in order to get a further PDE in terms of the
variable x, from which the optimal exercise boundary x() can be
deduced. The result is
x()
_
2 ln() ln
_

2
6r
2
ln()
+
_
+
In the following example we consider an American put option with
time to maturity = 3 and S
0
= K, r = q, = 0. 4. Figure 13.4 shows
ve simulated paths of the process
_
ln
_
S
K
__
[0;0.3]
311
MODEL RISK
Figure 13.4 Simulated paths of ln(S
/K) and x()

0.05 0.10 0.15 0.20 0.25 0.30
0.8
0.6
0.4
0.2
and the approximated optimal exercise boundary

x()
_
2 ln()
Four of the ve paths end with an exercise of the option.
AMonteCarlosimulationof 10,000paths shows that theestimated
exercise probability is 0.6. If we estimated this probability by (d
1
)
or (d
2
), we would receive the values 0.544 and 0.456, respectively.
Obviously, the exercise probability is underestimated by these val-
ues, as they are related to European options only. It can also be seen
by the simulation that the exercise time is always close to expiry.
In order to estimate the expected future cashow from an Ameri-
canoption, we also have to calculate the expectedexpiry date, which
could be done via the above-mentioned Monte Carlo simulation.
Credit derivatives
As an example for credit-derivatives-related cashows we consider
here credit default swaps (CDSs). CDSs canbe pricedinanintensity-
based model, where the probability of default Q(t) of the reference
asset up to time t is
Q(t) = 1 S(t) = 1 exp
_
_
t
0
(s) ds
_
Here, (s) is the time-dependent intensity rate and S(t) is the sur-
vival probability. In practice, the intensity rate is often chosen to
be piecewise constant, so for a CDS with premium payment dates
t
1
< < t
n
we have
Q(t) = 1 S(t) = 1 exp
_
_
t
i
t
i
_
312
where
i
denotes the number of days in the ith period. The amount
and timing of future cashow depend on possible credit events. For
the protection buyer, the following cases are possible from todays
perspective.
1. S(t) > 0. 5 for all t [0; t
n
]: premium payments sN
i
for all
periods (where N is the notional amount, s is the deal spread).
2. S() 0. 5 for some [t
k
; t
k+1
]: premium payments for
all periods up to [t
k1
; t
k
], accrued premium payment for the
period [t
k
; ] and inow of settlement payment (1 RR)N at
with the estimated recovery rate RR.
In the second case, the accrued premium payment and expected
default time have to be calculated. For the default time we write
0. 5 = S() = 1 exp
_
0
(s) ds
_
0
(s) ds = ln(0. 5)
_
t
i
t
k
i
+( t
k
)
k+1
ln(2)

1
k+1
_
ln(2)
_
t
i
t
k
i
_
+t
k
Now, the accrued payment can be calculated as
sN( t
k
)
The level of uncertainty in the default time can be simulated by
using stochastic intensity rates (s). Of course, there is also a level of
uncertainty in the recovery rate at time of default. Modelling of the
recovery rate can be done by assuming a statistical distribution for
that rate or by using an asset value model for the underlying asset
of the CDS.
It is much more difcult to estimate expected cashows from
structured, portfolio-related credit instruments, such as asset-back-
edsecurities (ABSs), because here we must consider the behaviour of
the underlying pool of assets and the tranche structure of the instru-
ment. In practice, the so-called weighted average lifetime (WAL)
is provided for an ABS structure. The WAL is the time-weighted
mean of future repayments in the pool and can be interpreted as the
expected time until full repayment of the ABS notional amount. It is
earlier by far than the legal nal maturity.
313
MODEL RISK
In general for true sale ABS the actual maturity can only be
estimate since it is determined by a number of factors, such as
prepayments,
reinvestment policy in the pool,
defaults and delinquencies,
performance triggers,
call features.
Due to the diversity and the complexity of the inuencing vari-
ables, only a simulation of future pool cashows can be proved to
be a reliable estimation of expected time and amount of cashows.
Some aspects, like the prepayment behaviour, canbe estimatedmore
easily than others: prepayment rates are derived from the so-called
CPR gure, which is the percentage of the outstanding portfolio
notional that is expected to be prepaid annually.
REFERENCES
Alobiadi, G., 2000, American Options and Their Strategies, PhD Thesis, University of
Western Ontario, Canada.
Alobiadi, G., and R. Mallier, 2001, On the Optimal Exercise Boundary for an American
Option, Journal of Applied Mathematics 1(1), pp. 3945.
Barles, G., J. Burdeau, M. Romano and N. Samsoen, 1995, Critical Stock Price Near
Expiration, Mathematical Finance 5(2), pp. 7795.
Kuske, R. E., and J. B. Keller, 1998, Optimal Exercise Boundary for an American Put
Option, Applied Mathematical Finance 5, pp. 10716.
Musiela, M., and M. Rutkowski, 1998, Martingale Methods in Financial Modelling (Berlin:
Springer).
Wilmott, P., 1998, Derivatives, the Theory and Practice of Financial Engineering (Chichester:
John Wiley & Sons).
314
14
Potential Future Market Risk
Manuela Spangler, Ralf Werner
Deutsche Pfandbriefbank
The 20089 nancial crisis has shown that most market risk models,
even if they deliver sufciently accurate risk gures over short time
horizons, are not able to provide reliable forecasts over longer time
horizons. This is due tothe fact that theyare usuallytailoredfor small
time horizons and do not account for long-term changes in market
environments. Therefore, even for unaltered portfolios, todays risk
gure cannot be used as a meaningful forecast for the portfolio risk
over a longer time horizon. The most important consequences of
this forecast risk are, among others, potential limit breaches and an
economic capital squeeze due to rising market risk, mainly triggered
by increasing volatilities.
As a potential remedywe introduce the concept of potential future
market risk, which is thought to be a supplement to traditional
market risk gures and closely related to the potential future expo-
sure concept in counterparty credit risk. For ease of presentation,
we focus on the special case of a delta-normal value-at-risk (VaR)
model, and thus on potential future value-at-risk (PFVaR), but the
concept can easily be generalised to arbitrary market risk measures.
We will demonstrate how PFVaR can be used for improving limit
and economic capital management.
The rest of this chapter is structured as follows. In the next sec-
tion, the delta-normal value-at-risk approach is briey introduced,
andsimplifyingmodel assumptions andrestrictions are pointedout.
The specic example of a two-factor delta-normal framework is then
used to show how VaR would have performed in the past for a
hypothetical portfolio of oating-rate notes with different maturi-
ties. Model risks of both the rst and second kind are dened, and
model risk of second kind (which covers the risk of changing risk
gures although positions are unchanged) is investigated in further
detail. A decomposition of the VaR evolution during the nancial
315
MODEL RISK
crisis shows the dominating effect of rising volatilities on market
risk gures. Next (see page 327), the concept of PFVaR is dened,
and we show how it can be used in practice for limit and economic
capital management. Furthermore, we give a general Monte Carlo
algorithm for PFVaR simulation and calculate PFVaR gures for the
hypothetical portfolio of oating-rate notes. Finally (see page 335),
we summarise the main results and identify the next steps towards
a real practical implementation of the concept.
THE DELTA-NORMAL VAR APPROACH
The delta-normal VaRapproach, as described, for example, byJorion
(2007), is a standard approach used in the nancial industry to mea-
sure market risk. It is based on the assumption that for a given port-
folio P the change in market value V
P
over a short holding period
h depends linearly on changes of a predened set of risk factors
RF
1
, . . . , RF
m
which in turn are multivariate normally distributed.
Although it is well known that this approach has several shortcom-
ings, anda varietyof extensions have beenproposedinthe literature,
this is still the method of choice for most small and medium-sized
banks.
The generic delta-normal VaR
Let us consider a portfolio P with n instruments, let V
P
denote the
value of the portfolioPat a givenpoint intime, andlet V
i
be the value
of instrument i 1, . . . , n. Then the change in portfolio value over
a xed holding period h is given by
V
P
=
n
_
i=1
V
i
To be able to calculate the -value-at-risk, ie, the -quantile of the
loss distribution
VaR
(P) = infz R Pr[V

p
> z] 1 = q
(V
P
)
a few simplifying assumptions on the distribution of V
P
have to
be made.
(i) The changes inasset values V
i
over the horizonh are assumed
to be linear in risk-factor changes, ie, there is an S
i
R
m
such
that
V
i
= S
T
i
RF =
m
_
j=1
S
i,j
RF
j
(A1)
316
POTENTIAL FUTURE MARKET RISK
(ii) Furthermore, the changes RF in risk factors over the time
horizon h are assumed to be normally distributed
RF J(0, C) (A2)
with mean zero and covariance matrix C.
Note that S
i
= (S
i,1
, . . . , S
i,m
)
T
is the vector of sensitivities of the
instrument value V
i
with respect to the risk factors, ie
S
i,j
:=
V
i
RF
j
On the portfolio level, the change in portfolio value is thus given by
V
P
=
n
_
i=1
S
T
i
RF = S
T
P
RF with S
P
:=
n
_
i=1
S
i
Under Assumptions A1 and A2, the loss V
P
is normally dis-
tributed with mean 0 and variance S
T
P
C S
P
. Therefore, the value-
at-risk can be calculated as
VaR
(P) = q
(V
P
) =
1
()
_
S
T
P
C S
P
As can be seen from this equation, the risk gure depends both
on the sensitivity vector S
P
and on the covariance matrix C. In the
following, we will investigate a specic simple example in more
detail in order to derive some interesting conclusions about these
dependencies. In particular, we are concerned with model risk, for
which we distinguish two different types.
Model risk of the rst kind. This covers the risk that arises
from the fact that Assumptions A1 and A2 may be violated.
Usually, parameter uncertainty, ie, estimation risk, would be
included.
Model riskof thesecondkind. This covers the riskthat the risk
gure, ie, the VaR, of a position changes in an adverse fashion
over time, although the position itself remains unchanged, or
is even reduced. This kind of risk is usually not covered by any
risk measure.
On the one hand, model risk of the rst kindis usually an issue for
risk measurement on small time scales (see, for example, Figlewski
2003 or Hendricks 1996), but it can, for example, be identied and
controlled via back-testing procedures (Kupiec 1995), or it can be
317
MODEL RISK
handled via more sophisticated models.
1
On the other hand, the
secondkindof model riskis important for proper limit andeconomic
capital management over larger time horizons (eg, one year) andcan
only be mitigated by suitable management action.
Before the beginning of the 20089 nancial crisis, model risk of
the second kind was not an issue for banks or nancial institutions,
2
as anyunwantedshift or increase inVaRgures couldbe easilycoun-
tered by an appropriate hedging or risk reduction action. Since the
second half of 2007, however, signicant parts of trading portfolios
became more and more illiquid and risks could no longer be hedged
adequately. Even worse, due to deteriorating market conditions,
risk gures increased in an unpredicted fast and threatening fash-
ion. Eventually, increased risk gures could lead to limit breaches
or, in a worst-case situation, to additional capital requirements to
keep the nancial institution solvent. These developments induced
our subsequent examination of this topic, which makes three main
contributions.
We will discuss in detail why todays VaR is not a reliable
estimate for VaR gures in the future, even if the portfolio
remains constant, which is due to
ageing effects which change sensitivities,
the impact of risk factor levels on sensitivities,
changing volatility and correlation gures.
This is a key observation for both limit management and the
economic capital framework within a nancial institution.
We will introduce the concept of potential future VaR, basedon
similar ideas from counterparty credit risk modelling. We will
then elaborate on how this concept can be used to handle the
second kind of model risk for the purpose of limit and capital
management.
We will illustrate how the potential future VaR concept,
together with known (marginal or incremental) risk gures,
can also be used for a proactive identication of the vulner-
ability of a portfolio, which can result in suitable early risk
management, ie, hedging activities.
318
Figure 14.1 Time series of risk factors from February 11, 2004, to
July 17, 2009
2004 2005 2006 2007 2008 2009
0
1
2
3
4
5
1
M

L
i
b
o
r

(
%
)

20
0
20
40
60
80
100
120
140
1
0
Y

I
t
a
l
i
a
n

s
p
r
e
a
d

(
b
p
)

10Y Italian spread
1M Libor
For this purpose, our exposition will focus on model risk of the
second kind only. In contrast to the small existing literature on this
topic, we introduce a complete framework to measure and han-
dle estimation risk for practical purposes, instead of just describing
condence intervals around VaR estimates as in Christoffersen and
Goncalves (2005) or Jorion (1996), in which the statistical properties
of VaR gures were investigated in more detail. Taking the ideas
fromChristoffersen and Goncalves (2005) further, our procedure for
calculating potential future VaR is based on their resampling idea.
A specic example of a two-factor delta-normal framework
Risk factors
For ease of presentation, we consider a two-factor delta-normal VaR
approach, assuming that market movements can be explained by
one interest rate risk factor r and one credit-spread risk factor s.
Both risk factors are assumed to have a at term structure. The at
term structure itself, however, is assumed to change over time. At
time t it is represented by the corresponding zero rates r
t
and s
t
. We
have chosen the one-month Euribor as interest rate risk factor and a
10-year asset swap spread time series for Italian government bonds
as credit-spread risk factor (Figure 14.1).
As rather similar results are obtained if a more general frame-
work (ie, varying term structures) is considered, we focus on this
simplied setting for brevity and clarity of presentation.
319
MODEL RISK
Figure 14.2 Historical values of (a) volatilities and (b) correlations
2005 2006 2007 2008 2009
0
2
4
6
V
o
l
a
t
i
l
i
t
y

(
b
p
)

2005 2006 2007 2008 2009
0.05
0.10
0.10
0
0.05

C
o
r
r
e
l
a
t
i
o
n
(a)
(b)
Values estimated from a 250d history. Black line in (a) denotes interest rate
volatility; grey line denotes credit-spread volatility.
The covariance matrix C
t
at time t is normally estimated from
N = 250 previous daily observations with risk-factor changes
(RF
t
)
=0,...,N1
, where RF
t
:= RF
t
RF
t1
. Under the assump-
tion that
1
N
N1
_
=0
RF
t
0
the (unbiased) maximum likelihood estimator

C
t
, which is usually
chosen for estimating the correlation matrix C
t
, becomes
C
t
=
1
N 1
N1
_
=0
RF
t
RF
T
t
To be able to consider the uctuation of parameters in a more prac-
tical way, let us decompose the covariance matrix C into a diagonal
matrix of volatilities and a correlation matrix K, ie, C
t
=
t
K
t
t
.
Using this decomposition, the historical estimates for volatilities
3
and correlation of the two risk factors r
t
and s
t
are illustrated in
Figure 14.2.
From there it becomes clear that risk parameters such as volatili-
ties and correlations do not remain constant over time, but change
considerably and thus impact on future VaR gures, which leads to
model risk of the second kind. Furthermore, it can be observed that
these changes take place in different fashions: while the interest rate
volatility jumps by a very large amount, the credit-spread volatility
increases rather steadily. Independent of the type of change, over
time windows of one year, both the interest rate volatility and the
320
credit-spread volatility take values which range over several basis
points.
Based on a simple toy example which is introduced in the next
section, wewill investigate(seepage325) inmoredetail howchanges
in volatilities and correlation impact on VaR numbers, and we will
compare the magnitude of their inuence with the effect of ageing
and risk-factor levels (both examined later; see page 323 onwards).
Toy portfolio
Together with the assumed simple market consisting of two risk
factors only, we consider a hypothetical portfolio of 15 oating-rate
Italian government bonds with notional N
i
and maturity M
i
= i
years (i = 1, . . . , 15), paying one-month Euribor plus a spread c
i
. We
assume that, at the date of issuance
4
t
I
, the spread c
i
is set in such a
way that each bond is at par, ie
V
i
(t, r
t
, s
t
, c
i
) = N
i
for t = t
I
where V
i
(t, r
t
, s
t
, c) denotes the value
5
of instrument i at time t for
given market risk factors r
t
and s
t
.
As the aim of this exposition is to investigate the model risk of
the second kind in more detail, it is necessary to thoroughly under-
stand the inuence of ageing and risk-factor levels on portfolio sen-
sitivities. Before we do so (see the next subsection), we start with
an illustration of the impact of the risk-factor levels on the present
value of the individual instruments. As can be seen fromFigure 14.3,
which shows the example of a 10-year bond, the inuence of the
interest rate on the bond value depends on the credit-spread level
observed in the market. While for low credit-spread levels the bond
value is not sensitive to changes in interest rates, this is different
for higher credit-spread levels, where the fair value depends on the
interest rate level in an increasing and concave fashion. The curves
in Figure 14.3(b) illustrate this phenomenon in more detail for two
exemplary credit-spread levels. It can also be observed that the fair
value depends decreasingly and convex on the credit-spread level,
and that this dependence is, in turn, related to the level of interest
rates. These effects will be investigated in more detail in the next
section.
Finally, inFigure 14.4 it is shownhowselectedinstruments as well
as the whole portfolio have evolved over time after issuance.
321
MODEL RISK
Figure 14.3 Joint inuence of interest rate and credit-spread level on
bond value at issuance
110
100
90
80
70
120
100
80
60
60
V

(
%

o
f

n
o
t
i
o
n
a
l
)
V

(
%

o
f

n
o
t
i
o
n
a
l
)
100
10
5
0
0
0
200
300
400
Credit spread (bp)
Interest
rate (%)
Interest rate (%)
(a)
(b)
1 2 3 4 5 6 7 8 9 10
(a) Inuence of interest rate and credit-spread; (b) inuence of interest rate (black
line, original credit spread (5.4bp); grey line, high credit spread (400bp)).
Figure 14.4 Bond and portfolio values over time
2005 2006 2007 2008 2009
85
90
95
100
105
V

(
%

o
f

n
o
t
i
o
n
a
l
)
V

(
%

o
f

n
o
t
i
o
n
a
l
)
2005 2006 2007 2008 2009
85
90
95
100
105
85
90
95
100
105
V

(
%

o
f

n
o
t
i
o
n
a
l
)
V

(
%

o
f

n
o
t
i
o
n
a
l
)
85
90
95
100
105
2005 2006 2007 2008 2009 2005 2006 2007 2008 2009
(a) (b)
(d) (c)
(a) 5y bond value, (b) 10y bond value, (c) 15y bond value, (d) portfolio value.
322
Portfolio sensitivities
The sensitivity of the bond values to changes in the risk factors r
t
and s
t
are usually expressed in terms of one basis point, ie
S
i,r,t
=

r
t
[V
i
(t, r
t
, s
t
, c)]
1
10, 000
and
S
i,s,t
=

s
t
[V
i
(t, r
t
, s
t
, c)]
1
10, 000
We will now investigate how much sensitivities may change due
to varying market conditions, and how these changes compare
with known portfolio ageing effects. In more detail, we want to
understand
byhowmuchinterest rate andcredit-spreadsensitivitychange
over time compared to changes in volatilities or correlations,
how strong interest rate level and credit-spread level impact
sensitivities, and
how these level effects compare against known and easy-to-
forecast ageing effects of the portfolio.
For this purpose let us start withFigure 14.5, whichshows howthe
interest rate and credit-spread sensitivity of the hypothetical port-
folio evolve over time after issuance. In Figure 14.5(a) we see that,
immediately after each xing, interest rate sensitivity jumps due
to the increased interest rate risk of the rst xed coupon. There-
fore, we prefer to focus on the xing dates only, as in Figure 14.5(b),
where noise from the rst coupon disappears. Surprisingly, it can
be observed that the sign of the interest rate sensitivity changes
6
and that interest rate sensitivity increases by a signicant factor over
time. Note that the change is comparable in scale with the change in
volatilities, ie, this does not represent a negligible effect.
If we look at the evolution of the credit-spread sensitivities
over the same time horizon (Figure 14.5(c)), we can observe that
credit-spread sensitivities are a factor 100 larger than interest rate
sensitivities for the particular portfolio.
Furthermore, in Figure 14.6 we illustrate interest rate risk, credit-
spread risk and total risk of the hypothetical portfolio P over time.
Credit-spread risk is the main driver of the portfolio risk and, there-
fore, interest rate risk is neglected for the following investigations.
323
MODEL RISK
Figure 14.5 Portfolio sensitivity over time for N
i
equal to 1 million
2005 2006 2007 2008 2009
200
0
200
400
200
0
200
400
2005 2006 2007 2008 2009

2005 2006 2007 2008 2009
12,000
10,000
8,000
S
e
n
s
i
t
i
v
i
t
y

(

)
S
e
n
s
i
t
i
v
i
t
y

(

)
S
e
n
s
i
t
i
v
i
t
y

(

)
(a)
(b)
(c)
(a) Portfolio interest rate sensitivity; (b) portfolio interest rate sensitivity at xing
dates (crosses, including credit-spread risk; circles, without credit-spread risk);
(c) portfolio credit-spread sensitivity.
Figure 14.6 Decomposition of total risk in interest rate risk and
credit-spread risk (at xing dates)
2005 2006 2007 2008 2009
0
10
20
30
40
50
60
R
i
s
k

(
0
.
0
1
%

o
f

n
o
t
i
o
n
a
l
)

Interest rate risk
Credit-spread risk
Of course, for other types of portfolio, interest rate risk may be the
dominant type of risk.
To understand how ageing, interest rate level and credit-spread
level impact the credit-spread sensitivity, let us refer to Figure 14.7,
where the impact on credit-spread sensitivity is given as the ratio
324
Figure 14.7 Impact of ageing, interest rate and credit-spread levels on
portfolio credit-spread sensitivity
2005 2006 2007 2008 2009
I
m
p
a
c
t
/

o
r
i
g
i
n
a
l

s
e
n
s
t
i
v
i
t
y

0.5
0.4
0.3
0.2
0.1
0
0.1
Thick black line, credit-spread sensitivity; light-grey line, credit-spread sensitivity
without ageing; dark-grey line, credit-spread sensitivity without ageing and interest
rate effects; dashed line, credit-spread sensitivity without ageing and interest rate
and credit-spread effects.
to the original portfolio sensitivity, (S
P,s,t
S
P,s,t
I
)/S
P,s,t
I
. As S
P,s,t
0
for t t
I
, a negative impact means decreasing absolute sensitiv-
ity values. As can be seen from the gure, most of the changes in
credit-spread sensitivity are due to ageing, which reduces the orig-
inal credit-spread sensitivity signicantly over time. Large changes
in interest rate and credit-spread levels, however, can offset ageing
effects, as observed starting from the end of September 2008. Sur-
prisingly, observed changes in interest rates seem to have a much
higher impact on credit-spread sensitivity than observed changes in
credit spreads.
7
While the portfolio value itself is not very sensitive
to changes in interest rates, the credit-spread sensitivity still is.
Model risk of the second kind
Summarising the ndings from the previous two sections we can
conclude that, over longer time horizons like one year, volatilities
and thus the covariance matrix C may change signicantly, and,
moreover, sensitivities S may change due to changes in levels of risk
factors and due to ageing.
Although it is obvious that these effects lead to model risk, the
strength of the impact of the individual changes on VaR numbers
325
MODEL RISK
Figure 14.8 Impact of sensitivity and covariance parameters on
portfolio VaR
2005 2006 2007 2008 2009
L
o
g
a
r
i
t
h
m
i
c

i
m
p
a
c
t

Sensitivity impact
Covariance impact
Total impact
2.0
1.5
1.0
0.5
0
1.0
0.5
remains unclear. One way to obtain more information about the vul-
nerability (ie, sensitivity) of the VaR to changes in the parameters is
to look at incremental or marginal VaR, or, more generally, to exploit
gradient information of the VaR with respect to parameters (see, for
example, Rosen 2001), as follows.
Sensitivity of VaR to changes in volatilities is given by
,j
:=
VaR
(P)
j
Sensitivity of VaR to changes in sensitivities S is given by
S,j
:=
VaR
(P)
S
P,j
Then, the maximum component of
,j
or
S,j
shows to which
volatility or risk position the VaR is most vulnerable and by how
much. However, as gradients only allow for the consideration of
instantaneous effects, an alternative way to investigate the impacts
is to look at the historical evolution of the VaR.
InFigure 14.8, the logarithmic change inVaRgures fromissuance
t
I
to time t is decomposed into two components: the logarithmic
change in VaRcaused by changes in the covariance matrix C and the
326
logarithmic change in VaR caused by changes in the sensitivities
ln
_
VaR
(P, t)
VaR
(P, t
I
)
_
. , .
total impact
= ln
_
_
_
S
T
P,t
C
t
S
P,t
_
S
T
P,t
C
t
I
S
P,t
_
_
. , .
impact of covariance
+ln
_
_
_
S
T
P,t
C
t
I
S
P,t
_
S
T
P,t
I
C
t
I
S
P,t
I
_
_
. , .
impact of sensitivity
As can be seen fromthe gure, the impact of sensitivities on VaRg-
ures over time is signicant; VaRis reduced by up to 52%of the orig-
inal value.
8
Compared with the changes in VaR caused by changes
in the covariance matrix, it can be observed that these effects even
account for amuchlarger factor, of 4.9.
9
Wefurthermoreobservethat,
from September 2008 onwards, changes in volatilities dominate all
other changes. From these numbers it becomes clear that a once fea-
sible portfolio (ie, within market risk limits) could easily exceed its
limit over time by a large or actually economic-capital-threatening
amount, if it does not age (ie, mature) fast enough.
THE CONCEPT OF POTENTIAL FUTURE MARKET RISK
Instead of looking at the decomposition of historical movements as
above, potential future evolutions of market environments could be
taken into account. This immediately leads to the new concept of
potential future value-at-risk (PFVaR), which we introduce in the
following. The concept of PFVaR is supplemental to traditional VaR
gures andis closely relatedto the potential future exposure concept
in counterparty credit risk (Pykhtin 2005). As it is an add-on to an
existing market risk framework, banks can use it without the need
to amend their current market risk systems.
Denition and motivation
Potential future VaRgures can be expressed by the potential future
sensitivities S and the future covariance matrix C. Therefore, from
todays perspective, these known parameters become uncertain ran-
dom variables, thus making the future VaR gure a random vari-
able as well.
10
Since the random variable S depends on the future
levels of the risk factors (besides deterministic ageing effects), the
risk factors will be the main stochastic drivers of the future VaR.
Meanwhile, the future covariance matrix C depends not only on the
future risk-factor levels, but on their whole evolution (ie, random
path) until then. This makes potential future VaR a path-dependent
327
MODEL RISK
random variable, which is coupled to the whole path of risk-factor
evolutions.
For a proper denition of potential future VaR, let us x a static
portfolio P and a reference time t
R
(ie, today). Then, similarly to
Pykhtin (2005), a few versions of potential future value-at-risk can
be introduced, as follows.
11
The expected VaRat time T > t
R
is the mean of the distribution
of the (random) future VaR at time T
EVaR
(P, T) := E[VaR
(P, T)
t
R
] (14.1)
The peak VaR at time T > t
R
is the maximum VaR that is
expected to occur at time T at a given condence level
(0, 1)
PVaR
,
(P, T) := q
[VaR
(P, T)
t
R
] (14.2)
The maximum peak VaR until time T > t
R
is the maximum
peak VaR that is expected to occur in [t
R
, T]
MPVaR
,
(P, t
R
, T) := max
t[t
R
,T]
PVaR
,
(P, t) (14.3)
Although we would prefer to switch the order of maximisa-
tion and percentile in the denition of the maximumpeak VaR
in Equation 14.3, as it is more meaningful for practical appli-
cations, we have kept the above denition, as it resembles its
twin concept in counterparty credit risk.
As we have dened three different notions of PFVaR, we nowdis-
cuss what these different gures can be used for in practice. Firstly,
the expected VaRmay be used for the steering of portfolios, as it tells
the portfolio manager or trader how quickly or slowly the VaR will
decay due to ageing effects. This can be used in addition to standard
analysis on the maturity prole of trading portfolios. Secondly, the
peak VaR represents merely an auxiliary quantity used to dene the
muchmore relevant maximumpeak VaR. We propose the maximum
peak VaR as the right measure for both limit management and an
economic capital framework, but with different condence levels
and time horizons T. In addition to limits on ordinary VaR gures,
we suggest the introduction of limits on MPVaR, for example with
a of 95% and a time horizon of one to three months, as no port-
folio manager or trader should be allowed to enter into positions
328
which might violate the VaR in the near future. This can be seen as a
limit on a stress case, as it is assumed that there is no liquid market
where the risk can be reduced. Using the appropriate from the
economic capital framework (usually around 99.95%) together with
a longer time horizon of one year, the MPVaR represents the max-
imum VaR for which a nancial institution should be capitalised.
As it can be reasonably assumed that, in stress scenarios which are
presumed for economic capital calculations, the market is broken
down and no management intervention is possible, we believe that
this number is the correct number a regulator should want to see in
an economic capital framework.
12
Analogously to the section on model risk of the second kind (see
page 325 onwards), all the above-mentioned newrisk gures can be
decomposed in the same way as we have decomposed the historical
VaR gures, which allows us to identify if the potential future VaR
is driven by effects from volatilities, correlations, ageing or level
impact on sensitivity. This in turn allows us to develop potential
countermeasures toprepare for anunfavourable evolutionof market
risks. For example, long option positions could be used to hedge
against an increase in volatility, or positions without convexity can
be switched into positions with higher convexity.
Calculation of potential future VaR
So far, we have introduced the general framework for potential
future VaR, but we have not yet given details how the new risk
gures (Equations 14.114.3) can be calculated. Before we specify
the differences of possible distinct implementations, let us describe
the general methodology for the calculation of PFVaR.
Algorithm 14.1. For a given portfolio P at time t
R
with given risk-
factor values RF
t
R
,1
, . . . , RF
t
R
,m
(including sufciently large historical
time series), nal time horizon T and a xed number of simulation
runs K, the PFVaR gures can be approximated by the following
Monte Carlo algorithm.
(1) Path simulation. For each scenario k = 1, . . . , K, simulate for
each time step = t
R
+ 1, . . . , T future risk-factor values
RF
k
,1
, . . . , RF
k
,m
based on given risk-factor values RF
k
up to
time .
(2) VaR calculation. For each scenario k = 1, . . . , K, calculate
for each time step = t
R
+ 1, . . . , T the corresponding VaR
329
MODEL RISK
gure VaR
k
(P, ) based on the appropriate risk-factor time
series up to time , ie, based on recalculated sensitivities and
covariances.
(3) PFVaR calculation. Based on the K Monte Carlo samples of
VaR
k
(P, ), the expected VaR, the peak VaR and the maximum
peak VaR can be easily calculated.
For the calculation of each individual VaR
k
(P, ) in each scenario
and time step in step 2 of the algorithm, the usual VaR calculation
methodology should be used. For practical applications this means
a high computational effort, as in each of the K(Tt
R
+1) sample
points both the sensitivity vector and the covariance matrix need
to be calculated. Since in practice sensitivity calculation rather than
the risk calculation itself is usually the bottleneck of this calculation,
two different approaches for the VaR calculation in step 2 can be
distinguished as follows.
Approach 1 (full run): at each time , the sensitivity S
k
P,
of
the portfolio is calculated for each scenario k, depending on
the simulated risk factors at time and the remaining time
to maturity as seen from time . The covariance matrix C
k
is
estimated from a history of risk-factor observations, ie, C
,k
depends not only on the time but also on the whole path of
historical risk-factor observations and simulations.
Approach 2 (reduced run): at each time and in each scenario
k, the sensitivity of the portfolio from the reference time t
R
is
used, ie, S
k
P,
= S
t
R
. The covariance matrix is estimated in the
same way as in approach 1.
While approach 1 can be usedto quantify the complete model risk
of the second kind, approach 2 only accounts for changing covari-
ances. For short time horizons such as one day or 10 days, changing
risk parameters are the main drivers of risk, but as we have seen
earlier (page 325 onwards), the ageing of positions and the impact
of risk-factor levels on sensitivities is important when considering
longer time horizons. If the calculation of sensitivities is not feasi-
ble due to unavailable computer resources, we suggest that future
sensitivities should be estimated by an at-least linear model in time
and risk-factor levels.
330
In step 1, the simulation of future risk factors has not been spec-
ied in detail. In principle, a wide range of methods, from fully
parametric to completely non-parametric, may be used.
Historical bootstrapping represents the fully non-parametric
methodology for the simulation of risk factors. By sampling
risk-factor changes from historical data, future risk factors
can be simulated without specifying any joint probability
distribution of the risk factors.
Stochastic differential equations or some type of Garch process
are the fully parametric alternative to bootstrapping. In this
case, some model for the evolution of risk factors is specied
and the parameters are estimated from historical data.
Mixed models, where historical bootstrapping results are
blended with simulated risk-factor evolutions (eg, by some
type of regime switching concept) can also be implemented.
Depending on the type of application and the availability of data,
all approaches have their pros and cons. Generally speaking, the
situation can be compared to the pros and cons of historical VaR
versus delta-normal VaRor a full Monte Carloriskcalculation. Using
the following toy portfolio described earlier (see page 321), we will
see howa bootstrappingapproachcompares witha mixedapproach,
especially in times of a nancial crisis.
Application of PFVaR to the toy portfolio
Before we discuss the PFVaR results in detail, we start with an illus-
tration of a fewbootstrapped scenarios. In Figure 14.9 we showhow
a sample of 100 bootstrapped scenarios of length 250 days compares
against the actual historical evolution of the risk factors.
Figure 14.10 shows the quantiles of the simulated interest rates
and credit-spread sensitivities over time as ratio to the original sen-
sitivity at the simulation start date. The gure clearly demonstrates
both the impact of the ageing effect (downward trend) and interest
rate and credit-spread levels (widening of distribution) on credit-
spread sensitivities, while interest rate sensitivities are not effected
by ageing but vary on a much larger scale.
Finally, Figure 14.11 shows the quantiles of simulated interest
rate and credit-spread volatilities over time as a ratio to the original
volatility observed at the simulation start date. It can be observed
331
MODEL RISK
Figure 14.9 Interest rate and credit-spread paths obtained by historical
bootstrapping
Jan 07 Apr 07 Jul 07 Oct 07Jan 08Apr 08 Jul 08 Oct 08Jan 09
0
2
4
6
8
I
n
t
e
r
e
s
t

r
a
t
e

(
%
)

Jan 07 Apr 07 Jul 07 Oct 07Jan 08 Apr 08 Jul 08 Oct 08 Jan 09
100
50
0
50
100
C
r
e
d
i
t

s
p
r
e
a
d

(
b
p
)

(a)
(b)
Black line, actual path; grey lines, simulated paths. (a) Simulated interest rate path;
(b) simulated credit-spread path.
Figure 14.10 Quantiles of simulated sensitivities over one year
Jan 08 Mar 08 May 08 Jul 08 Sep 08 Nov 08 Jan 09
5
0
5

Jan 08 Mar 08May 08 Jul 08 Sep 08 Nov 08 Jan 09
0.7
0.8
0.9
1.0
1.1
S
e
n
s
i
t
i
v
i
t
y

/

o
r
i
g
i
n
a
l
s
e
n
s
i
t
i
v
i
t
y
S
e
n
s
i
t
i
v
i
t
y

/
o
r
i
g
i
n
a
l
s
e
n
s
i
t
i
v
i
t
y
1
10
20
30
40
50
60
70
80
90
99
Original
Sensitivity (%)
(a)
(b)
(a) interest rate sensitivities at xing dates; (b) credit-spread sensitivities.
that volatilities go down if a jump falls out of the observation time
window, but otherwise volatilities tend to go up only. This upward
332
Figure 14.11 Quantiles of simulated volatilities over one year
V
o
l
a
t
i
l
i
t
y

/
o
r
i
g
i
n
a
l

v
o
l
a
t
i
l
i
t
y
V
o
l
a
t
i
l
i
t
y
/

o
r
i
g
i
n
a
l

v
o
l
a
t
i
l
i
t
y

Jan 08 Mar 08May 08 Jul 08 Sep 08 Nov 08 Jan 09
1
10
20
30
40
50
60
70
80
90
99
Original
Volatility (%)
2.5
2.0
1.5
1.0
0.5
1.5
1.4
1.3
1.2
1.1
1.0
0.9
(a)
(b)
(a) interest rate volatilities at xing dates; (b) credit-spread volatilities.
Figure 14.12 Potential future credit-spread VaR and actual
credit-spread VaR evolution under the bootstrapping model
V
a
R
/

o
r
i
g
i
n
a
l

V
a
R
True VaR
EVaR
PVaR
MPVaR
0.8
1.0
1.2
1.8
1.4
2.0
2.2
1.6
trend in volatility is due to the bootstrapping methodology, where
the older, low variation observations are not included in the esti-
mation window, but are replaced by resampled, usually higher
variation observations.
All in all, we can see in Figure 14.12 that, before September 2008,
the concept of potential future VaR works well. First of all, the
333
MODEL RISK
expected VaR gives the information that due to ageing effects the
credit-spread VaR is expected to decrease slightly on average. Sec-
ondly, the (maximum) peak VaR gives a reasonable upper bound on
VaR, as long as there is no regime switch in the market. An MPVaR
of approximately 1.3 means that the VaR may increase by 30%, if
history repeats itself in an unfavourable fashion.
However, the regime switch due to the nancial crisis in Septem-
ber 2008 cannot be predicted by the model as such large jumps in
credit spreads had not been observed in the past and are therefore
not covered by the historical bootstrapping methodology.
To overcome this drawback from pure historical bootstrapping,
we suggest moving fromthe bootstrapping methodology to a mixed
model, where ideas from regime-switching models are used in con-
junction with features from extreme value distributions. Historical
data can be used to estimate the probability of a regime shift to an
extreme scenario and the parameters of such an extreme scenario or,
alternatively, expert opinion may be incorporated. Although a boot-
strapping model may be still sufcient for limit management and
portfolio steering, we believe that, for the purpose of economic cap-
ital management, a more sophisticated model should be preferred
over an ordinary bootstrapping model. For brevity of presentation,
we demonstrate this idea by means of a very simple approach for
the simulation of paths under a potentially stressed market environ-
ment: for each scenario k and each date , a switch to an extreme
scenario is assumed to occur with a probability of p, and in this
case the risk-factor change obtained from ordinary bootstrapping is
multiplied by a factor f > 1.
Figure 14.13 shows the simulation results in the case in which
an extreme scenario ( f = 5) may be reached with a probability of
5%, which in times of a nancial crisis is a rather mild assumption.
After the incorporation of extreme scenarios in the PFVaR model,
the MPVaR increases dramatically to almost 250%. This means that
the nancial institution needs to have enough capital available to
cover an increase of 150% in market risk, or that suitable counter-
measures (eg, option positions to hedge for volatility increase) need
to be considered. Furthermore, it is interesting to note that the model
predicts not an instantaneous explosion of VaR, but rather a steady
increase, which is exactly what we can observe with the beginning
of the crisis in September 2008.
334
Figure 14.13 Potential future credit-spread VaR and actual
credit-spread VaR evolution under the bootstrapping model with
stressed scenarios
V
a
R
/

o
r
i
g
i
n
a
l

V
a
R

True VaR
EVaR
PVaR
MPVaR
1.5
1.0
2.5
2.0
SUMMARY AND OUTLOOK
In this chapter we have shown that value-at-risk gures cannot be
expected to be constant, but may vary over time for several rea-
sons; the most important reasons are changes in volatilities, changes
in sensitivities due to ageing and shifts in market environments.
We have shown that all effects are similarly important, but that
increasing volatilities dominated during the 20089 nancial crisis.
As a potential remedy against model risk arising from these
changes, the new concept of potential future market risk has been
introduced. We have discussed how new risk gures may be used
for improvedportfolio steering andlimit management. We have fur-
ther pointed out why and how the concept should be included in
economic capital frameworks to account for model risk.
We have also seen that PFVaR numbers depend crucially on
the simulated risk factors, ie, a certain dependence on models and
assumptions cannot be avoided. Therefore, we have discussed dif-
ferent alternatives, rangingfromparametric tonon-parametric mod-
els, and we have indicated what we believe to be a reasonable set-
up. In our opinion, a nal detailed specication of the PFVaR frame-
work, especially inconjunctionwitheconomic capital models, needs
more investigation in order to be able to give guidance for practical
implementations. Nonetheless, we hope and believe that the PFVaR
concept will gain in popularity due to its easy application and its
valuable additional information.
335
MODEL RISK
The views expressed in this chapter are the authors own and
should not be construed as being endorsed by Deutsche Pfand-
briefbank AG.
1 See, for example, Kamdem(2005) for an extension to elliptical distributions, Alexander (2001)
for a non-parametric linear historical or Monte Carlo VaR or Bams and Wielhouwer (2001)
for adjustment factors for model risk of the rst kind.
2 As an exception, the exposition by Danielsson (2002) points out that VaR gures are volatile
and not reliable in general.
3 The signicant jumpininterest rate volatilityat the endof 2007 is due toanupwardjumpinthe
one-month Libor. At the endof 2008, the interest rate volatility falls down again because of the
aforementioned observation dropping out of the 250-day historical period contributing to the
volatility estimate. Using the similarly popular exponentially weightedcovariance estimation
scheme instead of the maximumlikelihood estimator (see, for example, Alexander 2001), this
phenomenon would be smoothed out.
4 The date of issuance, February 11, 2005, is chosen such that there is a one-year history of
risk-factor observations available to estimate volatilities and correlations fromhistorical data.
5 Aderivation of the valuation for credit-risky bonds can, for example, be found in Choudhry
(2004).
6 Note that the signof the sensitivitybecomes positive, whichis indeeda surprisingobservation
for a oating-rate note. Adetailedinspection reveals that this is due to increasedcredit-spread
levels.
7 Although the credit-spread sensitivity of the portfolio is a factor 100 higher than the interest
rate sensitivity, the credit-spread gamma and the cross gamma have the same magnitude.
As changes in interest rates since issuance were considerably higher than changes in credit
spreads, this explains why observed changes in interest rates have a higher impact on credit-
spread sensitivity than observed changes in credit spreads.
8 exp(0. 65) 0. 52.
9 exp(1. 59) 4. 9.
10 We assume that a suitable probability space with corresponding ltration is chosen such that
all formulas appearing exist and are well dened.
11 Note that the concept of potential future VaR can easily be generalised to potential future
market risk when replacing VaR by any arbitrary market risk measure.
12 If a nancial institution can prove that it has a catalogue of countermeasures, these should of
course be acknowledged by setting suitable factors less than 1.
REFERENCES
Alexander, C., 2001, Market Models: A Guide to Financial Data Analysis (Chicago, IL: John
Wiley & Sons).
Bams, D., and J. L. Wielhouwer, 2001, Empirical Issues in Value-at-Risk, Astin Bulletin
31(2), pp. 299315.
Choudhry, M., 2004, Analysing and Interpreting the Yield Curve (Singapore: John Wiley &
Sons).
Christoffersen, P., and S. Goncalves, 2005, Estimation Risk in Financial Risk Manage-
ment, The Journal of Risk 7(3), pp. 128.
Danielsson, J., 2002, The Emperor Has No Clothes: Limits to Risk Modelling, Journal of
336
Figlewski, S., 2003, Estimation Error in the Assessment of Financial Risk Exposure,
Working Paper, SSRN, URL: http://ssrn.com/abstract=424363.
Hendricks, D., 1996, Evaluation of Value-at-Risk Models Using Historical Data,
Economic Policy Review, April, pp. 3969.
Jorion, P., 1996, Risk
2
: Measuring the Risk in Value at Risk, Financial Analysts Journal,
November/December, pp. 4756.
Jorion, P., 2007, Value at Risk: The New Benchmark for Managing Financial Risk, Third
Edition (New York: McGraw-Hill).
Kamdem, J. S., 2005, Value-At-Risk and Expected Shortfall for Linear Portfolios with
Elliptically Distributed Risk Factors, International Journal of Theoretical and Applied Finance
8(5), pp. 53751.
Kupiec, P., 1995, Techniques for Verifying the Accuracy of Risk Measurement Models,
Journal of Derivatives 3, pp. 7384.
Pykhtin, M. (ed.), 2005, Counterparty Credit Risk Modelling (London: Risk Books).
Rosen, D., 2001, Marginal VaRAnalysis and Symmetry in Parametric VaR, Algo Research
Quarterly 4(4), pp. 4952.
337
15
Market Risk Modelling:
Approaches to Assessing
Model Adequacy
Carsten S. Wehn
DekaBank
Like other risk models, market risk models often have their weak-
nesses. A risk model relies on two opposite and often off-trading
incentives: relevancy and credibility. Whereas relevancy means, in
a market risk model context, whether the information given by the
model is in time, complete and understandable, credibility means
whether the information is adequate to the portfolios and risk posi-
tion. As a model is always an approximation to reality, different
assumptions have to be made while constructing it. Nowadays, it
is even more crucial to identify and communicate the modelling
assumptions; it is essential to work continuously to improve the
qualityof market riskmodels withaviewtowards their adequacyfor
risk measurement and management purposes. This chapter draws
on previous publications by the author (Wehn 2005, 2008, 2010) and
other relevant papers. Serving as an overview of the different meth-
ods from a practitioners point of view, it presents in a systematic
and concise manner the building blocks of a market risk model and
describes in detail two approaches for assessing the models ade-
quacy: byassessingthe individual buildingblocks andsubsequently
by backtesting the model as a whole. Embedded in a regular vali-
dation process, this can help to improve the model continuously.
Regulatory issues are also considered. Finally, we discuss how to
quantify the model risk of the risk model and how to treat poten-
tial model weaknesses. The chapter closes with a conclusion and
outlook.
339
MODEL RISK
BUILDING BLOCKS OF MARKET RISK MODELS
The impact of market risk factors on the gains and losses associ-
ated with a portfolio of investments is referred to as market risk.
These risk factors include interest rates, equity prices, currencies or
impliedvolatilities. The starting positionfor applying adequate stat-
istical means to model andmeasure market risk is relatively good, as
market risk factors are generally readily observable. The measure-
ment of market risk is preceded by a prediction of the distribution
of the future (and therefore unknown) gains and losses. These gains
and losses of a portfolio are denoted by a random variable G
t
.
1
This
is completely analogous to the situation encountered when pric-
ing a nancial asset, as the risk modelling of portfolio gains and
losses takes all relevant information into account. This information
is usually given by observed risk-factor prices or risk-factor returns,
respectively. Thus, the relevant (univariate) predictive distribution
is
F
t
:= F
t
(G
t
R
t1
, R
t2
, . . . )
where R
t1
, R
t2
, . . . denote (multivariate) past risk-factor returns.
Several buildingblocks are necessarytobe denedfor the derivation
of a predictive distribution for portfolio gains and losses.
Firstly, an identication of relevant risk factors (such as inter-
est rates for different maturities and rating classes, equity indexes
or equity prices, exchange rates, implied volatilities and so on) is
required.
Secondly, we must choose a relevant multivariate distribution
F
t
(R
t
R
t1
, R
t2
, . . . ) for the risk-factor returns R
t
R
t1
, R
t2
, . . . .
Here, note that we are dealing with a conditional return distribu-
tion (ie, the distribution is conditional on previous observations).
Widely used models include Garch and Arch models for returns of
time series.
2
Time series of risk factors experience so-called stylised
facts. In the model set-up above, this is reected in the conditional
distribution of the risk-factor returns. The stylised facts comprise
autocorrelation of risk-factor values and serially almost uncorre-
lated respective returns. Often volatility clusters are observable for
the returns and the (unconditional) distribution of the risk-factor
returns is leptokurtic. The daily returns distribution has a mean
around zero and is almost symmetrical.
Thirdly, risk models usually make assumptions concerning the
relationship (ie, a mapping) between risk factors (or risk-factor
340
MARKET RISK MODELLING: APPROACHESTO ASSESSING MODEL ADEQUACY
returns) and the portfolio gain and loss function (eg, sensitivities
and Greeks, full valuation, present value grids, etc).
Most risk models show parameters of the conditional distribu-
tion F
t
, such as a quantile for the value-at-risk or conditional means
for the expected shortfall. Value-at-risk, for example, is dened as
the -quantile
3
of the conditional distribution of a portfolios gains
and losses: q
t
= F
1
t
(1 ). Nevertheless, the following meth-
ods focus mainly on the entire conditional distribution and in some
parts also rely on the respective quantiles. For a daily risk mea-
surement process, this mapping between risk factors and portfolio
gains and losses changes, due to new or matured trades and other
effects. Moreover, new information that occurs has to be taken into
account for the derivation of the conditional distribution of the risk
factors returns F
t
(R
t
R
t1
, R
t2
, . . . ). For a current discussion of
calculating the relevant gains and loss gures see, for example, Fin-
ger (2005). As a nancial institution commonly reserves economic
capital for risks such as market risk, it is concerned with not reserv-
ing too much capital (which would be inefcient) or too little capital
(thus underestimating the risks). Therefore, the adequacy of the pre-
dictive distribution is of high importance for the acceptability and
credibility of a risk model in the bank to generate trading impulses
for the correct management measurements.
TYPICAL MODELLING ERRORS
The above-mentioned set-up for modelling market risk helps us to
identify typical errors due to the modelling. Some of these stem
directly from the building blocks introduced. When selecting risk
factors for the model, we face the problem of parsimonious mod-
elling. This means that not all risk factors can be taken into account
within the model. Some of this arises fromthe nature of the real
risk factor; the yield curve, for example, is continuous, implying
that key rates have to be identied as risk factors and consequently
interpolation methods must be dened. We also want the model
to include only risk factors with liquid quotations. Finally, if we
introduce too many risk factors, we face statistical (and potentially
numerical) issues, because the noise associatedwithestimatingadis-
tributionfor theriskfactors increases andtheexplanationof themain
risk drivers in the portfolio becomes more and more complicated.
341
MODEL RISK
Many commondistributions for F
t
(R
t
R
t1
, R
t2
, . . . ) include the
normal distribution. This may only be consistent with the stylised
facts if the time-series properties like heteroscedasticity are also
taken into account. Nevertheless, many distributions used in the
risk modelling context are not sufciently able to reect rare events
with high impact. The estimation of the distribution comes along
with a respective estimation error.
Concerning the mapping between risk factors and the portfolios
gains and losses, it must be admitted that, in practice, approxima-
tions have to be incorporated, as valuation models might be very
complex on a product level and valuation under a certain scenario
might be time-consuming or involve many simulations, thus giv-
ing rise to numerical issues. The potential recalibration of valuation
models is another issue to be considered. Thus, common approx-
imations include linear or quadratic approximation, present value
grids or analytical functions instead of numerical simulations for
valuation functions.
From a regulatory point of view (BCBS 1996), typical errors
include: the basic integrity of the model in situations where a
banks risk system does not capture the risk; incorrect calcula-
tion/algorithmor implementation; the general necessity to improve
a models accuracy; insufcient precision in assessing the risk;
inferential or statistical errors.
Bearing these toeholds in mind, we will later provide insights into
how adequate means for improving the market risk model can be
identied.
RESULTS FROM MODEL RISK IN A MARKET RISK MODEL
Next, we briey discuss the impact of model risk on the risk man-
agement process. Model risk caused by the market risk model can
impact on the risk management process at all steps, which must be
taken into account. During risk analysis and identication, a poor
model may not reect risk adequately or may even completely fail
to take into account an important source of risk. This can lead to
missing impulses for risk management decisions, eg, when the risk
measure does not reect the real risk in the resulting gures.
Potential under- or overestimationof riskondifferent levels might
undermine the models credibility and lead to an inefcient use of
342
riskcapital. Portfolioeffects might be misleadingandresult inwrong
decisions owing to either a false sense of security or insecurity.
The potential impact of model risk in a market risk model makes
it obvious that a tough regular validation and backtesting process is
a necessary requirement for the institution.
A PRIORI VALIDATION
Risk-factor selection
As discussed above, the implementation of the risk model faces the
problem of parsimonious modelling that makes a concentration to
key risk factors and risk factor groups necessary. This raises ques-
tions such as which key rates should be taken out of the yield curve
or whether we are able to model different parts of a volatilitysurface,
etc.
The selection of the relevant risk factors depends to a large extent
onthe strategyof the portfolio, as the followingconsiderationshows.
From a purely stochastic modelling perspective, more than 95% of
the yield curve moves can be explained by parallel shifts, twists and
butteries, ie, the rst three eigenvectors. This enables us to reduce
the number of risk factors dramatically. But if, on the other hand, our
portfolio follows, say, a steepening strategy of between ve and six
years, we need to introduce these risk factors into the model even
if they are highly correlated, since even small differences will show
up in gains or losses.
Hence, the typical means of selecting risk factors rely on com-
mon statistical tools such as regression analysis (for calculating the
respective contributions of the individual risk factors to the port-
folios gains and losses), comparison of different sets of risk factors
and tests on homoscedasticity and autocorrelation.
Risk-factor distribution
Focusing on the modelling of the risk-factors distribution, we recall
the model above. Bearing in mind that we are modelling a condi-
tional distribution, we distinguish between the time-series model
and the distribution of the noise term.
Time-series models used for market risk should incorporate the
stylised facts. Some examples include white noise and (geometric)
Brownianmotion, Arch(autoregressive conditional heteroscedastic-
ity) models with stochastic volatility and variance h
t
= +
2
t1
,
343
MODEL RISK
where
t1
is the residual termfromthe last time step, Garch (gener-
alisedArch) models with h
t
= +
2
t1
+h
t1
and the special case
of the exponential weighting scheme usedby RiskMetrics (Mina and
Xiao 2001).
The relation between risk factors and risk-factor returns is typ-
ically modelled by logarithmic returns (eg, for equities), relative
returns (eg, for interest rates) or absolute returns (eg, for spreads
on interest rates). The respective selection depends on the risk
factor and the observation that the noise term should consist of
independent, identically distributed (iid) random variables.
Common distributional assumptions for the noise term in the
time-series model are the normal distribution or Student-t distri-
butions. As tools for the validation of the composite distribution,
ordinary statistical tests in combination with graphical means like
PP plots or QQ plots seem useful.
Transformation
The third building block of the market risk model is the mapping
between the risk factor (returns) and the portfolios gains and losses.
This mapping is usually done on an instrument level. Here, if the
model (like historical simulation) allows a complete evaluation, no
further approximations (besides potentially some parameters) have
tobe made. If, due tocomplex valuationfunctions or numerical com-
plexity, we have to establish an approximation, this can be done by a
Taylor series expansion (thus by sensitivities, ie, the Greeks). The
derivation is often done numerically by applying a small shift to the
risk factor (differential quotient), thus requiring that an appropri-
ate shift size be determined. Often, larger shifts seem plausible on
an individual basis, as the approximated values are located in the
tails of the distribution. Other practices include a quadratic func-
tion obtained by a three-point approximation, interpolation, use of
present value grids, etc. The goodness-of-t can be judged by some
measure like the L
2
-metric (or L
1
-metric) between the true valuation
function and the approximation.
A POSTERIORI BACKTESTING
By backtesting, we complement the validation of the main building
blocks of the market risk model by a retrospective view. We compare
344
the results of a portfolios risk estimation with the respective gains
and losses.
Theoretical foundation for backtesting methods
Common statistical tests most often require observations stemming
from iid random variables. The market risk measurement set-up
described above highlights that this is clearly not the case within
the backtesting framework. This gives rise to the question of howto
transformthe given data to enable a further application of statistical
methods and an in-depth analysis of howto improve the risk model.
Diebold et al (1998) present a basis for validating the predictions
when observing realised values and consequently for healing the
missing iid property. This approach dates back to an idea by Rosen-
blatt (1952) and uses the so-called Rosenblatt transformation (for a
proof see, for example, Diebold et al (1998) or Wehn (2010)).
Theorem 15.1. Let

f
tX
t1
,X
t2
,...
be the conditional densities gener-
ating the time series (X
t
)
tN
and let f
t
be the respective predictive
densities. These predictive densities are assumed to be continuous
and f
t
(x) > 0 for all x. For the probability integral transform
U
t
:= F
t
(X
t
) =
_
X
t
f
t
(u) du
it holds that
P(F
t
(X
t
) ) = P(U
t
) =
_
f
tX
t1
,X
t2
,...
(F
1
t
(x))
f
t
(F
1
t
(x))
dx
Furthermore, if
f
tX
t1
,X
t2
,...
= f
t
for all t
then F
t
(X
t
) = U
t
U
[0,1]
.
Now, we are in a position to drawconclusions regarding the time
series serial dependence. This is given by the following theorem.
Theorem 15.2. With the same notation and assumptions as in
Theorem 15.1, it holds that
P(U
t

t
, . . . , U
1

1
)
=
_
[0,
t
][0,
1
]
f
tX
t1
,X
t2
,...
(F
1
t
(x
t
))
f
t
(F
1
t
(x
t
))

f
1X
0
,X
1
,...
(F
1
1
(x
1
))
f
1
(F
1
1
(x
1
))
d
t
Furthermore, if

f
tX
t1
,X
t2
,...
= f
t
for all t, then
F
t
(X
t
) = U
t
iid
U
[0,1]
.
345
MODEL RISK
We can then derive a corollary that is helpful for the next steps in
backtesting the model.
Corollary 15.3. Let [a
t
, b
t
] be an interval of the domain of the pre-
dictive distribution with F
t
(b
t
) F
t
(a
t
) = for all points in time t.
If

F
tX
t1
,X
t2
,...
= F
t
for all t, it holds that
1
[a
t
,b
t
]
(X
t
)
iid
Ber
()
These results allowa treatment of the daily pairs of predicted dis-
tributions and realised gains (F
t
, g
t
)
T
t=1
over a period of T trading
days and yields the possibility to adopt mathematical and statis-
tical backtesting methods for the validation of the risk model. The
distribution F
t
() is the above-mentionedconditional predictive dis-
tribution, whereas

F
t
is the non-observable true distribution of the
portfolios gains and losses.
The Rosenblatt transformation forces us to focus on the standard-
ised gains U
t
:= F
t
(G
t
) or, if value-at-risk is the chosen risk measure
on the exceedances, O
t,
:= 1
(,q
t
)
(G
t
). In an ideal risk measure-
ment framework, due to the Rosenblatt transformation, the stan-
dardised gains U
t
should be iid uniformly distributed, ie, U
t
iid
U
[0,1]
and the exceedances should be Bernoulli distributed according to
the corollary, ie, O
t,
iid
Ber
(1)
.
Useful statistical tests for backtesting
This sectionintroduces some statistical tests that canbe usedtoback-
test the risk model. For a more in-depth description of the tests and
thederivationof adecisiontreethat canbeusedinapractical context,
we refer the reader to Wehn (2005, 2008, 2010).
Bearing the Rosenblatt transformation introduced above in mind,
we will examine rst statistical tests based on exceedances and then
statistical tests based on standardised gains. Every statistical test
is characterised by its null hypothesis, critical values and a brief
discussion of advantages and disadvantages.
4
Tests based on exceedances
Probably the best known statistical test for backtesting is the so-
called trafc-light approach (TLA)
TLA
. This test was introduced in
the context of rst allowing internal market risk models for reg-
ulatory purposes (BCBS 1996). The TLA has the null hypothesis
of
H
0
: (1 ) [0, 1 ]
346
meaning that it can be used to judge whether the observed rel-
ative frequency of exceedances is signicantly lower than the
requiredlevel of . As the individual exceedances are to be Bernoulli
distributed, their sum is binomial, with
T
_
t=1
O
t,
Bin
(T,1)
leading to a critical zone of
K
T,
=
_
x 0, 1
T
T
_
t=1
x
i
k
T,
_
withrespective critical values of k
T,
= minm 1F
Bin
(T,1)
(m1)
.
The Basel Committee on Banking Supervision, in a landmark
paper (BCBS 1996), laid down two different values for , namely
y
= 0. 05 and
r
= 0. 0001 for the rst-order error. The interpre-
tation then is as follows: if the null has to be rejected by
y
, then a
yellowlight is shown, if it has to be rejected by
r
, then a red light is
shown. Entering the yellow or red zone is directly linked to higher
regulatory capital requirements and can even lead to a refusal of the
use of the internal risk model for regulatory purposes.
Some of the oft-discussed advantages of the TLAinclude the fact
that it is a simple and understandable concept as well as its use
in setting incentives towards a conservative modelling. However,
the last point could also be considered a disadvantage, as there is
no penalty imposed for overestimating the risks (a conservative risk
model is not necessarilyanadequate one). Exceedances are relatively
rare events (especially on the regulatory level of = 99%), leading
to a small number of observations, whereas the values for
r
and
y
are rather high.
Awhole class of tests is based on the likelihood ratio. This allows
the statistical treatment of the exceedances as well as the whole dis-
tribution, as we will see later on. Likelihoodratiotests (LRTs) are best
uniform selective tests. Kupiec (1995) proposes a two-sided exten-
sion of the TLAnull by H
0
: = . He rst concludes that according
to Theorems 15.1 and 15.2 (including Corollary 15.3), the statistic
t
1
:= mint O
t,
= 1 is geometrically distributed, meaning that
P(t = t
1
) = (1 )
t
1
1
. The respective LRT for the time until
347
MODEL RISK
rst failure (TUFF) is given by
TUFF
= 2 ln
_
(1 )
t
1
1
_
1
t
1
_
1
1
t
1
_
t
1
1
_
1
_
By Wilkss (1938) Theorem (asymptotically), it holds that
TUFF
asympt

2
1
Thus, if
TUFF
> F
1
2
1
(1 ) for a given , the null has to be rejected.
The TUFF test is characterised by a relatively low discriminatory
power and the need for large observation samples for small values
of and large values of (for example, considerably more than 600
when using = 0. 01 and = 0. 99).
5
Kupiec (1995) also proposes a second LRT with the same null. The
test statistic for the proportion of failure (POF) test is given by
POF
= 2 ln
_
(1 )

T
__
T
_
_
1

T
_
T
_
1
_
where :=
T
t=1
O
t,
is again assumed to be binomially distributed.
It again holds that
POF
asympt

2
1
. This LRT corresponds to a two-
sided binomial test, and hence serves as an extension of the above-
mentioned TLA. As for the TLA, the second-order error for the POF
test can also be reduced signicantly by using longer observation
periods. Due to the two-sided property, the errors of the second
kind are clearly higher than those in the TLA, but the POF test
consequently removes the disadvantage of the one-sided TLA.
The rst LRT focusing on the serial dependence of exceedances
basedonthe corollaryis a statistic proposedbyChristoffersen(1998).
It takes into account a transition matrix of the kind
_
p
00
p
01
p
10
p
11
_
where p
ij
:= P(O
t,
= j O
t1,
= i). Independence of sequential
exceedances results in
p
ij
:= (p
0i
+p
1i
) (p
0j
+p
1j
) = P(O
t,
= j)P(O
t1,
= i)
Accordingly, the null hypothesis is given by
(p
01
, p
11
)
T
(x, x)
T
x (0, 1) =
0
= (0, 1)
2
348
and, with the denition of the number of observations capturing
rst state i and then state j,
t
ij
:=
T
_
t=1
1
j
(O
t,
)1
i
(O
t1,
) for i, j 0, 1
the respective statistic is given by
ind
= 2 ln
_
1
(1 p
01
)
t
00
p
t
01
01
(1 p
11
)
t
10
p
t
11
11
_
1
t
01
+t
11
T 1
_
(t
00
+t
10
)
_
t
01
+t
11
T 1
_
(t
01
+t
11
)
_
with the estimator
p
i1
:=
t
i1
t
i1
+t
i0
For the respective LRT statistic, it follows that
ind
asympt

2
1
. A very
similar test is the
2
statistic, with
2
(ind)
=
1
_
i=0
1
_
j=0
_
t
ij
(t
i0
+t
i1
)(t
0j
+t
1j
)
T 1
_
2
T 1
(t
i0
+t
i1
)(t
0j
+t
1j
)
and
2
(ind)
asympt

2
1
.
Christoffersen(1998) combines the POFtest withthe depictedLRT
test on independence, which results in a null hypothesis of
(, p
01
, p
11
)
T
(, x, x)
T
x (0, 1) =
0
= (0, 1)
3
and derives an LRT test
combined
= 2 ln
_

(t
00
+t
10
)
(1 )
(t
01
+t
11
)
(1 p
01
)
t
00
p
t
01
01
(1 p
11
)
t
10
p
t
11
11
_
with
combined
asympt

2
2
. It follows that
combined
=
ind
+
POF
(Christoffersen 1998). This test, by construction, has a lower discrim-
inatory power than other tests because of the joint null hypotheses.
The test yields goodresults to differentiate betweenonaverage good
predictions (the POF test would not reject this) and timely dynamic
(ie, heteroscedastic processes) predictions.
At this point, we can derive a natural extension to the binomial
test by approximating the binomial distribution with a Poisson dis-
tribution with the estimator :=
T
t=1
O
t,
for the intensity and its
349
MODEL RISK
expected value = (1 ) T. The respective LRT statistics follow
from
Poisson
= 2 ln
_
(
k
/k!) exp()
(
k
/k!) exp( )
_
= 2 ln
_
((1 ) T)
T
t=1
O
t,
exp((1 ) T)
(
T
t=1
O
t,
)
T
t=1
O
t,
exp(
T
t=1
O
t,
)
_
The behaviour of the Poisson LRT is very similar to the POF
test, stemming from the fact that the POF test assumes the binomial
distribution and the Poisson LRT assumes the Poisson distribution.
Tests based on the entire predictive distribution
Tests based on exceedances can only give a rst impression of the
adequacy of the risk modelling due to the fact that they focus on
the main parameter of the distribution, ie, the respective quantile,
the value-at-risk. As mentioned above, the aim is not just to pre-
dict a certain (conditional) parameter but rather to predict a whole
conditional distributionof a portfolios gains andlosses. Hence, tests
focusingonthe whole distributionrelyingonTheorems 15.1and15.2
come to mind. These tests comprise goodness-of-t tests and LRTs
as well as others.
Starting with the fact, that F
t
(G
t
) = U
t
iid
U
[0,1]
when the right
distribution F
t
() is predicted, we can easily apply goodness-of-t
tests like the test by KolmogorovSmirnov and the statistic
KS
= sup
0x1
T
t=1
1
[0,x]
(U
t
)
T
x
= max
i=1,...,T
i
T
U
i:T
for the case of a U

[0,1]
-distribution. Crnkovic and Drachman (1996)
apply a certain version of the KolmogorovSmirnov test dating back
to the Kuiper statistic with
CD
= max
i=1,...,T
_
i
T
U
i:T
_
+ max
i=1,...,T
_
U
i:T

i
T
_
Whereas the KolmogorovSmirnov test is very sensitive around the
median of the distribution, the test by Crnkovic and Drachman
weights the entire distribution almost equally. Crnkovic and Drach-
manalso propose a weighting scheme (like w(x) =
1
2
ln(x(1x)))
to set more weights to the tails of the distribution with a statistic
max
i=1,...,T
_
w(U
i:T
)
_
i
T
U
i:T
__
+ max
i=1,...,T
_
w(U
i:T
)
_
U
i:T

i
T
__
350
The distribution of the KolmogorovSmirnov test statistic as well as
that of the Kuiper test statistic can be derived analytically. For the
case of considering a weighting scheme the respective distribution
will be derived by a numerical simulation. The disadvantage of the
test by Crnkovic and Drachman is that it has to take into account a
large number of realisations (1,000 or more).
Afurther well-knowngoodness-of-t test is the
2
goodness-of-t
test that is given for the preceding case of a U
[0,1]
-distribution by
2 =
m
_
i=1
__
T
_
t=1
1
i
(U
t
)
T
m
_
2
m
T
_
where
i
:=
_
i 1
m
,
i
m
_
for i = 1, . . . , m1
and
m
:=
_
m1
m
, 1
_
for a number of m classes and P(
i
) = 1/m. These statistics are
asymptotically
2
-distributed with
2
asympt

2
m1
.
Berkowitz (2001) proposes a modied test based on a transfor-
mation by the normal distribution and testing for autoregressive
(AR(1)) properties of thetransformedtimeseries Z
t
:=
1
(0,1)
(F
t
(G
t
)).
The likelihood function for a Gaussian noise in an AR(1) series
follows by
L
(Z
1
,...,Z
T
)
(, , ) =
(/(1),
2
/(1
2
))
(Z
1
)
T
t=2
+Z
t1
,
2 (Z
1
)
and thus the LRT follows by
NT
= 2 ln
_
L
(Z
1
,...,Z
T
)
(0, 1, 0)
L
(Z
1
,...,Z
T
)
( , , )
_
with the respective estimators , and .
6
This test by Berkowitz
with property
NT
asympt

2
3
only rejects deviations from the mean
and variance of a distribution; it does not identify other distribu-
tions with a mean of 0 and a variance of 1. Also, it cannot recognise
heteroscedasticity. Thus, several authors (see, for example, Dowd
(2004)) recommend applying a test on the hypothesis of a normal
distribution in addition to the test by Berkowitz.
To conclude, we provide a test for the serial independence prop-
erty of the standardised returns U
t
. Dening
i
as above, we dene
351
MODEL RISK
ij
:= 1
i
(U
t1
)1
i
(U
t
) for t = 2, . . . , T as the realisation of a stan-
dardised return rst in class i and then in class j. This consideration
yields, in conjunction with
i
:=

m
j=1
ij
and
j
:=

m
i=1
ij
, the
following
2
(m1)
2
-distributed test statistics
2
(ind2)
=
m
_
i=1
m
_
j=1
__
ij
i
T 1
_
2
T 1
i
_
Obviously, the test
2
(ind)
is a special case of
2
(ind2)
for only two
classes:
0
= [0, ) and
1
= [, 1].
EMBEDDINGVALIDATION AND BACKTESTING PROCEDURES
Regular validation and backtesting process
During the validation and backtesting process, the bank regularly
examines the adequacy and validity of the predicted risk values and
toa certainextent the predicted(conditional) distributionof the port-
folios gains and losses. If any doubts about the adequacy exist, we
can modify the model to better cope with the risks by re-examining
its construction, by introducing new risk factors, changing distri-
butional assumptions about the risk-factor returns or by a better
mapping between returns and portfolio gains and losses.
A systematically formulated backtesting and validation process
canbesketchedinseveral steps (seeFigure15.1). Thedenitionof the
model and the respective approach used to calculate the risk (step 1
in the gure) include the assumptions such as risk-factor selection,
distribution and transformation. This set-up is regularly updated
through the backtesting and validation of results. Step 2 is the pro-
cedure of backtesting itself; the risk model must be analysed using a
set of backtestingtechniques. Anex post analysis takes place instep3,
leading to potential modications in the risk model, and the poten-
tial impact on capital adequacy is judged. Also, new backtesting
methods must be considered from time to time (see step 4).
Inpractice, botha validationpart anda backtesting part shouldbe
maintained on a regular basis. By using the validation procedures,
we can improve the different building blocks individually, whereas
in a backtesting context the impact of the test results can become
more complicated. Wehn (2008) derives a decision tree to improve
the model gradually. In short, in some of the tests we observe a mis-
estimation(under- or overestimation) of risk(eg, the TLA, the
POF
or
KS
,
CD
or
2 ), which might be healed by inspecting the risk-factor

352
Figure 15.1 Embedding the results of backtesting in a regular
validation and backtesting process
Market risk
model and
assumptions
Risk measure
Observed P&L
Capital adequacy
demands & market
risk managment
Exceedences and
standardised gains
Backtesting methods for VaR assessment
VaR efficiency analysis, reselection
and sustainable development
Constant update and
upgrade of backtesting
methods in use
4
3
2
1
Adapted from Lehikoinen (2007).
distribution. Other tests focus on the serial information (eg,
ind
,
2
(ind)
or
2
(ind2)
). This might be caused by missing information
due to missing risk factors, a coarse modelling of time pattern or a
crude mapping between risk-factor returns and the portfolio gains
and losses. For a practical example applied to a portfolio, we refer
the reader to Wehn (2008).
353
MODEL RISK
Regulatory concerns
In addition to several qualitative standards, BCBS (2009) formu-
lates the current requirements for a regular validation process for
approved market risk models which requires, beyond the TLAtests,
a demonstration that the assumptions made (eg, distribution, pric-
ing models etc) are appropriate. The backtests should use hypothet-
ical changes in portfolio value with end-of-day positions to remain
unchanged, it should use different condence intervals tests on a
sub-portfolio level. Regulators, among others, require the use of
hypothetical portfolios toassess structural properties suchas insuf-
cient data histories, where a mapping to proxies must be done. These
procedures must be carried out on a periodical basis and especially
when there have been any signicant structural changes.
Market risk models approved by the regulatory authorities have
to meet capital requirements (BCBS 2009), which are determined
by a qualitative factor and a quantitative factor. The quantitative
factor is directlylinkedtothe backtestingresults fromthe TLA. Thus,
the regulatory capital requirements have a high dependence on the
adequacy of the model, as Figure 15.1 suggests.
DEDUCING MODEL RISK FROM BACKTESTING RESULTS
As discussed above, model errors in a market risk model can lead to
undesiredresults andhave great impact to the management process.
Thus, we should think about how to treat potential model risk in a
market risk context.
Bearing Theorem 15.1 in mind, we are motivated to start our
approach by using the standardisedreturns U
t
. Methods of this kind
are well establishedinthe context of weather forecasts. Dawid(1982)
introduces the concept of recalibration which, applied to our case,
means that if there exist and such that
U
t
= F
t
_
G
t
_
iid
U
[0,1]
we call U
t
recalibratable and we would be enabled to just scale the
risk measurement G
t
or the respective distributional function by
and . This idea can be extended to the case of an abstract transfor-
mation T, with property U
t
= T(U
t
) U
[0,1]
. If this transformation
exists, we can easily use the transformeddistribution to calculate the
risk gures. In other words, the difference between the U
t
and U
t
354
(eg, interms of economic capital) is a goodmeasure of the model risk
for the market risk model andis directly linkedto a way of healing
potential misleading effects.
In addition, the regulatory requirements by the BCBS (1996) intro-
duce the idea of a potential recalibration on the basis of the outliers.
Here, there is a direct link between the number of outliers occurred
and the regulatory capital requirements by a multiplication factor.
Other researchintroduces the ideaof anindividual utilityfunction
that reects a banks trade-off between low capital requirements on
the one hand and exact risk measurement methods on the other.
The eld for adjustments to the model is quite broad, but we can
easilydrawthe conclusionthat amelioratingthe model is muchmore
useful than healing potential malicious effects ex post.
CONCLUDING REMARKS
In this chapter we have illustrated and systemised the main mod-
elling assumptions associatedwith a market risk model. The process
for assessing the quality of the model is now twofold: the building
blocks assumptions can be validated individually and the models
output can be judged by means of backtesting. Embedded in a reg-
ular process, the results can help to communicate the market risk
models strengths and weaknesses with a view to steadily improv-
ing the adequacy of the model. This is a crucial step in every risk
measurement and management cycle.
The views expressed in this chapter are those of the author and
do not necessarily reect the views of his employers. None of
the methods described herein is claimed to be in actual use at
DekaBank.
1 We consider only the discrete case of t 1, 2, 3, . . . with a time step of one day.
2 For example, in the popular RiskMetrics model (a special univariate Garch(1,1) case with
exponential weighting scheme), it holds that R
t
= h
t
t
with a stochastic volatility of h
2
t
=
h
2
t1
+(1 )R
2
t1
(Mina and Xiao 2001).
3 Usually, is set to a high level, eg, 0.95 or 0.99.
4 Acomprehensive complementary discussion of backtesting techniques is given in Campbell
(2005).
5 Apotential extension of the TUFF test is given by Haas (2001), where in addition to the time
until the rst exceedance, t
1
:= mint O
t,
= 1. Also, the intervals between the individual
exceedances
t
i
:= min
_
t
t
_
j=1
O
j,
= i
_
t
i1
355
MODEL RISK
for i := 2, . . . ,
T
j=1
O
j,
are taken into account. Under independence, it holds that P(t = t
i
) =
(1
i
)
t
i
1
i
and the respective LRT follows by
TUFF+
= 2
T
j=1
O
j,
_
i=1
ln
_
(1 )
t
1
1
_
1
t
1
_
1
1
t
1
_
t
1
1
_
1
_
with
TUFF+
asympt

2
T
j=1
O
j,
Further on, this LRT can be combined with the proportion of failure test, leading to
mix
:=
TUFF+
+
POF
asympt

2
T
j=1
O
j,
+1
As theses statistical tests do not lead to signicant further or complementary insights, they
are not mentioned in the remainder of the chapter.
6 The test by Berkowitz can easily be extended to a test on autocorrelation by the following
statistics:
NT(uncorr)
= 2 ln
_
L
(Z
1
,...,Z
T
)
( , , 0)
L
(Z
1
,...,Z
T
)
( , , )
_
REFERENCES
BCBS, 1996, Supervisory Framework for the Use of Backtesting in Conjunction with
the Internal Models Approach to Market Risk Capital Requirements, Basel Committee
on Banking Supervision, URL: http://www.bis.org.
BCBS, 2009, Revisions to the Basel II Market Risk Framework, Final Paper, Basel
Committee on Banking Supervision, http://www.bis.org.
Berkowitz, J., 2001, Testing Density Forecasts, with Applications to Risk Management,
Journal of Business & Economic Statistics 19(4), pp. 46574.
Campbell, S., 2005, A Review of Backtesting and Backtesting Procedures, Working
Paper, Federal Reserve Board, Washington, DC.
Christoffersen, P., 1998, Evaluating Interval Forecasts, International Economic Review
39(4), pp. 84162.
Crnkovic, C., and J. Drachman, 1996, Quality Control, Risk Magazine 9(9), pp. 13843.
Dawid, A. P., 1982, The Well-Calibrated Bayesian, Journal of the American Statistical
Association 77, pp. 60513.
Diebold, F. X., T. A. Gunter and A. S. Tay, 1998, Evaluating Density Forecasts with
Applications to Financial Risk Management, International Economic Review39(4), pp. 863
83.
Dowd, K., 2004, AModied Berkowitz Back-Test, Risk Magazine 17(4), p. 86.
Finger, Ch., 2005, Back to Backtesting, RiskMetrics Monthly Research, May.
Haas, M., 2001, New Methods in Backtesting, Working Paper, Center for Advanced
European Studies and Research.
Kupiec, P. H., 1995, Techniques for Verifying the Accuracy of Risk Measurement Models,
Journal of Derivatives 2, pp. 7384.
Lehikoinen, K., 2007, Development of Systematic Backtesting Processes of Value-at-
Risk, Masters Thesis, Helsinki University of Technology.
356
Mina, J., and J. Y. Xiao, 2001, Return to RiskMetrics: The Evolution of a Standard,
Technical Document, URL: http://www.riskmetrics.com.
Rosenblatt, M., 1952, Remarks on a Multivariate Transformation, Annals of Mathematical
Statistics 23, pp. 4702.
Stahl, G., C. S. Wehn and A. Zapp, 2006, Backtesting Within the Trading Book, The
Journal of Risk 8(2), pp. 116.
Wehn, C. S., 2005, Anstze zur Validierung von Marktrisikomodellen: Systematisierung,
Anwendungsmglichkeiten und Grenzen der Verfahren (Aachen: Shaker).
Wehn, C. S., 2008, Looking Forward to Back Testing, Risk Magazine, May.
Wehn, C. S., 2010, Evaluating the Adequacy of Market Risk Models, in G. N. Gregoriou,
C. Hoppe and C. S. Wehn (eds), Model Risk Evaluation Handbook (McGraw-Hill).
Wilks, S. S., 1938, The Large-Sample Distribution of the Likelihood Ratio for Testing
Composite Hypotheses, Annals of Mathematical Statistics 9, pp. 602.
357
16
Estimation of Operational
Value-at-Risk in the Presence of
Minimum Collection Threshold:
An Empirical Study
Anna Chernobai; Christian Menn;
Svetlozar T. Rachev; StefanTrck
Syracuse University; DZ Bank AG; Universitt Karlsruhe,
Finanalytica Inc, University of California at Santa Barbara;
Macquarie University
The Basel II Capital Accord, nalised in July 2009, requires banks
to adopt a procedure to estimate the operational risk capital charge.
In particular, the advanced measurement approaches that are cur-
rently mandated for all large internationally active US banks require
the use of historic operational loss data. Operational loss databases
are typically subject to a minimum recording threshold of roughly
US$10,000. We demonstrate that ignoring such thresholds, and thus
choosing an inappropriate model, leads to biases in correspond-
ing parameter estimates. Using publicly available operational loss
data, we analyse the effects of model misspecication on resulting
expected loss, value-at-risk, and conditional value-at-risk gures
and show that underestimation of the regulatory capital is a con-
sequence of such model error. The choice of an adequate loss distri-
bution is conducted via in-sample goodness-of-t procedures and
backtesting, using both classical and robust methodologies.
INTRODUCTION
The three major sources of risk in nancial institutions are market
risk, credit risk and operational risk. While the rst two are well
understood, the research on operational risk is still a growing area.
The scope of risks grouped under operational risk is quite large: the
359
MODEL RISK
Basel Capital Accorddenes operational riskas the riskof loss result-
ing frominadequate or failedinternal processes, people andsystems
or fromexternal events (BCBS 2006a). Examples of large operational
losses due to unauthorised trading, fraud and human error include
Orange County (US, 1994), Barings (Singapore, 1995), Daiwa (Japan,
1995), Socit Gnrale (France, 2008); losses due to natural disas-
ters include those due to hurricanes Andrew and Katrina (US, 1992
and 2005); losses were also caused by the terrorist attack of Septem-
ber 11, 2001. Arguably, while the 20089 nancial crisis is commonly
credited to credit risk, many of its roots can be traced to operational
risk. Specically, the failings of prominent mortgage and nancial
service companies could have been averted had lending practices
been founded upon forward-looking market expectations based on
fundamentals rather thanshort-termmarket movements. As aresult,
years of improper lendingpractices ledtothe mortgage crisis around
2007 and subsequent bailout of a series of US nancial institutions
by the US government. The crisis brought to light the grave con-
sequences of inadequate business practices and model errors: yet
another type of operational risk.
The signicance of operational risk as a major contributor to
banks andinsurance companies riskpositions is increasinglyrecog-
nised by the industry and regulators. Current estimates suggest that
the allocation of total nancial risk of a bank is roughly 60%to credit,
15% to market and liquidity and 25% to operational risk (Jorion
2000).
1
Under the Basel II Capital Accord (BCBS 2001a, 2006a), each
bank is required to adopt a methodology to determine the opera-
tional risk capital charge to account for unexpected losses. The US
banks are mandated to use the advanced measurement approaches.
Under the loss distribution approach (LDA), one of the suggested
approaches, banks compute separately the loss severity and fre-
quency distribution functions for each business line and risk-type
combination, over a one-year period. The total capital charge is then
determinedas the sumof one year value-at-risk(VaR) measures with
the condence level 1 (eg, = 0. 1%), across all combinations,
2
based on the compounded losses.
A number of issues remain in modelling operational risk. One
problem is related to internal operational loss data. Data record-
ing is a subject to lower recording thresholds, which for internal
databases are set at roughly US$10,000 (BCBS 2003).
3
We refer to
360
OPERATIONAL VALUE-AT-RISK INTHE PRESENCE OF MINIMUM COLLECTIONTHRESHOLD
such data that is not recorded as non-randomly missing data, and
refer to the recorded data as left-truncated and incomplete. If such
datatruncationis unaccountedfor, model errors cancarrysignicant
material consequences.
The choice of loss collection thresholds can signicantly affect the
calculation of expected loss and, to some extent, the shape of the
estimated-loss distribution and estimates of unexpected loss.
BCBS (2006b)
Recent (2009) guidelines by the Basel Committee make a clear
recommendation that data truncation be adequately accounted for.
Chernobai et al (2006b) showed that, if the truncation is ignored,
ttingunconditional distributiontotheobserved(incomplete) losses
wouldleadtobiasedestimates of the parameters of bothseverityand
frequency distributions. The magnitude of the effect is dependent
on the threshold level and the underlying loss distribution. Under
the compound Poisson process model, the severity and frequency
distributions of the operational risk are inter-related: while severity
of recorded data is biased toward higher losses, for as long as the
fraction of missing data (under the truncation point) is estimated
to be non-zero, the frequency parameter(s) require a proportional
increase. As a result, the resulting VaR measure would be under-
estimated. Several studies examined this issue further, including
Shevchenko and Temnov (2007, 2009) and Crama et al (2007). The
former two studies provided a theoretical extension to the model of
Chernobai et al (2006b) and the latter study investigated the effects
of data truncation on the loss distribution using real operational
loss data while ignoring the impacts of such a truncation on the fre-
quency distribution. In this chapter, we ll this void and attempt
to answer the following question: how substantial is the effect of
ignoring minimum collection threshold on the operational risk cap-
ital reserves? Namely, if the consequences are severe, then indeed
this issue warrants special attention on the part of regulators and
quantitative analysts. Specically, we extend the theoretical frame-
work of Chernobai et al (2006b) to apply it to real operational loss
data and test empirically the implications of the model error associ-
ated with misspecied severity and frequency distributions and the
magnitude of this misspecication on the estimates of operational
risk regulatory capital.
4
361
MODEL RISK
The aim of this chapter is twofold. We analyse the effects of miss-
ingdata onloss severityandfrequencydistributions, andthenexam-
ine the impact of model misspecication on the operational risk
capital charge, determined by two alternatives: the VaR and con-
ditional VaR (CVaR) measures. The chapter is organised as follows.
In the next section we explain the truncation problem and discuss
the methodology for the correct estimation of the severity distribu-
tionandthe necessaryadjustment tothe frequencyof loss events. We
then (see page 369) present the results of an empirical study using
publicly available operational loss data for 19802002 and exam-
ine the effects of misspecied and correctly dened distributions on
the capital charge, before carrying out goodness-of-t tests to deter-
mine an optimal lawfor the loss severity. We showthat ignoring the
missing data leads to misleading (underestimated) VaR estimates.
The following section (see page 383) provides a robustness check
in which we estimate the operational risk regulatory capital using
the principle of robust statistics. The nal section (see page 387)
concludes and states nal remarks.
COMPOUND MODEL FOR OPERATIONAL RISK
Compound Poisson process model
Following the recommendation by the Basel Committee, we assume
that the aggregated operational losses follow a stochastic process
S
t
t0
over the time interval t expressedbythe followingequation
S
t
=
N
t
_
k=0
X
k
, X
k
iid
F
(16.1)
in which the loss magnitudes are described by the randomsequence
X
k
assumed to follow the cumulative distribution function (CDF)
F
that belongs to a parametric family of continuous probability

distributions, the density is denoted by f
and the counting process

N
t
is assumed to take the form of a homogeneous Poisson process
(HPP) with intensity > 0 or a non-homogeneous Poisson process
(NHPP) with intensity (t) > 0
5
. Depending on the distribution,
is a parameter vector or a scalar. For simplicity, we shall refer to it
as a parameter throughout the chapter. We assume that the distri-
bution family is sufciently well behaved that can be estimated
consistently by the maximumlikelihood estimation (MLE). To avoid
362
the possibility of negative losses we restrict the support of the dis-
tribution to the positive half-line R
>0
. Equation 16.1 assumes inde-
pendence between frequency and severity distributions. The CDF
of the compound Poisson process is given by
P(S
t
s) =
_
_
n=1
P(N
t
= n)F
n
(s), s > 0
P(N
t
= 0), s = 0
(16.2)
where F
n
denotes the n-fold convolution with itself.

In practice, the model in Equation 16.1 can be used to determine
the required capital charge imposed by regulators. It is measured
as the (1 )th quantile of the cumulative loss distribution (Equa-
tion 16.2) over a one-year period, which denes VaR. VaR
t,1
, for
the tolerated risk level and the time interval of length t (gener-
ally = 1%5% and t is one year), is dened as the solution of the
equation
P(S
t+t
S
t
> VaR
t,1
) = (16.3)
and the CVaR (also called expected tail loss (ETL) or expected
shortfall (ES)) is dened by
CVaR
t,1
:= E[S
t+t
S
t
S
t+t
S
t
> VaR
t,1
]
=
E[S
t+t
S
t
; S
t+t
S
t
> VaR
t,1
]
(16.4)
Given a sample x = (x
1
, x
2
, . . . , x
n
), containing n losses which have
occurred during some time interval t = T
2
T
1
, under the imposed
assumptions on the structure of F
, the task of estimating and

canbe performedwiththe MLEprinciple (or, inthe case of anNHPP,
(t) is estimated by directly tting a deterministic function)
MLE
(x) =
n
t
and
MLE
(x) = arg max
n
_
k=1
log f
(x
k
) (16.5)
The task of operational-loss data analysis is complicatedby the pres-
ence of missing data that fall to the left of the left truncation point
(minimum collection threshold). The estimates in Equation 16.5
would be misleading in the presence of truncation. The question
addressed in subsequent analysis is whether ignoring the missing
data has a signicant impact on the estimation of the frequency
parameter ((t)) and the severity parameter . From a statistical
363
MODEL RISK
viewpoint, ignoringnon-randomlymissingdatawouldleadtoabias
in all estimates. However, in practical applications, a possible rea-
son why such thresholds are ignored would be that, since the major
bulk of losses is in excess of the threshold, the small losses cannot
have a signicant impact on the operational VaR that is determined
by the upper quantiles of the loss distribution. This chapter presents
empirical evidence to disprove this argument. In the following sec-
tionwe reviewthe methodologyfor consistent estimationof loss and
frequency distributions, as suggested by Chernobai et al (2006b).
Unbiased estimation of severity and frequency distributions
In the presence of missing data, we conclude that the observed oper-
ational losses followa truncated compound Poisson process. We fol-
low similar notation to that in Chernobai et al (2006b). The available
dataset collected in the time interval [T
1
, T
2
] is incomplete, due to
the non-negative pre-specied thresholds u that dene a partition
on R
>0
through the events A
1
= (0, u) and A
2
= [u, ). Realisa-
tions of the loss distribution belonging to A
1
will not enter the data
sample; neither the frequency nor the severity of losses below u is
recorded(missingdata). Realisations inA
2
are fullyreported, ie, both
the frequency andthe loss amount are specied(observeddata). The
observed sample is of the form z = (n, x), where n is the number of
observations in A
2
and x are the values of these concrete observa-
tions. Given that the total number of observations in the complete
sample is unknown, one possible joint density of z (with respect to
the product of counting and Lebesgue measures) consistent with
the model specication in Equation 16.1 is given by the following
expression
g
,
(z) =
(t
)
n
n!
e
t
k=1
f
(x
k
)
q
,2
(16.6)
where t is replaced with (t) for an NHPP and q
,j
denotes the
probability for a random realisation to fall into set A
j
, j = 1, 2, the
observed intensity is given by

:= q
,2
and t := T
2
T
1
is the
length of the sample window. In Equation 16.6, the Poisson pro-
cess

N
t
of intensity

(or

(t)) that counts only the observed losses
exceedinguinmagnitude canthus be interpretedas a thinningof the
original process N
t
of intensity ((t)) that counts all events in the
complete data sample. The maximisation of the corresponding log-
likelihood function with respect to (for the HPP case) and can be
364
divided into two separate maximisation problems, each depending
on only one parameter

MLE
= arg max
log g
(z) = arg max
log
_
n
k=1
f
(x
k
)
q
,2
_
(16.7)
MLE
= arg max
log g
,
MLE
(z) =
n
tq
MLE
,2
(16.8)
The MLE of the unknown parameter can be done in two ways:
by performing direct numerical integration or by using the two-step
expectation-maximisation algorithm, developed by Dempster et al
(1977). The expectation-maximisation algorithm has been used in a
variety of applications such as probability density mixture models,
hidden Markov models, cluster analysis, factor analysis and sur-
vival analysis. References include McLachlan and Krishnan (1997),
MengandvanDyk(1997), WulfsohnandTsiatis (1997), DeCanioand
Watkins (1998), among many others and, in the framework of the
operational risk modelling, Chernobai et al (2006b) and Bee (2005).
Implications of data misspecication on the operational risk
capital charge
The Basel Capital Accord requires banks to provide operational risk
capital charges that cover the unexpected losses. At the same time
they suggest using VaR for computing the capital charge. Some con-
fusion arises from such a denition of the capital charge, because
providing the capital charge for the unexpected losses would mean
that the expected aggregated loss (EL) has to be subtracted from
VaR. We therefore analyse the impact of data misspecication on
all relevant components: aggregated expected loss, VaR and also
CVaR.
6
For a compound Poisson process, the aggregated expected loss
is computed as a product of the expected frequency and loss
distributions
ES
t
= EN
t
EX (16.9)
(VaR
t,1
was previously dened in Equation 16.3.) We x a toler-
ated risk level , a time horizon of length t (the Basel Committee
suggests using t = 1 year (BCBS 2001b)) and x , eg, at 0.1%. By
denition, VaR equals the capital charge that must be maintained
in order to protect against potential operational losses in t in the
365
MODEL RISK
future that can occur with probability (1 ). Generally, no closed-
formexpressionfor the cumulative loss distributionis available. The
upper quantiles have to be determinednumericallythroughapprox-
imations such as the recursive PanjerEuler scheme, Fast Fourier
Transforminversion of the characteristic function or simulation (we
use the Monte Carlo method in this chapter). For the special case of a
sub-exponential loss distributions F :, such as lognormal, Pareto
and the heavy-tailed Weibull, relevant in the context of operational
risk modelling, the tail of the compound process is approximated by
(Embrechts et al 1997)
P(S
t
> s) EN
t
P(X > s), s (16.10)
For an example when the losses X follow a lognormal(, ) dis-
tribution, combining Equations 16.8 and 16.7 with Equations 16.3
and 16.10 results in the following expected aggregated loss and
approximate VaR estimates
ES
t
=

MLE
t exp
_

MLE
+

2
MLE
2
_
(16.11)
VaR
t,1
exp
_

MLE
+
MLE

1
_
1

MLE
t
__
(16.12)
for the HPP case, with

MLE
t replaced by

(t) for the NHPP
case, where () and () denote the density and the distribution
function, respectively, and
1
() denotes the quantile of a standard
normal distribution. Aclosed-form expression for CVaR exists only
for Gaussianloss severity. If the missingdatais ignoredinthe estima-
tion of loss severity and frequency, then for the lognormal example
the bias of the parameters can be expressed analytically as
E
observed
=
_
1
_
log u
__
bias(
observed
) =
_
log u
_
< 0
_
_
(16.13)
E
observed
= E
_
1
n
_
log X
k
X
k
> u
_
= +
((log u )/)
1 ((log u )/)
bias(
observed
) =
((log u )/)
1 ((log u )/)
> 0
_
_
(16.14)
366
and
E
2
observed
= E
_
1
n
_
log
2
X
k

2
observed
X
k
> u
_
=
2
_
1 +
log u
((log u )/)
1 ((log u )/)
_
((log u )/)
1 ((log u )/)
_
2
_
bias(
2
observed
) =
2
_
log u
((log u )/)
1 ((log u )/)
_
((log u )/)
1 ((log u )/)
_
2
_
< 0 since log u is small
_
_
(16.15)
where, in Equation 16.13
1
, is replaced by (t) for the NHPP case.
Figure 16.1 gives an illustration of the biases of the three parameters
for a wide range of initial (complete-data) true values of and. The
distances between the ratios of the estimated parameters to the true
parameters represent the relative biases for each case. For this exam-
ple, a threshold level of H = 50 in nominal value was considered,
which corresponds to the cut-off levels given in Table 16.1.
For the same example, Figure 16.2 demonstrates the biases (rep-
resented by the ratios) of the estimated fractions of missing data in
the naive and conditional scenarios. The fraction being equal to 1
indicates the absence of bias.
Combining Equations 16.12 and replacing the estimates for and
by their expectations from Equations 16.14 and 16.15, we obtain
an approximate estimate of expectedaggregatedloss andVaRunder
the data misspecication
ES
t
= ( +bias(
obs
))
exp +bias(
obs
) +
1
2
( +bias(
obs
))
2
< true ES
t
(16.16)
VaR
t,1
exp
_
+bias(
obs
) +( +bias(
obs
))
1
_
1

( +bias(
obs
))t
__
< true VaR
t,1
(16.17)
367
MODEL RISK
Figure 16.1 Ratios of estimated parameters to the true (complete-data)
parameter values, for the lognormal example, u = 50
1.5
2.0
2.5
4
5
6
0.6
0.8
1.0
1.2
1.4
n
a
i
v
e

/

0
1.5
2.0
2.5
4
5
6
0.6
0.8
1.0
1.2
1.4
1.5
2.0
2.5
4
5
6
0.4
0.6
0.8
1.0
1.5
2.0
2.5
4
5
6
0.4
0.6
0.8
1.0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
0.2
0.4
0.6
0.8
1.0
1.2
Fraction of missing data
0
/
0
obs
/
0
(a) (b)
(c) (d)
(e)
c
o
n
d

/

0
n
a
i
v
e

/

0
c
o
n
d

/

0
o
b
s

/

0
(a) , naive; (b) , conditional; (c) , naive; (d) , conditional; (e) .
with appropriate adjustments for an NHPP case. The direction of
the last inequality (it also holds for CVaR) generally depends on the
threshold u and the underlying distribution. For practical purposes
in the context of operational risk, the < inequality is valid in gen-
eral cases (Chernobai et al 2006b). Figure 16.3 illustrates the biases
(represented by the ratios) of the expected aggregated loss, VaR and
CVaR gures, under the naive and conditional scenarios, for the
= 100 example. We note that the value of has no effect on the
ratio of the expected loss (this follows directly from Equation 16.16)
368
Table 16.1 Fraction of missing data, F
0
(u), for the lognormal(
0
,
0
)
example with nominal threshold of u = 50
0

0
=4
0
=5
0
=6.5
1.5 0.48 0.23 0.04
2 0.48 0.29 0.10
2.7 0.49 0.34 0.17
Figure 16.2 Ratios of estimated fraction of missing data (Q) to the true
(complete-data) fraction, for the lognormal example, u = 50
1.5
2.0
2.5
4
5
6
1.5
2.0
2.5
4
5
6
0
Q
n
a
i
v
e

/

Q
0
Q
c
o
n
d

/

Q
0
0
0.2
0.4
0.6
0.8
1.0
1.2
0
0.2
0.4
0.6
0.8
1.0
1.2
(a) (b)
(a) F(u), naive; (b) F(u), conditional.
and increasing to 150, 200 or more has a very negligible impact
(bias increases) on the biases of VaR and CVaR (gures are omitted
here).
APPLICATIONTO OPERATIONAL RISK DATA
Purpose of study and data description
In this section we apply the model to real operational risk data,
obtained from the FIRST database of Zurich IC Squared (IC
2
),
an independent consulting subsidiary of Zurich Financial Services
Group. The external database is comprisedof operational loss events
throughout the world. The original loss data covers losses in the
period 19502002. Afewrecorded data points were belowUS$1 mil-
lion in nominal value, so we excluded these from the analysis in
order to make it more consistent with the conventional threshold
(US$1 million) for external databases. Furthermore, we excluded
the observations before 1980 because relatively few data points
were available (most likely due to poor data-recording practices).
The nal dataset for the analysis covered losses in US dollars for
the time period between 1980 and 2002. It consists of ve types
369
MODEL RISK
Figure 16.3 Ratios of estimated one-year EL, 95% VaR and 95%
CVaR to the true (complete-data) values, for the lognormal example,
u = 50, = 100
1.5
2.0
2.5
4.0
4.5
5.0
5.5
6.0
6.5
1.5
2.0
2.5
4.0
4.5
5.0
5.5
6.0
6.5
1.5
2.0
2.5
4.0
4.5
5.0
5.5
6.0
6.5
1.5
2.0
2.5
4.0
4.5
5.0
5.5
6.0
6.5
1.5
2.0
2.5
4.0
4.5
5.0
5.5
6.0
6.5
1.5
2.0
2.5
4.0
4.5
5.0
5.5
6.0
6.5
0
0.5
1.0
1.5
0
0.5
1.0
1.5
0
0.5
1.0
1.5
0
0.5
1.0
1.5
0
0.5
1.0
1.5
0
0.5
1.0
1.5
(a) (b)
(c) (d)
0
0
0
0
0
(e) (f)
0
0
0
E
L
n
a
i
v
e

/

E
L
0
E
L
c
o
n
d

/

E
L
0
C
V
a
R
n
a
i
v
e

/

C
V
a
R
0
C
V
a
R
c
o
n
d

/

C
V
a
R
0
V
a
R
n
a
i
v
e

/

V
a
R
0
V
a
R
c
o
n
d

/

V
a
R
0
(a) EL, naive; (b) EL, conditional; (c) VaR
0.95
, naive; (d) VaR
0.95
, conditional;
(e) CVaR
0.95
, naive, (f) CVaR
0.95
, conditional.
of loss: relationship (such as events related to legal issues, neg-
ligence and sales-related fraud), human (such as events related
to employee errors, physical injury and internal fraud), processes
(such as events related to business errors, supervision, security and
transactions), technology (such as events related to technology
andcomputer failure andtelecommunications) andexternal (such
as events related to natural and man-made disasters and external
fraud). The loss amounts have been adjusted for ination using the
370
Consumer Price Index from the US Department of Labor. The num-
bers of data points n of each type are 849, 813, 325, 67 and 233,
respectively.
Note that, since the dataset is external, the estimates of the param-
eters and VaR and CVaR values are not applicable to any particular
bank. The purpose of the empirical study is to apply the model pro-
posed earlier (see page 362) and demonstrate the results, and we
recommend risk managers to apply the technique to their internal
databases.
In the empirical study we focus on two scenarios. The rst sce-
nario we refer to as a naive approach, in which no adjustments
for the missing data are made to the data. The second scenario is the
rened approach, in which the losses are modelled with truncated
(conditional) distributions, given that the losses are larger than or
equal to US$1 million, the MLE estimates are obtained according
to Equation 16.8, and the frequency functions parameters of the
Poisson counting process are adjusted according to Equation 16.7.
Operational frequency distributions
We consider two types of Poisson process, homogeneous and non-
homogeneous, withcumulative intensityt for the HPPandcumu-
lative intensity (t) for the NHPP. For this particular dataset, visual
inspection (Figures 16.4 and 16.6) of the annually aggregated num-
ber of losses suggests that the accumulation is somewhat similar to a
CDF-like process. We hence consider the two following tted cubic
functions for the NHPP, each with four parameters.
Cubic I: a Lognormal CDF-like process of the form
(t) = a +
1
2c
_
b exp
_
(log t d)
2
2c
2
__
Cubic II: a log-Weibull CDF-like process of the form (t) =
a b expc log
d
t.
We obtainthe four parameters a, b, c, d sothat the mean-squarederror
(MSE) is minimised(minimising mean absolute error (MAE) instead
often led to higher error estimates). For the HPP, the estimate for
is obtained by simply averaging the annual total number of opera-
tional loss events. Other deterministic functions were tried for the
cumulative intensity (sinusoidal, tangent, etc), but did not result in a
goodt. Table 16.2 shows the estimatedparameters andthe MSEand
371
MODEL RISK
Table 16.2 Fitted frequency functions to the external"-type losses
Process
Parameter
, ..
estimates Cubic I Cubic II Poisson
a 2.02 237.88
b 305.91 236.30
c 0.53 0.00026
d 3.21 8.27
10.13
MSE 16.02 14.56 947.32
MAE 2.708 2.713 24.67
Figure 16.4 Annual accumulated number of external operational
losses, with tted cubic and Poisson models
5 10 15 20
0
50
100
150
200
250
300
Year
C
u
m
u
l
a
t
i
v
e

n
u
m
b
e
r

o
f

l
o
s
s
e
s
Actual
Cubic I
Cubic II
Poisson
MAE for both the cubic cumulative intensities and a simple homo-
geneous Poisson process with a constant intensity factor. Figure 16.4
shows the three ts plottedtogether withthe actual aggregatednum-
ber of events. The cubic ts appear to be superior to the standard
Poisson, as illustrated by Figures 16.4 and 16.6, and conrmed by
the MSE and MAE comparison from Tables 16.2 and 16.12.
Inthe subsequent analysis, we will assume the deterministic cubic
(I or II) forms for the operational loss frequency distributions, and
will no longer consider the HPP case.
372
Operational loss distributions
We restrict our attention to the loss distributions that can be used to
model the losses that lie on the positive real half-line. The following
distributions for loss severity are considered in the study.
Exponential: denoted by Exp()
f
X
(x) = e
x
, x 0, > 0
Lognormal: denoted by 1J(, )
f
X
(x) =
1
2x
exp
_
(log x )
2
2
2
_
, x 0, , > 0
Gamma: denoted by Gam(, )
f
X
(x) =

x
1
()
expx , x 0, , > 0
Weibull: denoted by Weib(, )
f
X
(x) = x
1
expx
, x 0, , > 0
Log-Weibull: denoted by log-Weib(, )
f
X
(x) =
1
x
(log x)
1
exp(log x)
, x 0, , > 0
Generalised: denoted by 11(, )
f
X
(x) =
1
(1 +x
1
)
(1+1/)
Pareto: x 0, > 0
Burr: denoted by Burr(, , )
f
X
(x) =
x
1
( +x
)
(+1)
, x 0, , , > 0
log- stable: denoted by log :
(, , )
f
X
(x) =
g(ln x)
x
,
g :
(, , ), x > 0, (0, 2), [1, 1], , > 0

(no closed-form density).
Symmetric stable: denoted by :
:()
f
Y
(y) = g(y), g :
(0, , 0), x = y, (0, 2), > 0

(no closed-form density).
373
MODEL RISK
Table 16.3 Estimated and F

operational loss data
, F
(u) Naive Conditional

Exp 9.675610
9
9.770110
9
F
(u) 0.0096 0.0097

1J 16.5789 15.7125
1.7872 2.3639
F
(u) 0.0610 0.2111

Gam 0.3574 1.539210
6
3.458510
9
1.657110
9
F
(u) 0.1480 1
Weib 1.161310
4
0.0108
0.5175 0.2933
F
(u) 0.1375 0.4629

log-Weib 3.193310
12
2.816910
8
9.2660 6.2307
F
(u) 0.1111 0.3016

11 1.2481 1.5352
1.258810
7
0.706010
7
F
(u) 0.0730 0.1203

Burr 0.0987 0.1284
2.509810
26
3.249710
20
4.2672 3.3263
F
(u) 0.0145 0.0311

log :
1.8545 1.3313
1 1
1.1975 2.7031
16.6536 10.1928
F
(u) 0.0331 0.9226

:
: 0.6820 0.5905
1.139510
7
0.707310
7
F
(u) 0.0715 0.1283

Table 16.3 demonstrates the parameter values of the distri-
butions tted to the external dataset and the estimated frac-
tion of the missing data F
(u), under the naive and the correct,

conditional, approaches. The results for the remaining four loss
types are presented in the appendix. The tables demonstrate that,
under the truncated t, more weight is put on the lower magni-
tude losses, including the missing losses, than is predicted by the
naivemodel, as indicatedbytheF
(u) estimates. The fractionindi-

cates the true information loss due to data misspecication. The
374
Figure 16.5 Upper quantiles of tted truncated loss distributions to the
external-type losses, together with the empirical distribution
0 1 2 3 4 5 6
x 10
9
0.95
0.96
0.97
0.98
0.99
1.00
Data
C
D
F
Empirical
Exponential
Lognormal
Weibull
log Weibull
GPD
Burr
log Stable
Symm. stable
location parameters (if relevant) are decreased, the scale parameters
are increased and the shape parameters (if relevant) are decreased
under the correct model (for the GPDdistribution, the shape param-
eter corresponds to 1/) in most cases. Furthermore, the change
in the skewness parameter of the log- stable law from 1 to 1
for the external, relationship and human types indicates that
the right tail of the loss distribution under the correct model has a
near-exponential decay, comparable to that of the lognormal.
Based on the estimation fraction of missing data, we agree to
exclude the gamma distribution from further analysis. The propor-
tion of observed data being nearly zero is highly unrealistic.
Goodness-of-t tests for operational loss distributions
Visual tests
Figure 16.5 illustrates the upper quantiles of the considered condi-
tional distributions (except for Gamma) plottedagainst the empirical
distribution, for the external-type losses. The remaining four cases
are given in the appendix (see Figure 16.7).
375
MODEL RISK
Table 16.4 Results of in-sample GOF tests for external-type
operational losses
KS V AD AD
2
W
2
Exp 6.5941 6.9881 4.410
6
128.35 17.4226
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
1J 0.6504 1.2144 2.1702 0.5816 0.0745
[0.326] [0.266] [0.469] [0.120] [0.210]
Weib 0.4752 0.9498 2.4314 0.3470 0.0337
[0.852] [0.726] [0.384] [0.519] [0.781]
log-Weib 0.6893 1.1020 2.2267 0.4711 0.0563
[0.296] [0.476] [0.481] [0.338] [0.458]
11 0.9708 1.8814 2.7742 1.7091 0.2431
[0.009] [<0.005] [0.284] [<0.005] [<0.005]
Burr 1.3266 2.0385 2.8775 2.8954 0.5137
[0.050] [0.048] [0.328] [0.048] [0.048]
log :
7.3275 7.4089 37.4863 194.74 24.3662

[0.396] [0.458] [0.218] [0.284] [0.366]
S
S 0.7222 1.4305 1.110

5
1.7804 0.1348
[0.586] [0.339] [0.990] [0.980] [0.265]
p-values (in square brackets) were obtained via 1,000 Monte Carlo
simulations.
Around the 95th quantile, the lognormal distribution suggests a
good t for the external-type losses. Overall, Weibull, log-Weibull
and lognormal appear to be close to the empirical distribution
function. For relationship-type, lognormal and log-stable appear
to be the best, for human-type lognormal, log-stable, Weibull
and log-Weibull are the best, for process-type Burr, Weibull and
log-Weibull are the best, and for technology-type Weibull and
log-Weibull are the best, respectively.
EDF goodness-of-t tests
We test a composite hypothesis that the empirical distribution func-
tion (EDF) belongs to an entire family of hypothesised truncated
distributions. Since we estimate the parameters via MLE, we do not
specify the parameter values in the null expression. The null and
alternative hypotheses are summarised as
H
0
: F
n
(x)

F(x), H
A
: F
n
(x)

F(x), (16.18)
376
Table 16.5 Estimates of expected aggregated loss, VaR and CVaR for
external-type losses
EL VaR
0.95
VaR
0.99
CVaR
0.95
CVaR
0.99
Exp N 0.0207 0.0618 0.0897 0.0790 0.1064
C 0.0306 0.0798 0.1100 0.0985 0.1283
1J N 0.0157 0.0613 0.1697 0.1450 0.3451
C 0.0327 0.1126 0.4257 0.3962 1.1617
Weib N 0.0151 0.0613 0.1190 0.0975 0.1628
C 0.0208 0.0885 0.2494 0.2025 0.4509
log-Weib N 0.0611 0.1309 0.1059 0.1940
C 0.0839 0.2489 0.2046 0.4909
11 N 0.1190 0.8381 2.7082 12.4017
C 0.2562 2.6514 63.4969 314.41
Burr N 0.4072 8.7417 366.32 1823.79
C 0.7165 15.8905 1502.42 7498.81
log :
N 0.1054 3.7687
C 0.3879 0.8064 0.6750 1.2641
:
: N 0.1730 1.8319 35.7423 176.64

C 0.4714 7.6647 206.49 1025.16
N, naive; C, conditional. Figures must be further scaled by 10
10
. Figures
are based on 50,000 Monte Carlo samples.
where F
n
(x) is the EDF, and

F(x) is the tted distribution function
for this sample. For the naive scenario,

F(x) is the naively tted
distribution function with unconditional parameters. After neces-
sary adjustments for the missing data,

F(x) for the truncated sample
is dened as
F(x) =
_
c (x)

F
c (H)
1

F
c (H)
, x H
0, x < H
(16.19)
where
c
refers to the parameter(s) of the complete loss distribution.
We consider seven statistics for the measure of the distance between
the empirical and hypothesised distribution function: Kolmogorov
Smirnov (D), Kuiper (V), supremumAndersonDarling (A), supre-
mum upper tail AndersonDarling (A
up
), quadratic Anderson
Darling (A
2
), quadratic upper tail AndersonDarling (A
2
up
) and
377
MODEL RISK
Cramrvon Mises (W
2
), computed as
D = maxD
+
, D
V = D
+
+D
A =
nsup
x
F
n
(x)

F(x)
_
F(x)(1

F(x))
A
up
=
nsup
x
F
n
(x)

F(x)
1

F(x)
A
2
= n
_
(F
n
(x)

F(x))
2
F(x)(1

F(x))
d
F(x)
A
2
up
= n
_
(F
n
(x)

F(x))
2
(1

F(x))
2
d
F(x)
W
2
= n
_
(F
n
(x)

F(x))
2
d
F(x)
where
D
+
=
nsup
x
F
n
(x)

F(x) and D
nsup
x
F(x) F
n
(x)
and

F(x) is dened in Equation 16.19. Note that the supremumclass
statistics are multiplied by
n and the quadratic class by n, to make

them comparable across samples of different size. The limiting dis-
tributions of the test statistics are not parameter-free, so the p-values
and the critical values were obtained with Monte Carlo simulations
(Ross 2001). The A
up
and A
2
up
statistics are introduced and studied
in Chernobai et al (2005), and designed to put most of the weight on
the upper tail. Results for the external losses for the conditional
approach are presented in Table 16.4. The remaining four cases are
presented in the appendix. For the external-type losses, the log-
normal, Weibull, and log-Weibull distributions show the best t in
terms of low statistic values and high p-values.
Expected loss, value-at-risk and conditional value-at-risk
In this section, we estimate the EL, VaR and CVaR and examine the
impact of ignoring the missing data on the operational risk capital
charge. We use a forward-looking approach, and use the functional
formof thefrequencyandtheparameters of theseveritydistribution,
obtained from the historical data over a 23-year period, to forecast
expected total loss, VaRand CVaRone year ahead. We only consider
378
the Cubic I case for the frequency. Table 16.5 provides the estimates
of expected loss (whenever applicable), VaR and CVaR estimates for
2003, obtained via 50,000 Monte Carlo samples, and compares the
gures obtained using the naive approach and the conditional
approach. The remaining four cases are covered in the appendix.
Table 16.5 provides sufcient evidence to conclude that in most
cases, the expected aggregated loss, VaR and CVaR gures appear
highly underestimated if the naive approach is used incorrectly
or instead of the conditional. Some gures also indicate that the
effect is more severe for heavier-tailed distributions. We exclude the
exponential distribution fromsubsequent consideration due to poor
performance of in-sample goodness-of-t tests.
Backtesting
In this section, we conduct an out-of-sample backtesting of the mod-
els. We have demonstrated in the theoretical part of the chapter that
tting unconditional distributions to the operational loss frequency
and severity functions would inevitably result in biased estimates.
We have provided (see page 378) empirical evidence to verify that
indeed the capital charge is signicantly underestimated (in most
cases) if the naive approach is wrongly used.
The goal of this section is to determine which loss distribution
ts our ve samples best. Examining how well or how badly vari-
ous considered models predict the true future losses, is, we believe,
the key to determining which of the loss distributions is the best to
be used for practical purposes. For this purpose, we split our data
samples into two parts:
the rst sample consists of all data points in the 198095 time
frame, and will be used for forecasting;
the second sample consists of the remaining data in the 1996
2001 time frame.
We use the rst sample and the obtained truncated loss distribu-
tions parameter estimates to analyse our models predicting power
regarding the data belonging to the second sample. We conduct the
analysis as follows. We assume that our model has a one-step-ahead
predictingpower, withone stepequal toone year (due toa scarcityof
data, it wouldbe unreasonable touse smaller intervals). The window
length of the sample used for calibration is taken to be 16 years.
379
M
O
D
E
L
R
I
S
K
Table 16.6 Average estimates of forecast errors for external-type aggregated losses
Forecasted upper Forecasted upper Overall error:
quantiles v. upper quantiles forecasted
bootstrapped quantiles v. actual loss v. actual loss
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
1J 95 0.2284 0.4071 0.2665 0.4508
99 2.6679 1.4589 2.8631 1.5220 0.4161 0.1373
99.9 59.0091 6.9119 60.2081 6.9971
Weib 95 0.0756 0.2380 0.0981 0.2817
99 0.4529 0.6065 0.5358 0.6698 0.0350 0.0835
99.9 2.7728 1.5693 3.0653 1.6555
log-Weib 95 0.0843 0.2496 0.1078 0.2933
99 0.6127 0.6974 0.7058 0.7605 0.0523 0.0873
99.9 5.6387 2.1352 6.0582 2.2236
11 95 16.7032 3.0645 16.9948 3.1083
99 3287.19 41.7376 3292.66 41.8008 23.210
10
1877.31
99.9 1.110
7
2454.50 1.110
7
2454.59
Left, errors between relative upper quantiles; middle, errors of forecasted upper quantiles relative to realised loss; right, overall error between
forecasted and realised loss. Figures are based on 10,000 Monte Carlo samples for every year in the forecasting period.
3
8
0
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.6 (Cont.) Average estimates of forecast errors for external-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
Burr 95 1174.90 25.2473 1177.23 25.2909
99 0.310
7
1026.97 2.710
6
1027.03 3.310
12
9800.80
99.9 5.710
10
1.810
5
5.710
20
1.810
5
log :
95 0.0916 0.2684 0.1170 0.3121

99 0.7443 0.8117 0.8506 0.8747 0.0792 0.0929
99.9 6.8469 2.4760 7.2695 2.5635
:
: 95 22.0854 3.8156 22.4520 3.8594

99 5606.35 60.8946 5614.50 60.9580 55.810
10
4337.63
99.9 1.110
7
2758.87 1.110
7
2758.96
3
8
1
MODEL RISK
We start with the data from the rst sample, 198095, in order
to conduct forecasting about 1996. First, we estimate the unknown
parameters of truncated distributions. Next, to obtain the distribu-
tion of the annually aggregated losses, we repeat the following a
large number (10,000) of times: we use the estimated parameters to
simulate N losses exceeding the US$25 million threshold, where N
is the number of losses in the year on which we perform forecasting
as dictated by the tted frequency function, and aggregate them.
At each forecasting step (there are seven steps in total) we shift the
window forward by one year and repeat the above procedure. In
this way we test the model for both the severity and severity distri-
butions. We have observed that both Cubic I and Cubic II models
t the data very well. For simplicity, in this section we only focus
on the Cubic I model. Since the observed data is incomplete, we are
only able to compare the forecasting power regarding the truncated
(rather than complete) data. The analysis is composed in two parts.
In the rst part, we compare the high quantiles (95, 99 and 99.9) of
the forecasted aggregated loss distribution with the corresponding
bootstrappedquantiles of the realisedlosses.
7
Table 16.6 presents the
MSE and MAE estimates for the forecasted high quantiles relative to
the corresponding bootstrapped quantiles (left), realised total loss
(middle) and the errors of the simulated aggregate losses relative
to the actual total loss (right), for the external-type losses. The
remaining four cases are given in the appendix. For the external-
type losses, clearly the Weibull model provides the lowest estimates
for the errors, followed by the log-Weibull and log--stable models.
For the remaining four types, Weibull and log-Weibull are the best,
followed by log--stable and lognormal distributions. GPD, Burr
and symmetric stable overestimate the true losses, as is suggested
by very high error estimates.
In the second part of the analysis, we test the severity distribution
models (without checking for the frequency) via the likelihood ratio
(LR) test, proposed by Berkowitz (2000). We estimate the parame-
ters of the loss distribution from the historic data in the calibration
period; then we use this distribution and the estimated parameters
to conduct the forecast one year ahead. The parameters are fully
specied, so we are able to conduct a simple hypothesis testing. LR
tests are uniformly most powerful, so applying them would give us
an accurate estimate of how likely it is that the realised losses have
382
come froma particular distribution. Under the assumption that a t-
ted truncated loss distribution F is true, F 1[0, 1] under the null.
Asimple transformation Y =
1
(F) would transform the values of
CDF into a standard normal random variable Y. The LR test is then
applied to Y directly in the usual way
LR = 2(l
0
l
1
)
where l
0
and l
1
are, respectively, the log-likelihood under the null
parameters = 0, = 1 and under the parameters estimated
via MLE. The p-values are obtained by referring to the chi-squared
table, with two degrees of freedom. Table 16.7 presents the results
for the external losses, and the remaining four cases are given
in the appendix. The symmetric stable shows the highest average
p-values, with log-Weibull and GPD the next highest. The high-
est average p-values were obtained for the Weibull model for the
relationship and human losses, GPD was the best for pro-
cess and log--stable was the best for the technology losses. The
results are slightly surprising compared with the estimated forecast
errors, but conrm many conclusions drawn from the in-sample
goodness-of-t tests.
ROBUST APPROACH
In the previous section (see page 379 onwards), we tested the fore-
castingpower of the consideredloss models. Fromthe rst part of the
analysis, we concluded that moderately heavy-tailed distributions
such as lognormal and Weibull distributions possess a reasonably
good predicting power. The second part of the analysis suggested
that the losses considered for the forecast period are more likely to
be drawn from heavier-tailed distributions such as the GPD or log-
-stable distributions. It is very likely that such a difference between
therst andthesecondparts of theanalyses results fromthepresence
of high-magnitude outliers in the data, which leads to acceptance
of heavy-tailed models, whereas inclusion of such outliers in the
forecasting models can seriously overestimate the predicted losses.
In recent years, outlier-resistant or so-called robust estimates
of parameters have become more widespread in risk management.
Such models, called robust (statistics) models, were introduced
by Huber (1981) and applied to robust regression analysis; more
recent references on robust statistics methods include Huber (2004),
383
M
O
D
E
L
R
I
S
K
Table 16.7 LR statistic and p-values for external-type aggregated losses in the seven-year forecast period
Year
, ..
Average
1 2 3 4 5 6 7 p-value
1J 6.5946 2.6909 0.2571 0.0677 0.0682 6.6879 3.5182
[0.037] [0.260] [0.879] [0.967] [0.967] [0.035] [0.172] [0.474]
Weib 6.8197 2.7312 0.3206 0.1054 0.1319 6.7159 3.5155
[0.033] [0.255] [0.852] [0.949] [0.936] [0.035] [0.172] [0.462]
log-Weib 6.5458 2.3548 0.2483 0.1285 0.1255 6.3278 3.4601
[0.038] [0.308] [0.883] [0.938] [0.939] [0.042] [0.177] [0.475]
11 6.8378 3.1089 0.2408 0.0518 0.0462 7.2744 3.5594
[0.033] [0.211] [0.887] [0.974] [0.977] [0.026] [0.169] [0.468]
Burr 9.5129 5.6630 0.6205 0 0 9.6215 3.9581
[0.009] [0.059] [0.733] [>0.995] [>0.995] [0.008] [0.138] [0.421]
log :
23.2014 32.0407 15.3027 5.1930 20.0156 6.7474 22.5930

[<0.005] [<0.005] [<0.005] [0.075] [<0.005] [0.034] [<0.005] [0.016]
:
: 6.4584 2.5302 0.2017 0.0990 0.0879 7.4456 3.4455

[0.040] [0.282] [0.904] [0.952] [0.957] [0.024] [0.179] [0.477]
p-values are given in square brackets.
3
8
4
Rousseeuw and Leroy (2003), Martin and Simin (2003), Knez and
Ready (1997) and Hampel et al (1986). Robust models treat extreme
data points as outliers (or some standard procedure is used to detect
outliers in the data) which distort the main ow of the loss process.
Practitioners are more likely to be searching for a stable model that
would capture the mainstreamtendency of the operational loss pro-
cess. Under the robust approach, the focus is on modelling the major
bulkof the data that is drivingthe entire process. Robust models help
to protect against the outlier bias in parameter estimates and pro-
vide a better t of the loss distributions to the data than that under
the classical model. Moreover, outliers in the original data can seri-
ouslydrive future forecasts inanunwanted(eg, worst-case scenario)
direction, which is avoided by the robust approach models. Regard-
ing the use of robust methods for the operational risk modelling, the
Basel Committee stated that
datawill needtobe collectedandrobust estimationtechniques
will need to be developed.
BCBS (2001a)
Following the idea of robust statistics, for forecasting purposes
we offer a second methodology, which involves determining out-
liers and trimming the top 15% of the data. This data adjustment
would result in a more robust outlook regarding a general future
scenario. Excluding the outliers in the original loss data is likely to
noticeably improve the forecasting power of considered loss distri-
butions, and can be used for forecasting of the generic (most likely)
scenario of future losses within reasonable boundaries. The result-
ing operational capital charge estimates would be more optimistic
thanotherwisepredictedbytheclassical model. Weemphasise, how-
ever, that we are not recommending the use of only one of the two
approaches (classical or robust) instead of the other. Rather, in the
presence of outliers, we encourage the use of both models for the
analysis, anduse the robust model as the complement tothe classical.
We here consistently exclude the highest 5% of each dataset. We
reproduce the results for the parameter estimates, capital charge
estimates and out-of-sample goodness of t tests, for the external-
type losses.
8
Tables 16.916.11 indicate the following: the estimates
of expected loss, VaR and CVaR are much more realistic; the accu-
racy of the forecasts has remarkably improved, as indicatedby much
385
MODEL RISK

operational loss data, under the robust approach
, F

Exp 2.515610
8
2.580510
8
F
(u) 0.0248 0.0255

1J 16.3676 15.8095
1.5680 1.9705
F
(u) 0.0518 0.1558

Gam 0.5532 0.0491
1.391610
8
6.124410
9
F
(u) 0.1052 1
Weib 1.1900 0.0012
0.6606 0.4178
F
(u) 0.1036 0.3185

log-Weib 1.522410
14
2.138910
10
11.2079 7.9597
F
(u) 0.0879 0.2254

11 0.8995 1.1813
1.296810
7
7.747410
6
F
(u) 0.0718 0.1132

Burr 0.5477 1.1642
1.889010
9
8.612810
5
1.3784 0.8490
F
(u) 0.0502 0.1451

log :
2 2
0.8736 0.4377
1.1087 1.3992
16.3674 15.7960
F
(u) 0.0522 0.1593

:
: 0.7693 0.6598
1.040410
7
6.778510
6
F
(u) 0.0708 0.1208

lower error estimates; both tests (forecast error estimates and the
LR test) converge in their indication of the best model (the robust
approachconrms that the log-Weibull distributionhas the best fore-
cast power for the external-type loss, with Weibull, lognormal,
log-stable, GPD, symmetric stable and Burr next, ordered from best
to poor).
386
external-type losses, under the robust approach
EL VaR
0.95
VaR
0.99
CVaR
0.95
CVaR
0.99
Exp N 0.0080 0.0238 0.0346 0.0304 0.0409
C 0.0116 0.0299 0.0414 0.0371 0.0481
1J N 0.0088 0.0331 0.0804 0.0661 0.1364
C 0.0154 0.0580 0.1642 0.1397 0.3334
Weib N 0.0076 0.0278 0.0466 0.0393 0.0581
C 0.0088 0.0354 0.0715 0.0599 0.1066
log-Weib N 0.0289 0.0507 0.0428 0.0675
C 0.0395 0.0865 0.0704 0.1318
11 N 0.0258 0.0463 0.1834 0.2817 1.0771
C 0.0943 0.5604 2.1575 9.9486
Burr N 0.0751 0.5666 9.8732 48.6049
C 0.0676 0.3246 0.7372 3.1567
log :
N 0.0341 0.0818 0.0684 0.1446

C 0.0570 0.1695 0.1493 0.3841
:
: N 0.0854 0.6791 17.1402 84.8334

C 0.2234 2.4408 129.37 644.14
10
. Figures
CONCLUSIONS
The nalised Basel II Capital Accord acknowledges the importance
of operational riskmanagement andrequires banks toadopt a proce-
dure to also estimate their operational risk capital charge. While the
recent nancial crisis is mainly attributed to credit risk, some of its
roots can actually be traced to operational risks such as inadequate
or failed internal processes, people and systems or from external
events (BCBS 2001a). Obviously, these issues are highly correlated
to the origins of the nancial crisis and the failures that could be
observed in the mortgage, nancial services and banking industries.
In this chapter we examined the impact of model risk with respect
to ignoring collection thresholds and missing data in the estimation
of the operational risk capital charge. We proposed and empirically
investigated a methodology for consistent estimation of the loss and
frequency distributions for the assumed actuarial model of opera-
tional losses in the presence of minimum collection thresholds. A
387
M
O
D
E
L
R
I
S
K
Table 16.10 Average estimates of forecast errors for external-type aggregated losses, under the robust approach
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
1J 95 0.0185 0.1153 0.0260 0.1443
99 0.1532 0.3498 0.1838 0.3921 0.0143 0.0455
99.9 1.7033 1.2369 1.8506 1.2937
Weib 95 0.0037 0.0479 0.0070 0.0749
99 0.0136 0.1018 0.0236 0.1442 0.0019 0.0284
99.9 0.0555 0.2212 0.0838 0.2779
log-Weib 95 0.0035 0.0469 0.0069 0.0753
99 0.0140 0.1069 0.0245 0.1489 0.0019 0.0281
99.9 0.0619 0.2388 0.0918 0.2958
11 95 0.5389 0.5677 0.5729 0.5968
99 45.4984 4.9129 45.9027 4.9548 25964.56 1.6577
99.9 22951.07 108.00 22964.73 108.06
forecasted and realised loss. Figures are based on 10,000 Monte Carlo samples for every year in forecasting period.
3
8
8
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.10 (Cont.) Average estimates of forecast errors for external-type aggregated losses, under the robust approach
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
Burr 95 41.6391 4.3990 41.9815 4.4280
99 22565.03 92.0143 22572.25 92.0562 2.010
10
630.68
99.9 4.110
8
11255.38 4.110
8
11255.44
log :
95 0.0149 0.0951 0.0210 0.1243

99 0.1312 0.2961 0.1568 0.3382 0.0164 0.0413
99.9 1.6894 1.0598 1.7989 1.1166
:
: 95 3.8490 1.6133 3.9474 1.6423

99 604.07 19.7887 605.82 19.8306 1.510
6
10.6711
99.9 6.210
5
626.31 6.210
5
626.37
forecasted and realised loss. Figures are based on 10,000 Monte Carlo samples for every year in forecasting period.
3
8
9
M
O
D
E
L
R
I
S
K
Table 16.11 LR statistic and p-values for external-type aggregated losses in the seven-year forecast period, under
the robust approach
Year
, ..
Average
1 2 3 4 5 6 7 p-value
1J 5.7220 3.6199 0.1477 0 0.1078 6.5407 2.5591
[0.057] [0.164] [0.929] [>0.995] [0.948] [0.038] [0.278] [0.488]
Weib 5.1081 2.9975 0.0533 0 0.0601 5.7516 2.4772
[0.078] [0.223] [0.974] [>0.995] [0.970] [0.056] [0.290] [0.513]
log-Weib 4.6784 2.3294 0 0 0.1273 5.4843 2.4643
[0.096] [0.312] [>0.995] [>0.995] [0.938] [0.064] [0.292] [0.529]
11 6.2144 3.7330 0.1203 0.0289 0.2780 7.1288 2.5038
[0.045] [0.155] [0.942] [0.986] [0.870] [0.028] [0.286] [0.473]
Burr 9.2110 6.0794 0.3852 0 0.2343 8.9561 2.4562
[0.010] [0.048] [0.825] [>0.995] [0.890] [0.011] [0.293] [0.440]
log :
5.7938 94.1860 0.1588 0 125.9073 6.5408 2.5913

[0.055] [<0.005] [0.924] [>0.995] [<0.005] [0.038] [0.274] [0.327]
:
: 6.5552 4.4805 0.2889 0 0.0702 7.7629 2.4301

[0.038] [0.106] [0.866] [>0.995] [0.966] [0.021] [0.297] [0.470]
3
9
0
compound non-homogeneous Poisson process was considered for
the study. The analysis was conducted using losses of ve different
loss categories (relationship, human, processes, technology
and external) obtained from an operational risk loss database.
Our ndings demonstrated that ignoring such minimum thresh-
olds leads to severe biases in corresponding parameter estimates
under the naive approach, in which the thresholds are ignored.
As a consequence, EL, VaR and CVaR are underestimated under the
naive approach and are generally 1.25 times higher under the
conditional approach, in which truncated loss distributions were
tted to the loss data and frequency was adjusted to account for
information loss. A variety of goodness-of-t measures were used
to test the adequacy of different loss distributions. For example,
for the external-type losses the log-Weibull and Weibull distri-
butions showed the best overall t, while more heavy-tailed distri-
butions such as Burr and GPD better t the upper tail for practi-
cally all ve datasets, supporting the conjecture that the operational
loss data is severely heavy-tailed. The ndings were supported by
out-of-sample forecasting.
Analternative approach, the robust approach, was brieyintro-
duced and applied in the forecasting part of the study. Excluding the
fewhighest data points fromthe dataset allows us to investigate the
behaviour of the bulk of the data as well as to examine the sen-
sitivity of parameters and risk measures to the tail events. In this
chapter, applying the robust methodology resulted in signicantly
improved forecasts and conrmed the choice of loss distribution
obtained under the classical approach.
APPENDIX
The following tables and gures show supplemental data referred
to in the text.
391
MODEL RISK
Table 16.12 Frequency functions tted to the operational losses
Parameter estimates
, ..
Process a b c d MSE MAE
Relationship
Cubic I 34.13 1364.82 0.63 3.32 76.57 7.05
Cubic II 930.29 896.17 0.0010 6.82 69.08 6.57
Poisson 36.91 5907.45 65.68
Human
Cubic I 33.49 1436.56 0.65 3.43 68.05 6.89
Cubic II 950.20 917.11 0.0008 6.80 61.59 6.60
Poisson 35.35 6600.38 65.33
Processes
Cubic I 9.44 2098.96 1.04 4.58 22.50 3.64
Cubic II 2034.25 2024.77 0.0007 4.79 23.06 3.65
Poisson 14.13 1664.82 36.57
Technology
Cubic I 0.79 120.20 0.58 3.47 3.71 1.28
Cubic II 137.68 138.39 0.0006 6.32 4.89 1.67
Poisson 3.35 217.04 13.42
Figure 16.6 Fitted frequency functions to the operational losses
10 15 20
200
400
600
800
C
u
m
u
l
a
t
i
v
e

#

l
o
s
s
e
s
Actual
Cubic I
Cubic II
Poisson
10 15 20
200
400
600
800
5 10 15 20
200
300
400
Year
C
u
m
u
l
a
t
i
v
e

#

l
o
s
s
e
s
0
100
5 5
0 0
5 10 15 20
0
20
40
60
80
Year
Actual
Cubic I
Cubic II
Poisson
Actual
Cubic I
Cubic II
Poisson
Actual
Cubic I
Cubic II
Poisson
(a) (b)
(c) (d)
(a) Relationship, (b) human, (c) processes, (d) technology.
392

human, processes and technology-type operational loss data
Relationship
, F

Exp 1.112810
8
1.125310
8
F
(u) 0.0111 0.0112

1J 16.6771 16.1911
1.6956 2.0654
F
(u) 0.0457 0.1250

Gam 0.4018 7.909210
7
4.470810
9
1.961410
9
F
(u) 0.1281 1
Weib 6.103810
5
0.0032
0.5528 0.3538
F
(u) 0.1189 0.3479

log-Weib 0.512810
12
0.269410
8
9.8946 7.0197
F
(u) 0.0938 0.2386

11 1.0882 1.2852
1.551610
7
1.055810
7
F
(u) 0.0604 0.0855

Burr 0.4817 5.1242
3.483210
9
1.022110
4
1.4077 0.4644
F
(u) 0.0365 0.2575

log--stable 1.9097 1.9340
1 1
1.1584 1.5198
16.7182 15.9616
F
(u) 0.0303 0.1742

:
: 0.7377 0.6592
1.369510
7
0.996810
7
F
(u) 0.0558 0.0841

393
MODEL RISK
Table 16.13 (Cont.) Estimated and F

Human
, F

Exp 7.221610
9
7.274110
9
F
(u) 0.0072 0.0073

1J 16.5878 15.4627
1.8590 2.5642
F
(u) 0.0679 0.2603

Gam 0.3167 6.976310
8
2.286910
9
1.167910
9
F
(u) 0.1628 1
Weib 0.0002 0.0240
0.4841 0.2526
F
(u) 0.1501 0.5441

log-Weib 14.325410
12
30.734410
8
9.8946 7.0197
F
(u) 0.1221 0.3718

11 1.3761 1.6562
1.144110
7
0.613510
7
F
(u) 0.0792 0.1344

Burr 0.0938 0.0922
5.181910
27
2.846310
27
4.4823 4.4717
F
(u) 0.0131 0.0195

log--stable 1.6294 1.4042
1 1
1.1395 2.8957
16.8464 10.5108
F
(u) 0.0083 0.8793

:
: 0.6724 0.6061
1.112610
7
0.714310
7
F
(u) 0.0742 0.1241

394

Processes
, F

Exp 3.502010
9
3.514310
9
F
(u) 0.0035 0.0035

1J 17.5163 17.1600
2.0215 2.3249
F
(u) 0.0336 0.0751

Gam 0.3450 0.0247
1.208210
9
0.548010
9
F
(u) 0.1104 1
Weib 0.0001 0.0021
0.4938 0.3515
F
(u) 0.0923 0.2338

log-Weib 2.489410
12
0.109110
8
9.1693 7.1614
F
(u) 0.0687 0.1479

11 1.4754 1.6147
2.923010
7
2.288610
7
F
(u) 0.0328 0.0413

Burr 0.8661 14.3369
4.383510
6
1.198710
4
0.8884 0.3829
F
(u) 0.0405 0.2097

log--stable 2.0000 2.0000
0.9697 0.8195
1.4294 1.6476
17.5163 17.1535
F
(u) 0.0336 0.0760

:
: 0.5902 0.5478
2.719610
7
1.992510
7
F
(u) 0.0358 0.0536

395
MODEL RISK

Technology
, F

Exp 1.291410
8
1.308310
8
F
(u) 0.0128 0.0130

1J 16.6176 15.1880
1.9390 2.7867
F
(u) 0.0742 0.3112

Gam 0.4217 7.517610
6
5.445810
9
2.353810
9
F
(u) 0.1250 1
Weib 6.366810
5
0.0103
0.5490 0.2938
F
(u) 0.1177 0.4485

log-Weib 1.930910
12
11.064710
8
9.4244 5.7555
F
(u) 0.1023 0.3329

11 1.5823 2.0925
1.047010
7
0.344610
7
F
(u) 0.0851 0.2029

Burr 0.0645 0.0684
1.721010
35
8.740610
20
5.8111 5.2150
F
(u) 0.0227 0.8042

log--stable 2.0000 2.0000
0.7422 0.8040
1.3715 1.9894
16.6181 15.1351
F
(u) 0.0747 0.3195

:
: 0.1827 0.1827
0.167610
7
0.167610
7
F
(u) 0.3723 0.3723

396
Figure 16.7 Upper quantiles of tted truncated loss distributions to
operational losses, together with the empirical distribution
C
D
F
0 0.5 1.0 1.5 2.0
0 2 4 6 8 10 12
C
D
F
0 1 2 3 4 5 6 7 8
Empirical
Exponential
Lognormal
Weibull
log Weibull
GPD
Burr
log Stable
Symm. stable
0.95
0.96
0.97
0.98
0.99
1.00
0.95
0.96
0.97
0.98
0.99
1.00
0.95
0.96
0.97
0.98
0.99
1.00
0.95
0.96
0.97
0.98
0.99
1.00
C
D
F
C
D
F
x 10
9
0 1 2 3 4 5 6
x 10
10
x 10
8
x 10
9
Data
Data
(a)
(c)
Data
Data
(b)
(d)
(a) Relationship, (b) human, (c) processes, (d) technology.
397
MODEL RISK
Table 16.14 Results of in-sample GOF tests for relationship-type
operational losses
KS V AD AD
2
W
2
Exp 11.0868 11.9973 1.310
7
344.37 50.5365
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
1J 0.8056 1.3341 2.6094 0.7554 0.1012
[0.082] [0.138] [0.347] [0.043] [0.086]
Weib 0.5553 1.0821 3.8703 0.7073 0.0716
[0.625] [0.514] [0.138] [0.072] [0.249]
log-Weib 0.5284 1.0061 3.0718 0.4682 0.0479
[0.699] [0.628] [0.255] [0.289] [0.514]
11 1.4797 2.6084 3.5954 3.7165 0.5209
[<0.005] [<0.005] [0.154] [<0.005] [<0.005]
Burr 1.3673 2.4165 3.3069 3.1371 0.4310
[0.032] [<0.005] [0.309] [<0.005] [0.011]
log :
1.5929 1.6930 3.8184 3.8067 0.7076

[0.295] [0.295] [0.275] [0.290] [0.292]
S
S 1.1634 2.0695 1.410

5
4.4723 0.3630
[0.034] [<0.005] [>0.995] [0.992] [<0.005]
simulations.
398
Table 16.15 Results of in-sample GOF tests for human-type
operational losses
KS V AD AD
2
W
2
Exp 14.0246 14.9145 2.410
6
609.15 80.3703
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
1J 0.8758 1.5265 3.9829 0.7505 0.0804
[0.032] [0.039] [0.126] [0.044] [0.166]
Weib 0.8065 1.5439 4.3544 0.7908 0.0823
[0.103] [0.051] [0.095] [0.068] [0.188]
log-Weib 0.9030 1.5771 4.1343 0.7560 0.0915
[0.074] [0.050] [0.115] [0.115] [0.217]
11 1.4022 2.3920 3.6431 2.7839 0.3669
[<0.005] [<0.005] [0.167] [<0.005] [<0.005]
Burr 2.2333 3.1970 4.7780 7.0968 1.2830
[0.115] [0.115] [0.174] [0.115] [0.115]
log :
9.5186 9.5619 36.2617 304.61 44.5156

[0.319] [0.324] [0.250] [0.312] [0.315]
S
S 1.1628 2.1537 5.810

5
11.9320 0.2535
[0.352] [0.026] [0.651] [0.971] [0.027]
399
MODEL RISK
Table 16.16 Results of in-sample GOF tests for process-type
operational losses
KS V AD AD
2
W
2
Exp 7.6043 8.4160 3.710
6
167.60 22.5762
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
1J 0.6584 1.1262 2.0668 0.4624 0.0603
[0.297] [0.345] [0.508] [0.223] [0.294]
Weib 0.6110 1.0620 1.7210 0.2069 0.0338
[0.455] [0.532] [0.766] [0.875] [0.755]
log-Weib 0.5398 0.9966 1.6238 0.1721 0.0241
[0.656] [0.637] [0.832] [0.945] [0.918]
11 1.0042 1.9189 4.0380 2.6022 0.3329
[0.005] [<0.005] [0.128] [<0.005] [<0.005]
Burr 0.5634 0.9314 1.6075 0.2639 0.0323
[0.598] [0.800] [0.841] [0.794] [0.840]
log :
0.6931 1.1490 2.0109 0.4759 0.0660

[0.244] [0.342] [0.534] [0.202] [0.258]
S
S 1.3949 1.9537 3.310

5
6.5235 0.3748
[0.085] [0.067] [0.931] [0.964] [0.102]
simulations.
400
Table 16.17 Results of in-sample GOF tests for technology-type
operational losses
KS V AD AD
2
W
2
Exp 3.2160 3.7431 27.6434 27.8369 2.9487
[<0.005] [<0.005] [<0.005] [<0.005] [<0.005]
1J 1.1453 1.7896 2.8456 1.3778 0.2087
[<0.005] [0.005] [0.209] [<0.005] [<0.005]
Weib 1.0922 1.9004 2.6821 1.4536 0.2281
[<0.005] [<0.005] [0.216] [<0.005] [<0.005]
log-Weib 1.1099 1.9244 2.7553 1.5355 0.2379
[<0.005] [<0.005] [0.250] [<0.005] [<0.005]
11 1.2202 1.8390 3.0843 1.6182 0.2408
[<0.005] [<0.005] [0.177] [<0.005] [<0.005]
Burr 1.1188 0.9374 2.6949 2.0320 0.3424
[0.389] [0.380] [0.521] [0.380] [0.380]
log :
1.1540 1.7793 2.8728 1.3646 0.2071

[<0.005] [0.007] [0.208] [<0.005] [<0.005]
S
S 2.0672 2.8003 2.710

5
19.6225 1.4411
[>0.995] [>0.995] [>0.995] [>0.995] [0.964]
simulations.
401
MODEL RISK
relationship-type losses
EL VaR
0.95
VaR
0.99
CVaR
0.95
CVaR
0.99
Exp N 0.1348 0.2231 0.2704 0.2515 0.2959
C 0.1422 0.2322 0.2763 0.2598 0.3016
1J N 0.1105 0.2832 0.5386 0.4662 0.8685
C 0.1634 0.4662 1.0644 0.9016 1.9091
Weib N 0.1065 0.2203 0.2996 0.2700 0.3505
C 0.1284 0.3187 0.5121 0.4430 0.6689
log-Weib N 0.2235 0.3193 0.2845 0.3873
C 0.3332 0.5902 0.5049 0.8386
11 N 0.8240 4.1537 9.6367 41.5129
C 1.5756 11.3028 52.8928 249.17
Burr N 2.8595 31.5637 1234.95 6139.69
C 1.5713 11.5519 25.9142 114.20
log :
N 1.9124 7488.08
C 0.4359 0.9557 0.8277 1.7443
:
: N 2.1873 17.3578 329.99 1627.38

C 4.5476 56.2927 376.09 1822.93
10
. Figures
402
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.19 Estimates of expected aggregated loss, VaR, and CVaR for human-type losses
EL VaR
0.95
VaR
0.99
CVaR
0.95
CVaR
0.99
Exp N 0.0305 0.4657 0.5452 0.5145 0.5848
C 0.0316 0.4818 0.5618 0.5312 0.6041
1J N 0.1981 0.4970 0.9843 0.8534 1.6652
C 0.4171 1.2161 3.4190 3.3869 9.4520
Weib N 0.1993 0.4017 0.5507 0.4945 0.6456
C 0.2881 0.7997 1.5772 1.3232 2.3746
log-Weib N 0.4174 0.6184 0.5460 0.7732
C 0.8672 1.8603 1.5569 3.0576
11 N 3.9831 33.5741 3945.75 19685.73
C 12.1150 168.64 67596.68 3.410
5
Burr N 85.5620 2690.44 2.110
6
1.110
7
C 94.8281 3042.32 7.710
6
3.810
7
log :
N 1.910
7
7.210
24

C 2.2737 4.2319 3.6742 6.7179
:
: N 6.2811 77.4762 554.19 2691.79

C 14.5771 203.24 3922.83 19403.45
10
. Figures are based on 50,000 Monte Carlo samples.
4
0
3
MODEL RISK
process-type losses
EL VaR
0.95
VaR
0.99
CVaR
0.95
CVaR
0.99
Exp N 0.5140 0.8175 0.9664 0.9109 1.0504
C 0.5407 0.8522 1.0058 0.9483 1.0904
1J N 0.5622 1.5508 3.5665 3.1201 6.9823
C 0.8457 2.5610 6.5625 5.7823 13.9079
Weib N 0.4170 0.8800 1.2102 1.0891 1.4311
C 0.5131 1.2761 2.1308 1.8257 2.8578
log-Weib N 0.9611 1.4498 1.2794 1.8514
C 1.4780 2.6511 2.2575 2.8255
11 N 12.5930 131.25 1121.25 5467.11
C 20.8700 262.52 4384.77 21648.21
Burr N 6.8569 52.0391 206.06 962.88
C 1.7987 4.1859 3.9723 9.7191
log :
N 1.5613 3.5159 2.9589 6.0887

C 2.5394 6.7070 5.9289 14.3725
:
: N 38.7627 529.99 1.310

5
6.410
5
C 74.9073 1280.02 1.810
6
8.910
6
10
. Figures
404
technology-type losses
EL VaR
0.95
VaR
0.99
CVaR
0.95
CVaR
0.99
Exp N 0.0232 0.0598 0.0828 0.0741 0.0964
C 0.0306 0.0712 0.0963 0.0867 0.1102
1J N 0.0324 0.1202 0.3593 0.2970 0.7303
C 0.0958 0.2898 1.2741 1.5439 5.4865
Weib N 0.0226 0.0798 0.1368 0.1159 0.1795
C 0.0358 0.1454 0.3625 0.2958 0.6180
log-Weib N 0.0861 0.1683 0.1399 0.2408
C 0.1670 0.4747 0.3885 0.8817
11 N 0.4415 5.6954 56.3367 276.03
C 1.6249 54.4650 92471.16 4.610
5
Burr N 2.8840 158.94 5.3 10
6
2.610
7
C 9.0358 855.78 8.310
7
4.210
8
log :
N 0.1222 0.3560 0.3024 0.7435

C 0.2990 1.2312 1.6447 5.9933
:
: N 4.910
5
3.210
9
9.410
23
4.710
24
C 7.110
6
6.910
10
1.510
26
7.610
26
10
. Figures
405
M
O
D
E
L
R
I
S
K
Table 16.22 Average estimates of forecast errors for relationship-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
1J 95 0.1260 0.3356 0.1974 0.4259
99 0.9316 0.8823 1.4409 1.1805 0.1225 0.1780
99.9 10.5910 3.0870 13.1306 3.5322
Weib 95 0.0837 0.2262 0.0636 0.2260
99 0.1731 0.3680 0.2335 0.4681 0.0335 0.1348
99.9 0.5521 0.7183 0.8549 0.9056
log-Weib 95 0.0846 0.2368 0.0760 0.2508
99 0.2031 0.4193 0.3072 0.5387 0.0382 0.1401
99.9 0.7933 0.8496 1.3684 1.1500
11 95 8.6863 2.8066 9.9388 3.0156
99 516.75 22.0579 531.19 22.3719 2.610
4
3.4405
99.9 1.110
5
320.35 1.110
5
320.79
4
0
6
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.22 (Cont.) Average estimates of forecast errors for relationship-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
Burr 95 31.6051 5.1915 34.0472 5.4006
99 2892.27 48.9572 2926.30 49.2709 16.110
5
16.7084
99.9 31.110
5
1483.69 31.210
5
1484.14
log :
95 0.1246 0.3250 0.1886 0.4088

99 0.7544 0.7781 1.2114 1.0588 0.1378 0.1744
99.9 9.8656 2.9407 12.3724 3.3795
:
: 95 117.28 10.1415 121.96 10.3504

99 1.3510
4
109.31 1.3610
4
109.62 6.010
6
25.4668
99.9 5.010
6
2125.12 5.010
6
2125.56
4
0
7
M
O
D
E
L
R
I
S
K
Table 16.23 Average estimates of forecast errors for human-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
1J 95 0.0061 0.8881 1.9712 1.3778
99 16.0585 3.9059 22.1948 4.6600 13.7183 0.5331
99.9 478.23 21.2076 522.73 22.2719
Weib 95 0.2386 0.4025 0.5234 0.6901
99 1.1650 0.9607 2.8835 1.6854 0.2607 0.3286
99.9 11.6062 3.2353 18.8769 4.2933
log-Weib 95 0.3137 0.4948 0.7277 0.8196
99 2.3749 1.4061 4.7648 2.1604 0.4318 0.3577
99.9 29.3964 5.1937 40.6156 6.2559
11 95 694.06 25.8942 721.22 26.3880
99 1.510
5
373.03 1.510
5
373.78 14.310
10
1796.50
99.9 7.510
8
24126.68 7.510
8
24127.74
4
0
8
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.23 (Cont.) Average estimates of forecast errors for human-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
Burr 95 97987.89 285.09 98258.80 285.58
99 8.510
7
8688.99 8.510
7
8689.75 0.610
15
1.610
5
99.9 4.910
12
1.810
6
4.910
12
1.810
6
log :
95 0.8942 0.7879 1.7607 1.2803

99 11.9707 3.2636 17.2067 4.0170 2.6761 0.4797
99.9 289.35 15.3948 322.95 16.4575
:
: 95 671.54 25.3614 697.49 25.8550

99 1.410
5
365.68 1.410
5
366.43 6.310
9
473.22
99.9 7.710
8
2.210
4
7.710
8
2.210
4
4
0
9
M
O
D
E
L
R
I
S
K
Table 16.24 Average estimates of forecast errors for process-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
1J 95 2.5846 1.4030 2.1980 1.3718
99 19.7641 4.2121 26.0670 4.7600 3.6324 0.8224
99.9 290.95 16.0800 339.94 17.7763
Weib 95 2.2836 0.9438 0.5316 0.6703
99 3.2796 1.2998 1.6572 1.1983 0.6991 0.6419
99.9 9.0910 2.6978 11.2902 3.0559
log-Weib 95 2.2026 0.9875 0.6936 0.7703
99 4.0327 1.6497 3.4160 1.7046 0.7913 0.6661
99.9 15.7046 3.7568 21.4673 4.3466
11 95 384.25 17.4088 407.12 18.2343
99 97300.89 275.10 97809.74 276.29 1.710
13
15951.40
99.9 6.210
8
17619.28 6.210
8
17620.00
4
1
0
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.24 (Cont.) Average estimates of forecast errors for process-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
Burr 95 5.3751 2.0950 6.7041 2.2717
99 248.36 13.3752 277.35 14.6014 6.710
5
5.1213
99.9 33069.50 160.90 33685.29 162.63
log :
95 2.3570 1.2818 1.6089 1.1205

99 15.7130 3.3234 20.0033 3.8551 4.6436 0.7726
99.9 280.03 13.5974 319.64 15.2891
:
: 95 2176.66 44.5025 2238.36 45.3302

99 7.810
5
839.45 7.910
5
840.66 9.910
10
3033.42
99.9 0.610
10
62936.26 0.610
10
62937.95
4
1
1
M
O
D
E
L
R
I
S
K
Table 16.25 Average estimates of forecast errors for technology-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
1J 95 0.1879 0.4135 0.2029 0.4316
99 3.6536 1.8314 3.7435 1.8571 1.0549 0.1402.
99.9 136.90 11.1882 137.62 11.2242
Weib 95 0.0294 0.1624 0.0345 0.1806
99 0.1819 0.4168 0.2022 0.4426 0.0140 0.0535
99.9 1.1294 1.0467 1.1985 1.0824
log-Weib 95 0.0439 0.1979 0.0501 0.2162
99 0.3556 0.5804 0.3838 0.6063 0.0272 0.0620
99.9 3.1991 1.7705 3.3291 1.8065
11 95 45.7139 5.4619 45.9212 5.4803
99 77821.03 203.46 77829.78 203.49 1.810
13
2.010
4
99.9 5.1 40154.50 5.1 10
9
40154.53
4
1
2
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.25 (Cont.) Average estimates of forecast errors for technology-type aggregated losses
, .. , .. , ..
Quantile MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
) MSE ( 10
20
) MAE ( 10
10
)
Burr 95 1853.93 34.1046 1854.92 34.1228
99 4.210
7
4135.69 4.210
7
4135.71 1.510
20
5.410
7
99.9 1.610
13
2.310
6
1.610
13
2.310
6
log :
95 0.1901 0.4161 0.2053 0.4344

99 3.3955 1.7774 3.4864 1.8031 11.9464 0.1610
99.9 147.59 11.8642 148.36 11.8998
:
: 95 8.210
20
1.110
10
8.210
20
1.110
10
99 9.410
29
3.710
14
9.410
29
3.710
14
6.010
61
2.910
28
99.9 1.110
43
1.210
21
1.110
43
1.210
21
4
1
3
M
O
D
E
L
R
I
S
K
Table 16.26 LR statistic and p-values for relationship-type aggregated losses in the seven-year forecast period
Year
, ..
Average
1 2 3 4 5 6 7 p-value
1J 0.3579 2.3588 1.9911 1.4466 2.0630 0.3518 2.5297
[0.836] [0.308] [0.370] [0.485] [0.357] [0.839] [0.282] [0.497]
Weib 0.0619 1.5978 0.8392 1.2538 2.0261 0.1808 2.7085
[0.970] [0.450] [0.657] [0.534] [0.363] [0.914] [0.258] [0.592]
log-Weib 0.1383 1.8481 1.2071 1.3434 2.1209 0.2174 2.5593
[0.933] [0.397] [0.547] [0.511] [0.346] [0.897] [0.278] [0.559]
11 0.6618 3.0483 3.2423 1.6954 1.8640 0.7013 2.4631
[0.718] [0.218] [0.198] [0.428] [0.394] [0.704] [0.292] [0.422]
Burr 0.6511 2.8692 3.3531 1.6386 1.7583 0.7717 2.5161
[0.722] [0.238] [0.187] [0.441] [0.415] [0.680] [0.284] [0.424]
log :
0.3433 1.6031 2.0380 5.6319 2.1734 0.3646 9.4886

[0.842] [0.449] [0.361] [0.060] [0.337] [0.833] [0.009] [0.413]
:
: 0.2621 24.6599 2.5327 1.1871 1.3401 0.3538 2.2774

[0.877] [<0.005] [0.282] [0.552] [0.512] [0.838] [0.320] [0.483]
4
1
4
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.27 LR statistic and p-values for human-type aggregated losses in the seven-year forecast period
Year
, ..
Average
1 2 3 4 5 6 7 p-value
1J 0.5022 4.6756 0.1023 2.5790 0.9439 4.7796 0.7730
[0.778] [0.097] [0.950] [0.275] [0.624] [0.092] [0.679] [0.499]
Weib 0.2541 3.7590 0.1958 2.5951 0.6551 4.7877 0.7553
[0.881] [0.153] [0.907] [0.273] [0.721] [0.091] [0.686] [0.530]
log-Weib 0.3783 4.4179 0.0884 2.4636 0.8178 4.4334 0.7605
[0.828] [0.110] [0.957] [0.292] [0.664] [0.109] [0.684] [0.520]
11 0.8031 5.4367 0.1657 2.5592 1.3462 5.3795 0.8338
[0.669] [0.066] [0.921] [0.278] [0.510] [0.068] [0.659] [0.453]
Burr 0.6539 5.4566 0.2840 2.5620 0.8786 7.4635 0.8278
[0.721] [0.065] [0.868] [0.278] [0.645] [0.024] [0.661] [0.466]
log :
0.2067 4.5401 0.1326 2.6464 11.1638 50.2250 0.8028

[0.902] [0.103] [0.936] [0.266] [0.004] [<0.005] [0.670] [0.412]
:
: 19.5913 57.8253 16.5836 22.7072 20.0307 0.8759

[<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [<0.005] [0.645] [0.092]
4
1
5
M
O
D
E
L
R
I
S
K
Table 16.28 LR statistic and p-values for process-type aggregated losses in the seven-year forecast period
Year
, ..
Average
1 2 3 4 5 6 7 p-value
1J 4.7303 7.9488 20.5106 0 5.9690 1.2397 7.6575
[0.094] [0.019] [<0.005] [>0.995] [0.051] [0.538] [0.022] [0.246]
Weib 5.8358 9.0067 19.8166 0 5.3046 1.3434 7.1111
[0.054] [0.012] [<0.005] [>0.995] [0.071] [0.511] [0.029] [0.239]
log-Weib 5.4614 8.6240 20.0130 0 5.6225 1.2509 7.1481
[0.065] [0.013] [<0.005] [>0.995] [0.060] [0.535] [0.028] [0.243]
11 3.3761 7.3430 21.4076 0.0575 6.7186 1.2567 8.0513
[0.185] [0.025] [<0.005] [0.972] [0.035] [0.534] [0.018] [0.253]
Burr 4.4869 8.0195 21.2706 0.0192 6.4037 1.3078 7.1694
[0.106] [0.018] [<0.005] [0.990] [0.041] [0.520] [0.028] [0.243]
log :
4.7196 8.0265 36.6496 2.5550 14.3856 1.2909 7.7456

[0.094] [0.018] [<0.005] [0.279] [<0.005] [0.524] [0.021] [0.134]
:
: 4.1486 0.1850 40.7501 1.6664 23.4444

[0.125] [<0.005] [<0.005] [0.912] [<0.005] [0.435] [<0.005] [0.210]
4
1
6
O
P
E
R
A
T
I
O
N
A
L
V
A
L
U
E
-
A
T
-
R
I
S
K
I
N
T
H
E
P
R
E
S
E
N
C
E
O
F
M
I
N
I
M
U
M
C
O
L
L
E
C
T
I
O
N
T
H
R
E
S
H
O
L
D
Table 16.29 LR statistic and p-values for technology-type aggregated losses in the seven-year forecast period
Year
, ..
Average
1 2 3 4 5 6 7 p-value
1J 1.7031 0.7748 7.9165 1.8076 3.9816
[0.427] [0.679] [0.019] [0.405] [] [0.137] [] [0.333]
Weib 1.4969 0.9152 8.5419 1.9915 4.1399
[0.473] [0.633] [0.014] [0.370] [] [0.126] [] [0.323]
log-Weib 1.4175 0.9414 8.3460 2.3232 4.0876
[0.492] [0.625] [0.015] [0.313] [] [0.130] [] [0.315]
11 2.1113 0.7520 6.8921 1.8079 3.6926
[0.348] [0.687] [0.032] [0.405] [] [0.158] [] [0.326]
Burr 3.1543 0.4183 8.2779 1.3857 4.3545
[0.207] [0.811] [0.016] [0.500] [] [0.113] [] [0.330]
log :
1.7229 0.7572 7.8918 1.7649 3.9782

[0.423] [0.685] [0.019] [0.414] [] [0.137] [] [0.335]
:
: 4.3899 5.5676 14.5400 5.6148 6.4994

[0.111] [0.062] [<0.005] [0.061] [] [0.039] [] [0.055]
4
1
7
MODEL RISK
The opinions expressed in this article are those of the authors,
and not of the authors employers. Svetlozar Rachev gratefully
acknowledges research support by grants from the Division of
Mathematical, Life and Physical Sciences, College of Letters and
Science, University of California, Santa Barbara, the Deutschen
Forschungsgemeinschaft and the Deutscher Akademischer Aus-
tausch Dienst.
1 Cruz (2002) suggests 50%, 15%and 35%for credit, market and operational risks, respectively.
In the 2008 annual report of JPMorgan Chase, credit risk capital accounted for 63%, market
risk capital for 17.5% and operational risk for 11% of total regulatory capital.
2 Summing across business line and event types assumes perfect correlation between different
cells. In more recent Basel II guidelines (BCBS 2006a), it is recommended that banks use
appropriate dependence structures that exist between the cells to produce the aggregate risk
capital.
3 The 2002 Quantitative Impact Study (QIS) revealed that only ve out of 89 banks that par-
ticipated in the study had minimum cut-off thresholds below 10,000, 59 banks (or 66% of
banks) used a threshold of around 10,000 and 13 rms (or 15%) had thresholds exceeding
that amount (BCBS 2003). There could be various reasons for truncation. First, data recording
is costly. Second, data-entry errors that can occur while recording a large number of small
losses result in additional operational losses for a rm. Third, smaller losses are easier to hide,
while larger losses must be reported, which results in smaller losses being under-represented
from a complete database if all losses were recorded. And fourth, small frequent losses are
perceived as routing and immaterial, so banks often opt to leave them unrecorded.
4 Asimilar model was applied by Chernobai et al (2006a) to the natural catastrophe insurance
model. Using real insurance claims data, they showed that the model misspecication of the
claims leads to serious underestimation of the ruin probabilities.
5 The case of a sinusoidal rate function is considered in Chernobai et al (2006a).
6 It is also notable that EL does not exist for some very heavy-tailed distributions that possess
an innite mean.
7 The use of bootstrapping and Monte Carlo was suggested by the Basel Committee (BCBS
2001b, 2004).
8 We omit the remaining four cases for brevity. Key results and conclusions remain the same.
REFERENCES
BCBS, 2001a, Consultative Document: Operational Risk, URL: http://www.bis.org.
BCBS, 2001b, Working Paper on the Regulatory Treatment of Operational Risk, URL:
http://www.bis.org.
BCBS, 2003, The 2002 Loss Data Collection Exercise for Operational Risk: Summary of
the Data Collected, URL: http://www.bis.org.
BCBS, 2004, International Convergence of Capital Measurement and Capital Standards,
URL: http://www.bis.org.
BCBS, 2006a, International Convergence of Capital Measurement and Capital Stan-
dards, URL: http://www.bis.org.
BCBS, 2006b, Observed Range of Practice in Key Elements of Advanced Measurement
Approaches (AMA), URL: http://www.bis.org.
Bee, M., 2005, On Maximum Likelihood Estimation of Operational Loss Distributions,
Working Paper, University of Trento.
418
Berkowitz, J., 2000, Testing Density Forecasts with Applications to Risk Management,
Working Paper, University of California Irvine.
Chernobai, A., S. Rachev and F. Fabozzi, 2005, Composite Goodness-of-Fit Tests for
Left-Truncated Loss Samples, Working Paper, University of California Santa Barbara.
Chernobai, A., K. Burneki, S. Rachev, S. Trck and R. Weron, 2006a, Modelling Catas-
trophe Claims with Left-Truncated Severity Distributions, Computational Statistics 21,
pp. 53755.
Chernobai, A., C. Menn, S. Trck and S. Rachev, 2006b, ANote on the Estimation of the
Frequency and Severity Distribution of Operational Losses, Mathematical Scientist 30(2),
pp. 8797.
Crama, Y., G. Hbner and J. Peters, 2007, Impact of the Collection Threshold on the
Determination of the Capital Charge for Operational Risk, in G. Gregoriou, ed., Advances
in Risk Management, pp. 121 (Palgrave Macmillan).
Cruz, M., 2002, Modeling, Measuring and Hedging Operational Risk (New York: John Wiley
& Sons).
DeCanio, S., and W. Watkins, 1998 , Investment in Energy Efciency: Do the Character-
istics of Firms Matter?, The Review of Economics and Statistics 80(1), pp. 95107.
Dempster, A., N. Laird and D. Rubin, 1977, MaximumLikelihood fromIncomplete Data
via the EMAlgorithm, Journal of the Royal Statistical Society, Series B 39(1), pp. 138.
Embrechts, P., C. Klppelberg and T. Mikosch, 1997, Modeling Extremal Events for
Insurance and Finance (Springer).
Hampel, F., E. Ronchetti, R. RousseeuwandW. Stahel, 1986, Robust Statistics: The Approach
Based on Inuence Functions (Chichester: John Wiley & Sons).
Huber, P. J., 1981, Robust Statistics (New York: John Wiley & Sons).
Huber, P. J., 2004, Robust Statistics (Hoboken, NJ: John Wiley & Sons).
Jorion, P., 2000, Value-at-Risk: The NewBenchmark for ManagingFinancial Risk, SecondEditon
(New York: McGraw-Hill).
Knez, P., and M. Ready, 1997, On the Robustness of Size and Book-to-Market in Cross-
Sectional Regressions, Journal of Finance 52, pp. 135582.
Martin, R., and T. Simin, 2003, Outlier Resistant Estimates of Beta, Financial Analysts
Journal 59, pp. 5669.
McLachlan, G., and T. Krishnan, 1997, The EM Algorithm and Extensions, Wiley Series in
Probability and Statistics (Chichester: John Wiley & Sons).
Meng, X., and D. van Dyk, 1997, The EM Algorithm: An Old Folk-Song Sung to a Fast
New Tune, Journal of the Royal Statistical Society, Series B 59(3), pp 51167.
Ross, S., 2001, Simulation, Third Edition (Boston, MA: Academic Press).
Rousseeuw, P., and A. Leroy, 2003, Robust Regression and Outlier Detection (Hoboken, NJ:
John Wiley & Sons).
Shevchenko, P., and G. Temnov, 2007, Addressing the Impact of Data Truncation and
Parameter Uncertainty on Operational Risk Estimates, The Journal of Operational Risk 2,
pp. 326.
Shevchenko, P., and G. Temnov, 2009 , Modeling Operational Risk Data Reported above
a Time-Varying Threshold, The Journal of Operational Risk 4(2), pp. 1942.
Wulfsohn, M., and A. Tsiatis, 1997, A Joint Model for Survival and Longitudinal Data
Measured with Error, Biometrics 53(1), pp 3309.
419
17
Operational Risk and
Hedge Fund Performance:
Evidence from Australia
Robin Luo, XiangkangYin
La Trobe University
The rapidgrowthinhedge funds is a global trend. At the endof 2008,
global hedge fund assets under management were US$1.29 trillion
(Gentilini 2009). Global mutual fund assets shrank, from US$21.82
trillion at the end of 2006 to US$18.97 trillion at the end of 2008
(Investment Company Institute 2009). The Australian hedge fund
industry has also grown signicantly since the late 1990s, with par-
ticularly strong growthinthe past fewyears. Australias US$59.7 bil-
lion hedge fund industry is currently the largest in Asia-Pacic
(Australian Trade Commission 2008).
As the hedge fund industry has grown explosively, so too has
the list of fund failures. One of the famous examples of hedge
fund failure is the Long Term Capital Fund managed by Long-Term
Capital Management (LTCM). LTCM was founded by a number of
star traders and Nobel Laureate economists in 1994. The fund was
spectacularly successful until the middle of 1998, with an average
annual return of 33.4% during the period 199597. It had capital
of US$4.8 billion and assets of US$120 billion at the beginning of
1998. In the aftermath of the Russian crisis in August 1998, the fund
lost almost all its capital in one month. In September 2006, another
large hedge fund, Amaranth, reported losses of more than US$6 bil-
lion, apparently incurredin only one month, representing a negative
return over that month of roughly 66% (Stulz 2007).
To better understand why hedge funds fail and howthese failures
can be avoided, Kundro and Feffer (2004) conducted a study based
on a database containing over 100 failed hedge funds in the past
20 years, capturing details of losses, litigation and root causes. This
421
MODEL RISK
Figure 17.1 Distribution of operational issues contributing to
operational risk in hedge funds
P
e
r
c
e
n
t
0
5
10
15
20
25
30
35
45
40
M
i
s
r
e
p
r
e
s
e
n
t
a
t
i
o
n
o
f

i
n
v
e
s
t
m
e
n
t
s
M
i
s
a
p
p
r
o
p
r
i
a
t
i
o
n
o
f

f
u
n
d
s
U
n
a
u
t
h
o
r
i
s
e
d
t
r
a
d
i
n
g
I
n
a
d
e
q
u
a
t
e
r
e
s
o
u
r
c
e
s
O
t
h
e
r
41
33
14
6
9
Source: Hedges (2005).
study found that operational issues account for 50% of hedge fund
failures.
Hedges (2005) concludes that the primary cause of a funds failure
can be attributed to one of three categories of risk: investment risk,
business risk or operational risk. Investment risks are the market
and related risks associated with the investment style of the fund
or the securities it holds. This is the type of risk that fund investors
generally intendto take in exchange for the promise of performance.
Business risks are those associatedwith the management of the fund
company as a business but are not directly related to market move-
ments, eg, a failure to reach a base level of assets under management
or a change in management of the fund. They also include factors
that stemfromthe possibilitythat the fundmanager will be unable to
create a sustainable business. Operational risks are associated with
supporting the operating environment of the fund. The operating
environment includes trade processing, accounting, administration,
valuation and reporting. These are the types of risk that investors do
not intend to take as part of their investment strategy. For example,
it could be the risk that an investment might be fraudulent or that
managers might misrepresent performance.
422
OPERATIONAL RISK AND HEDGE FUND PERFORMANCE: EVIDENCE FROM AUSTRALIA
The most common operational issues related to hedge fund losses
include misrepresentation of fundinvestments, misappropriation of
investor funds, unauthorisedtrading andinadequate resources. The
most signicant operational issue is the misrepresentation of invest-
ments, which is dened as the act of creating or causing the genera-
tionof reports andvaluations withfalse andmisleadinginformation.
This may be due to deliberate deception or to operational errors.
Traditional mutual funds have to disclose a lot of information
to investors. They must report their holdings to the regulators (for
instance, the Securities and Exchanges Commission (SEC) in the US)
and must have audited statements. Hedge funds may agree contrac-
tually to disclose some types of information and to provide audited
nancial statements if they decide that it helps them to recruit
investors, but they are not legally required to do so (Stulz 2007).
For instance, referring to the LTCM, Lowenstein (2001, p. 32) states:
Long Term even refused to give examples of trades, so potential
investors had little idea of what they were doing.
Therefore, mandatory disclosure becomes an important regula-
tory tool, intended to allow investors to assess the operational risks
without necessarily constraining managerial actions. Brown et al
(2008) use the -score to explore the questions of whether Form
ADV information was redundant in the investment marketplace.
They test the potential value and materiality of operational risk and
conict of interest variables disclosed by a large number of hedge
funds in the US in February 2006. They nd that operational risk
indicators are conditionally correlated with conict of interest vari-
ables and with lower leverage and concentrated ownership. How-
ever, they argue that operational risk factors had no ex post effect on
the owperformance relationship, suggesting that investors either
lack this information or do not regard it as material.
This chapter attempts to investigate the relationship between
operational risk and fund performance in the Australian hedge fund
market, as it has grown rapidly in recent years to become the largest
inAsia-Pacic. The remainder of the chapter is organised as follows.
In the next section, we discuss the legal structure and risk disclosure
of Australian hedge funds. We then develop the hypotheses for the
relationship between operational risk and fund performance before
describing the data and main empirical results. Concluding remarks
are given in the nal section.
423
MODEL RISK
THE LEGAL STRUCTURE AND RISK DISCLOSURE OF
AUSTRALIAN HEDGE FUNDS
Australian hedge funds are most commonly structured as trusts,
although company structures (typically unlisted and domiciled in
offshore tax havens) are also used. The administration and invest-
ment decisions of the fund are handled by a hedge fund manager,
whois appointedbythe fundtrustees or boardof directors. There are
no specic regulations covering the legal structure of hedge funds in
Australia. Like other types of managedfunds, hedge funds fall under
the scope of the Corporations Act 2001; the provisions that apply
depend on whether they are structured as trusts or companies.
1
Inthe case of a trust, if a hedge fundis marketedto retail investors,
it must then be registered with the Australian Securities and Invest-
ments Commission (ASIC), and is subject to certain operational
anddisclosure requirements designedto protect investors interests.
Their requirements include the appointment of a responsible entity
charged with certain duciary duties, the provision of adequate
product disclosure statements and annual or semi-annual reporting
of nancial statements.
Hedge funds that do not accept funds from retail investors are
subject to fewer requirements, as their investors are considered to
be better placed to monitor and manage their investments without
government regulation. Hedge funds structured as companies must
comply with provisions covering capital raising, corporate gover-
nance anddisclosure requirements (Reserve Bank of Australia 2004).
Apart fromregistrationof the retail funds themselves, all hedge fund
managers (whether their funds are retail or wholesale) must hold an
Australian Financial Services Licence issued by ASIC under Chap-
ter 7 of the Corporations Act 2001. All licensees have ongoing compli-
ance requirements. Table 17.1 provides the statistics of legal struc-
ture of Australian hedge funds. Nearly 30% of the hedge funds in
Australia are structured as Caymans Corporation.
2
Australian hedge funds are required to conform to the same reg-
ulatory obligations as other managed funds; however, the nature
of hedge fund structures and trading strategies suggests additional
disclosure is required compared with more traditionally managed
funds. Australian hedge funds are subject to the Alternative Invest-
ment Management Association (AIMA) Australia Risk Disclosure
Guidelines.
3
424
Table 17.1 Australian hedge funds: legal structure
Legal structure Number %
Australia Corporation 4 5.71
Australia Investment Trust 3 4.29
Australia Management Accounts 2 2.86
Australia Proprietary LLC 1 1.43
Australia Trust 1 1.43
Australia Unit Trust 9 12.86
Bermuda Investment Company 1 1.43
BVI Limited Liability Company 3 4.29
Cayman Islands Exempted LLC 1 1.43
Caymans Corporation 19 27.14
Caymans Exempted Company 6 8.57
Caymans Investment Company 4 5.71
Caymans Investment Trust 1 1.43
Caymans LLC 7 10.00
Caymans Limited Partnership 2 2.86
Caymans open-ended LLC 1 1.43
Caymans Other 1 1.43
Delaware Limited Partnership 1 1.43
Total 70 100
Source: Hedge Fund Research (HFR) database.
AIMArequires hedge fund managers to provide a product disclo-
sure statement (PDS) or aninformationmemorandumtothe clients.
4
Thedisclosuredocuments, either PDSor informationmemorandum,
shouldgenerally disclose the following 11 categories of information:
(i) disclaimers;
(ii) key features of the fund;
(iii) details about the fund manager and other intermediaries;
(iv) investment approach, strategy and style of the fund;
(v) risks of investing;
(vi) fees and other costs;
(vii) information about investing in the fund;
(viii) valuation, unit pricing and distributions;
(ix) reporting;
(x) taxation; and
(xi) further information about the hedge fund.
425
MODEL RISK
Category (iv) requires the investment objectives and investment
strategies be detailed in the disclosure document. Other issues,
includingleverage, short selling, liquidity, market exposure andpast
performance, should also be discussed in the document. Some of the
Australian hedge fund characteristics will be further investigated in
the results section (see page 429).
METHODOLOGY AND HYPOTHESES
Brown et al (2008) employ the following probit regression in order
to examine the relationship between hedge fund performance and
operational risk in the US. In their regression, the funds appraisal
ratio is the dependent variable and the set of explanatory variables
includes those identied in the probit as determinants of legal or
regulatory problems
AppraisalRatio
i
=
0
+
1
LogAssets
i
+
2
StdDev
i
+
3
Onshore
i
+
4
LockupPeriod
i
+
5
IncentiveFee
i
+
6
HighWaterMark
i
+
7
Relationship
i
+
8
DirectDomestic
i
+
9
75%Ownership
i
+
10
_
j=1
j
StyleDummies
ji
+
i
Appraisal ratios are calculated using monthly fund returns with
respect to funds respective style indexes and US Treasury Bill
returns. LogAssetsis the logof the assets under management inUS
dollars. StdDev is the standard deviation of a funds returns over
the life of the fund. Onshore equals 1 if the fund is based in the US.
IncentiveFee is the funds incentive fee in percent. HighWater-
Mark equals 1 if the fund has a high-water mark. Relationship
equals 1 if the fund has any external conicts of interest. Direct-
Domestic is the number of domestic corporations listed as direct
owners. The 75% ownership is the percentage of direct owners who
own at least 75% of the company. Style dummies were included to
control for style differences.
The potential conicts-of-interest data of US hedge funds were
obtained fromthe FormADVlings, and operational risk character-
istics were takenfromthe Lipper TASS Inc (TASS) database inBrown
et als (2008) paper. A similar dataset is not available in Australia.
5
426
Therefore, we modify Brown et als (2008) model to suit the reg-
ulatory and disclosure requirement of the Australian hedge fund
industry. We compose the hypotheses (H1)(H3) below to test the
relationshipbetweenthe operational risk, hedge fundcharacteristics
and fund performance.
It is widely accepted that the Big Four rms
6
impose a high
level of earnings quality in order to protect their brand name repu-
tation from legal exposure and reputation risk which can arise from
misleading nancial reports by clients (DeAngelo 1981). Theoretical
studies have shownthat the rms auditedbythe larger or more pres-
tigious accounting rms are less exposed to operational risk because
larger or more prestigious accounting rms have greater incentives
not to perform a low-quality service at a high-quality price because
they have more wealth (Dye 1993) and more valuable reputations
(DeAngelo 1981). This is also evident in the mutual fund industry
(Luo 2009). We therefore expect that Australian hedge funds audited
by a Big Four rm are exposed to less operational risk and provide
higher returns than those not audited by a Big Four rm. This leads
to the following hypothesis.
(H1) A hedge fund that is audited by a Big Four rm and pro-
vides annual performance auditing report is exposed to less
operational risk and tends to provide higher return to the
investors.
Hedge funds typically charge a xed asset-based fee (management
fee) plus a performance-based fee (incentive fee). Some fund man-
agers take aperformance fee onlyonreturns that exceedsome hurdle
rate of return. The hurdle rate is usually dened in terms of some
short-term interest rate benchmark such as the rate of Treasury bills
or the London Interbank Offered Rate (Libor). Some managers offer
a high-water mark in calculating their performance fees, which
means that the performance fee is triggered only when the investor
has prots that exceed the level of prots that existed the last time a
performance fee was charged.
The high-water mark contract is quite typical in the hedge fund
industry. In such a contract, the hedge fund manager receives a per-
centage of the assets, eg, 2%, under management and also receives a
fraction of the growth in value of the funds assets if and only if the
funds assets rise invalue beyondthe previouslyachievedmaximum
427
MODEL RISK
level: the high-water mark. The current high-water mark serves as
the strike price on a call option. In a given evaluation period, if the
value of the funds assets exceeds the current high-water mark, the
manager receives payment and a newcall option with a newexpira-
tion date, and a new strike price equal to the new high-water mark
is granted to the manager. The compensation contract is similar to a
series of call options. In practice, the sequence will be terminated if
the value of the fund goes to zero or if the investor withdraws funds.
We may think that the option-like character of such contracts
could induce the manager to assume extremely risky positions in
the hope of huge payouts, especially as the manager does not share
directly in any loss of the funds assets. Panageas and Westereld
(2009) develop a formal model on these compensation contracts.
They assume that the hedge fund manager is risk neutral and max-
imises the net present value of expected future compensation fees,
and they argue that the risk-neutral manager chooses a portfolio
that coincides with the portfolio that a constant relative risk aver-
sion investor would choose. Because the current high-water mark
remains the effective strike price on the compensation contract and
all subsequent strike prices must lie above that level, the hedge fund
manager faces a trade-off between current and future payouts and
will therefore choose a portfolio with controlled risk. Panageas and
Westereld (2009) believe that the managers optimal portfolio will
be to place a constant fraction of wealth in the risky asset and the
remainder in a stock portfolio that is mean-variance efcient.
This hypothesis is formulated in an alternative form as follows.
(H2) There is a positive relationship between the hedge funds per-
formance and the management fee and incentive fee charged,
and the high-water mark offered by the fund.
Leverage has been widely used in hedge fund investment strategies.
The use of leverage is absolutely guaranteed to increase the risk of a
position. It is not, however, guaranteedto increase the return. Brown
et al (2008) investigate the relationship between the leverage and
operational risk by examining the difference in three leverage mea-
sures for problemandnon-problemhedge funds inthe US. Theynd
that leverage does increase the operational risk faced by the hedge
funds. A long lock-up period is another feature that makes hedge
funds distinct from traditional mutual funds. This implies that the
428
investors need to have sufcient information about the performance
of hedge funds over a long period before committing their money
to the fund. Agarwal and Naik (2000) propose that there is a posi-
tive relationship between the hedge fund performance and lock-up
period. This hypothesis is formulated as follows.
(H3) There is a positive relationship between the hedge fund per-
formance and the use of leverage, lock-up period and the
redemption notice period.
These hypotheses are then tested by running the following regres-
sion
Ret
i
=
0
+
1
StdDev
i
+
2
Auditor
i
+
3
Apaudit
i
+
4
MFee
i
+
5
IFee
i
+
6
HighWaterMark
i
+
7
Leverage
i
+
8
Lockup
i
+
9
Notice
i
+
i
(17.1)
Ret
i
is the average net-of-fee return of fund i following Agarwal
and Naik (2000). StdDev
i
is the standard deviation of funds return.
Auditor
i
and Apaudit
i
are used to test hypothesis (H1). Auditor
i
equals 1 if the fund is audited by one of the Big Four rms and
0 otherwise. Apaudit
i
equals 1 if the fund has an annual perfor-
mance audit and 0 otherwise. MFee
i
and IFee
i
are the manage-
ment fee and incentive fee, respectively. They are both reported in
percent. HighWaterMark
i
equals 1 if the fund offers a high-water
mark contract. These three variables are included to test the second
hypothesis. Leverage
i
equals 1 if the fund uses leverage. Lock-up
i
is
measured in months, while Notice
i
is measured in days.
DATA AND EMPIRICAL RESULTS
We conduct this research using data provided by Hedge Fund
Research Inc (henceforth HFR) which covers returns earned byAus-
tralian hedge funds from September 1997 to February 2009. The
HFR data set provides information about hedge funds both liv-
ing and dead, and is known to have a lower attrition rate than
other databases such as TASS (Liang 2000). The lower attrition rate
in HFR data suggests that it includes a smaller number of funds
that fail compared with other databases. This potentially exacer-
bates survivorship-bias-related problems in studies that employ
429
MODEL RISK
Table 17.2 Descriptive statistics of Australian hedge funds
N Mean Median StdDev
Return 4,250 0.47 0.56 0.89
StdDev 70 3.45 3.35 1.56
Skewness 70 0.66 0.44 1.82
Kurtosis 70 5.60 2.16 10.10
MFee 70 1.52 1.50 0.52
IFee 70 18.93 20.00 4.16
MinInvt 70 0.50 0.23 1.20
Leverage 70 0.57 1.00 0.50
Hurdle 70 0.40 0.00 0.49
HighWaterMark 70 0.94 1.00 0.23
Lockup 70 3.26 0.00 5.05
Notice 70 36.01 30.00 27.96
Auditor 70 0.74 1.00 0.44
Apaudit 70 0.89 1.00 0.32
Descriptive statistics are given for Australian hedge funds as of February
2009.Return is the average return over the life of the fund.StdDev is the
standard deviation of the funds returns. MFee is the management fee,
reported in percent.IFee is the incentive fee, reported in percent.MinInvt
is the minimum investment of the fund in millions of dollars. Leverage
and HighWaterMark are equal to 1 if the fund uses leverage or has a
high-water mark, respectively. Lockup is the lock-up period measured
in months. Hurdle is the hurdle rate reported in percent. Notice is the
advance notice measured in days. Auditor is equal to 1 if the fund is
audited by one of the Big Four rms. Apaudit is equal to 1 if the fund has
an annual performance audit.
HFR databases. We try to mitigate the problem of spurious infer-
ences caused by survivorship-related issues by including data on
both living and dead hedge funds (Brown et al 1992, 2001).
The characteristic and performance data of 70 living and dead
Australian hedge funds obtained fromthe HFRdatabase is reported
in Table 17.2.
We have the following results. Firstly, negative skewness implies
that the distribution of the funds return has a long left tail. The
mean kurtosis is 5.6, which is greater than 3, thus indicating that
the distribution of the funds return is peaked relative to normal.
Secondly, most of the Australian hedge funds charge a manage-
ment fee at 1.50% and an incentive fee at 20%. The high-water
mark contract is commonly used by the Australian hedge funds.
Thirdly, Australian hedge funds are divided in using leverage in
430
Table 17.3 Empirical results
Variable Coefcient StdErr t -stat p-value
Constant 0.4493 0.6674 0.6731 0.5035
StdDev
i
0.1793
0.0949 1.8893 0.0637

Auditor
i
0.3640
0.1981 1.8376 0.0711

Apaudit
i
0.2355 0.2670 0.8821 0.3812
MFee
i
0.2822
0.1444 1.9541 0.0554

IFee
i
0.0057 0.0207 0.2743 0.7848
HighWaterMark
i
0.0001 0.2004 0.0006 0.9995
Leverage
i
0.1208 0.1946 0.6208 0.5371
Lockup
i
0.0161 0.0240 0.6688 0.5062
Notice
i
0.0047 0.0045 1.0514 0.2973
Adjusted R
2
value is 0.06; DurbinWatson statistic is 1.59; Hannan
Quinn criterion is 2.79. Results are estimates using Equation 17.1. Using
the ordinary least-squares formula with NeweyWest heteroscedasticity-
consistent standard errors and covariance with the lag truncation equals 3.
Ret
i
is the average return over the life of the fund. StdDev
i
is the stan-
dard deviation of the funds returns. Auditor
i
is equal to 1 if the fund is
audited by one of the Big Four rms. Apaudit
i
is equal to 1 if the fund has
an annual performance audit. MFee
i
is the management fee, reported
in percent. IFee
i
is the incentive fee, reported in percent. Leverage
i
and HighWaterMark
i
are equal to 1 if the fund uses leverage or has a
high-water mark, respectively. Lockup
i
is the lock-up period measured
in months. Notice
i
is the advance notice measured in days.

indicates
signicance at the 10% level.
their investment strategies, and the lock-up period varies between
0 and 12 months.
The cross-sectional regression results are reported in Table 17.3.
The standard ordinary least-squares formula is estimated by using
the NeweyWest heteroscedasticity-consistent standard errors and
covariancewiththelagtruncationequal to3. TheadjustedR-squared
value is less than 0.1, which indicates the explanatory power of fund
characteristics on the hedge fund performance is rather low. How-
ever, hypothesis (H1) is partly supported because the coefcient of
Auditor
i
is positive and signicant at the 10% level. It indicates that
Australian hedge funds audited by a Big Four rm are exposed to
lower operational risk and therefore offer better return than those
audited by non-Big Four rms. The annual performance auditing
does not enhance the fundperformance, as the coefcient is negative
and insignicant. We do not nd any support for hypothesis (H2),
as the coefcients of IFee
i
and HighWaterMark
i
are insignicant,
431
MODEL RISK
while the coefcient of MFee
i
is negative. Hypothesis (H3) is also
not supported in the Australian hedge fund market as none of the
coefcients of Leverage
i
, Lockup
i
and Notice
i
is positive and sig-
nicant. The standard deviation is used as a control variable and is
signicantly positive.
CONCLUSION
The hedge fund industry has enjoyed tremendous growth over the
past few years. The number of hedge fund failures also increased
dramatically in the occurrence of the global nancial crisis and short
selling ban. Many of the hedge fund failures are attributed to oper-
ational risks such as misrepresentation of fund investments, misap-
propriation of investor funds, unauthorised trading and inadequate
resources, etc.
In this chapter we have investigated the legal structure and risk
disclosure of Australian hedge funds and modied Brown et als
(2008) model to examine the relationship with the operational risk
and hedge fund performance in the Australian hedge fund industry.
We have found that the hedge funds audited by Big Four rms are
exposed to less operational risk and therefore provide higher net-
of-fee returns. Other fund characteristics, such as fees, high-water
mark, leverage and lock-up period, do not have strong explanatory
power on the hedge fund performance. The results of this study will
help market participants to better understand operational risk expo-
sures and hedge fund performance, which should reduce investor
losses as well as limit risk to nancial markets.
1 In Australia, the Limited Partnership is governed by Part III of the Partnership Act 1958.
2 The two mainjurisdictions to establishanoffshore hedge fundare the CaymanIslands andthe
BVI. The attractionof offshore funds inthese locations stems fromtheir provisionof regulated,
tax neutral and exible investment vehicles, allowing for capital fromvarious countries to be
pooled together for deployment.
3 Other guidelines that Australian hedge funds are recommended to follow include the AIMA
Guide to Business Continuity Management for Hedge Fund Managers, the AIMA Guide to
Sound Practices for European Hedge Fund Managers, the AIMA Guide to Sound Practices
for Hedge Fund Valuation, the AIMAOffshore Alternative Fund Directors Guide, the Hedge
FundStandards Board(HFSB) HedgeFunds Standards (Final Report), International Organiza-
tionof Securities Commissions (IOSCO) Principles for the Valuationof Hedge FundPortfolios
and the Managed Fund Association (MFA) Sound Practices for Hedge Fund Managers.
4 Many hedge fund managers structure the offer of securities so that a PDS or other formal
disclosure document under the Corporations Act 2001 is not required. In these cases, issuers
of interests in hedge funds usually provide a document, known as an information memo-
randum.
432
5 OnDecember 2, 2004, the SECadopteda newrule andrule amendments under the Investment
AdviserAct of 1940that wouldrequirehedgefundmanagers toregister as investment advisers
by February 1, 2006. To comply with the new requirements, hedge fund managers were
required to le FormADVwith the SECand to comply with a variety of additional regulatory
requirements. Form ADV is a regulatory ling that is required of all types of fund manager,
including hedge fundmanagers, mutual fundmanagers andseparate account managers, who
fall under the denition of investment adviser in the Investment Advisers Act. However,
on June 23, 2006, the US Court of Appeals for the District of Columbia Circuit vacated the
rule changes that had required many newly registered hedge fund managers to register as
investment advisers under the Investment Advisers Act. Since the rule changes were vacated,
far fewer hedge fund managers have been required to register as investment advisors (Brown
et al 2008).
6 The Big Four accounting rms are Deloitte, Ernst & Young, KPMG and Pricewaterhouse-
Coopers. We use the termBig Four throughout the chapter to refer to the above set of large
international accounting rms.
REFERENCES
Agarwal, V., and N. Y. Naik, 2000, Multi-Period Performance Persistence Analysis of
Hedge Funds, Journal of Financial and Quantitative Analysis 35, pp. 32742.
AIMA, 2007, AIMA Australias Guidelines to Risk Disclosure for Australian Hedge Funds
(Sydney: Alternative Investment Management Association Australia Chapter).
Australian Trade Commission, 2008, The Hedge Funds Industry in Australia, ACT
Report.
Brown, S. J., W. N., Goetzmann, R. G. Ibbotson and S. A. Ross, 1992, Survivorship Bias
in Performance Studies, Review of Financial Studies 5, pp. 55380.
Brown, S. J., W. N. Goetzmann and J. Park, 2001, Careers and Survival: Competition
and Risk in the Hedge Fund and CTAIndustry, Journal of Finance 56, pp. 186986.
Brown, S. J, W. N. Goetzmann, B. Liang and C. Schwarz, 2008, Mandatory Disclosure
and Operational Risk: Evidence from Hedge Fund Registration, Journal of Finance 63,
pp. 2785815.
Civelek, E., 2009, Hedge Fund Due Diligence, in J. M. Longo (ed.) Hedge Fund Alpha: A
Framework for Generating and Understanding Investment Performance (Hackensack, NJ: World
Scientic).
DeAngelo, L., 1981, Auditor Size and Auditor Quality, Journal of Accounting and
Economics 3, pp. 18399.
Dye, R. A., 1993, Auditing Standards, Legal Liability and Auditor Wealth, Journal of
Political Economy 101, pp. 887914.
Gentilini, A., 2009, Lipper Tass Asset Flow Report: Hedge Fund, First Quarter 2009,
URL: http://www.lipperweb.com.
Hedges IV, J. R., 2005, Hedges on Hedge Funds: How to Successfully Analyze and Select an
Investment (Hoboken, NJ: John Wiley & Sons).
Investment Company Institute, 2009, Worldwide Mutual Fund Assets and Flows, First
Quarter 2009, URL: http://www.ici.org.
Jaeger, R. A., 2002, All About Hedge Funds (New York: McGraw-Hill).
Kundro, C., and S. Feffer, 2004, Valuation Issues and Operational Risk in Hedge Funds,
Journal of Financial Transformation 10, pp. 417.
Liang, B., 2000, Hedge Funds: The Living and the Dead, Journal of Financial and
Quantitative Analysis 35, 30926.
433
MODEL RISK
Lowenstein, R., 2001, When Genius Failed: The Rise and Fall of Long-TermCapital Management
(New York: Random House).
Luo, R. H., 2009, The Choice of Auditor and Investor Protection in the Investment Funds
Industry, Working Paper, La Trobe University.
Panageas, S., and M. M. Westereld, 2009, High-Water Marks: High Risk Appetites?
Convex Compensation, Long Horizons, and Portfolio Choice, Journal of Finance 64, pp. 1
36.
Rathore, S., 2009, Risk Management for Hedge Funds, in J. M. Longo (ed.), Hedge Fund
Alpha: A Framework for Generating and Understanding Investment Performance (Hackensack,
NJ: World Scientic).
Reserve Bank of Australia, 2004, Financial Stability Review: September, Report,
Reserve Bank of Australia.
Stulz, R. M., 2007, Hedge Funds: Past, Present, and Future, Journal of Economic Perspect-
ives 21(2), pp. 17594.
434
Part V
Risk Transfer and
Securitisation Models
18
Identication and Classication of
Model Risks in Counterparty
Credit Risk Measurement Systems
Marcus R. W. Martin
University of Applied Sciences, Darmstadt
Modellingcounterpartyexposure of over-the-counter (OTC) deriva-
tive portfolios and securities nancing transactions is a challenging
task at the borderline between market and credit-risk modelling.
For this purpose, so-called counterparty exposure proles are gen-
erated under numerous assumptions regarding the future dynamics
of asset prices and simplifying pricing models.
Counterparty exposure proles contain information about how
much of, say, a nancial institutions money is exposed to a certain
counterparty at a certain point in time. If the market value of all
outstanding contracts, eg, OTC derivatives or securities nancing
transactions, is positive, a default of the counterparty at a particular
point in time will lead to a close out of these positions, producing a
loss. Since the exposure is merely the positive market value of the
(possibly netted) counterparty portfolio, it depends on the market
environment at that particular time. Therefore, in contrast to classi-
cal credit exposures (of, say, loans), a direct modelling of potential
future market scenarios under which all contracts must be revalued
is necessary. Thereafter, according to applicable margin agreements
andcredit support annexes (CSAs), the resulting market values have
to be aggregated (in the respective netting sets) to come up with the
overall counterparty exposure at that particular point in time.
Consequently, counterparty exposure proles are generated in
three steps, whichwe followalongthe same lines as PykhtinandZhu
(2006); a somewhat rened viewis given by Tang and Li (2007), who
alsodeal withcredit valuationadjustments (CVAs), ie, incorporating
437
MODEL RISK
Figure 18.1 Sample paths for EuroStoxx50 generated by a GBM ESG
500
1,000
1,500
2,000
2,500
3,000
3,500
4,000
4,500
5,000
0
0
.
0
0
0
.
3
3
0
.
6
7
1
.
0
0
1
.
3
3
1
.
6
7
2
.
0
0
2
.
3
3
2
.
6
7
3
.
0
0
3
.
3
3
3
.
6
7
4
.
0
0
4
.
3
3
4
.
6
7
5
.
0
0
5
.
3
3
E
x
p
o
s
u
r
e
Time buckets
counterpartyriskintothe pricing, whichis not a subject we deal with
in this chapter.
Here, we discuss the model risks inherent in measuring counter-
party credit risk. In particular, we focus on those model risks that
arise in the rst two steps (scenario-generation and instrument pric-
ing) indetail, while anin-depthanalysis of those models risks arising
in the third step (aggregation, which is, in particular, more complex
for collateralisedexposures) will be dealt withinfuture publications.
Therefore, in this chapter we aim to identify the different sources of
model risk and try to classify and identify potential issues of these
risks for counterparty exposure measurement.
BASIC INGREDIENTS OF A COUNTERPARTY CREDIT-RISK
MEASUREMENT SYSTEM
A counterparty credit-risk measurement system (CRMS) usually
consists of three main components for calculating a total exposure
prole per counterparty portfolio.
First, an economic scenario generator (ESG) is used to produce
future realisations of the risk factors relevant for the institutions
positions on a certain set of dates dened by a pre-specied time
grid. For example, Figure 18.1 shows six sample paths generated
using geometric Brownian motion (GBM) calibrated to two years of
438
MODEL RISKS IN COUNTERPARTY CREDIT RISK MEASUREMENT SYSTEMS
Figure 18.2 Sample paths for a call option on EuroStoxx50
0
100
200
300
400
500
600
700
800
1,000
900
0
.
0
0
0
.
3
3
0
.
6
7
1
.
0
0
1
.
3
3
1
.
6
7
2
.
0
0
2
.
3
3
2
.
6
7
3
.
0
0
3
.
3
3
3
.
6
7
4
.
0
0
4
.
3
3
4
.
6
7
5
.
0
0
5
.
3
3
E
x
p
o
s
u
r
e
Time buckets
historical movements of the EuroStoxx50 in accordance with formu-
las (5.2) and (5.3) of Pykhtin and Zhu (2006) and used to forecast
the future behaviour of the index over ve years and four months
in time steps of equal length of four months.
Second, these risk factors are used to revalue the positions at cer-
tain future points in time to calculate the relevant exposure. Based
on the sample paths generated in Figure 18.1, a European call option
on the EuroStoxx50 with strike price 3.800 can easily be revalued
using the BlackScholes formula at every time step along each sam-
ple path. This yields market-value distributions in the future market
scenarios (along each path) as shown in Figure 18.2.
Finally, anaggregationandcollateral model is superposedtocome
up with an aggregated view on the total exposure of a counterparty
portfolio.
At the rst sub-step, the resulting market values of all positions in
the counterpartys portfolio for whichnetting is eligible (either regu-
latory or internally) will be netted, ie, its market value will simply be
added to get the market value of the netting set, while each position
for which netting is not eligible has to be treated as a single netting
set.
Counterparty credit exposure exists only when the market value of a
netting set is not negative. We still have to distinguish the following
two cases.
439
MODEL RISK
Figure 18.3 Counterparty exposure prole for a single uncollateralised
call option on EuroStoxx50
0
100
200
300
400
500
600
700
800
1,000
900
0
.
0
0
0
.
3
3
0
.
6
7
1
.
0
0
1
.
3
3
1
.
6
7
2
.
0
0
2
.
3
3
2
.
6
7
3
.
0
0
3
.
3
3
3
.
6
7
4
.
0
0
4
.
3
3
4
.
6
7
5
.
0
0
5
.
3
3
E
x
p
o
s
u
r
e
Time buckets
Figure 18.4 Portfolio of a European put and call on the EuroStoxx50
0
200
400
600
800
1,400
1,200
1,000
0
.
0
0
0
.
3
3
0
.
6
7
1
.
0
0
1
.
3
3
1
.
6
7
2
.
0
0
2
.
3
3
2
.
6
7
3
.
0
0
3
.
3
3
3
.
6
7
4
.
0
0
4
.
3
3
4
.
6
7
5
.
0
0
5
.
3
3
E
x
p
o
s
u
r
e
Time buckets
Netted (black diamonds) and unnetted (grey diamonds) counterparty risk exposure
envelopes are shown.
For uncollateralised exposure only, non-negative market val-
ues are replaced by a zero (exposure value) at each time grid
point. In the above example of the future market values of the
EuroStoxx50 option, the negative parts of the market-value
distribution per time step are cut off.
440
Instead of visualising all generated paths or scenarios per time
bucket, the resulting distribution of positive market values per
time grid point is usually described using quantiles (in most
cases the 5th percentile and 95th percentile) and the mean of
these market values. The resulting gure is usually called the
counterparty exposure prole (also called counterparty risk
cone or counterparty risk envelope by the industry), where
the 5th percentile and the 95th percentile are depicted in Fig-
ure 18.3 by thick black lines, while the mean of the market-
value distributions per time bucket is depicted as a thick grey
line.
Note that any eligible netting can help to mitigate the expo-
sure at this level (and is only dependent on the point at which
the positive part operator is applied). Consider a short Euro-
pean put position with the same trade characteristics (ie, the
same underlying, the same strike price andmaturity) for which
netting is eligible (black envelope) or not (grey envelope) as
depicted in Figure 18.4.
In the case of collateralised exposures, any collateral and mar-
gin agreements that are in place to reduce the exposure have
to be considered, which can lead to a diminished or even nega-
tive market value: the total exposure of the portfolio consisting
of OTC derivative or SFT positions including collateral is then
given as the maximum of this total market value and zero.
Therefore, the collateral market values also have to be calcu-
latedandthe marginmechanismis also to be incorporatedinto
the simulation, which dramatically increases the complexity
and workload of the whole process.
Summarising this step means that here all exposures are aggre-
gated according to netting set or counterparty level taking into
account any margin and collateral mechanisms, and a statistical
exposure measure is calculated, eg,
the mean of the future distributions of market values (which
yields the so-called expected exposure (EE), expected positive
exposure (EPE) or effective expected exposures (EEPEs)) or
a certain quantile (for the so-called potential future exposure
(PFE)
1
) as needed.
441
MODEL RISK
Figure 18.5 Three steps for generating exposure proles and
counterparty measures
IR term
structure
FX rates
Stocks/equity
indexes
Credit
spreads
Commodity
prices
...
Consisent
future
market
scenarios
Portfolio:
per counterparty,
per netting set,
per single trade
PFE
EE
EPE
EEE
EEPE
1
2
3
1. Correlated generation
of future market
scenarios
(ESG)
2. Revaluation of
all positions
per market scenario
and simulation date
3. Statistics/
aggregation/
reporting
Further insights, particularly with respect to credit-spread-driven

counterparty exposures and the regulatory framework for using
CRMSs according to the so-called internal model method conform-
ing to Basel II for measuring exposure at default (EAD) for OTC
derivatives, as well as securities nancing transactions, are provided
by the counterparty risk modellers bible (Pykhtin 2005), while we
will follow broadly the lines set up in Martin (2010).
Each of the three main components depicted in Figure 18.5 is sub-
ject to different aspects of model risk and, therefore, has to be treated
separately in the following discussion. For this purpose, we will
focus on the rst two components in this chapter, starting with short
descriptions of typical implementations for these following Pykhtin
and Zhu (2006).
Inparticular, calculationtime constraints have to be activelytaken
intoaccount since, incontrast toa single present value calculationfor
prot andloss (P&L) andmarket riskpurposes, all revaluations have
to be performed at any future time step to calculate a present value
(PV) at that time bucket, given the risk-factor scenarios produced by
the ESG.
Hence, in most cases it is still not possible to treat complex prod-
ucts (completely) in full detail in CRMSs because such products
would require Monte Carlo-on-Monte Carlo calculations.
442
For these reasons, model risk inherent to CRMSs is far more com-
plex than, for example, that in classical market risk models (which
is to some extent the single-time-bucket subset of this problem).
In order to analyse the different sources of model risk we start by
describing the respective steps in more detail.
Methods of generating future economic risk-factor scenarios
(economic scenario generation)
Potential market scenarios are generated in all relevant market risk
factors at different times in the future on a pre-specied time grid
of dates. This simulation-time grid is usually chosen to have daily
and, after a few weeks, weekly or monthly time steps in the short
term, while the grid points become sparser in the mid and long
term. The xed simulation-time grid can usually also be enriched
by additional, portfolio- or trade-specic time buckets. By this, risk-
relevant cashows can be incorporated according to the underlying
payment schedule(s) because certain cashows may signicantly
inuence the total balance of the counterparty exposure.
Typically, the evolution of interest rates, FX rates, equity prices,
credit spreads and commodity prices can be generated in various
ways; sometimes, factor models are used (eg, by modelling equity
indexes and using beta regressions for deriving single share prices),
while in other cases evolution dynamics similar to those used in
models for instrument pricing are used (eg, by choosing short rate
or Libor market models to model the evolution of the termstructure
of interest rates in the required currencies; cf. Pykhtin and Zhu 2006;
Hegre 2006; Beck 2010).
The dynamics of risk factors are in most cases based upon lognor-
mal or mean-reverting stochastic processes governed by Brownian
motions evolving under the empirical measure.
This modelling approach is in general in concordance with the
way market risk measurement systems, which are built under the
real measure based on historical data (and not necessarily con-
strained to a risk-neutral framework (see the next section)), are
designed. For example, in a front-ofce pricing model, the inter-
est rate scenarios are usually generated by bootstrapping zero
rates or discount factors using the market prices of government
bonds or swap rates. However, such a construction is in most cases
computationally too expensive for counterparty risk purposes and
443
MODEL RISK
the forward rates or forward volatility surfaces implied from this
method may be too sticky to mimic possible real-world changes
over the considered time horizon as a consequence of arbitrage-free
constraints.
Note, however, that the stochastic dynamics specied for market
risk measurement, which are constructed only for a short-termfore-
cast horizon, do not necessarily coincide with those designed for
counterparty risk measurement systems, which have to be built to
recognise long-dated future forecasts and risk-factor movements.
The dependence betweenthe evolvingriskfactors is usuallymod-
elled by correlating the driving stochastic processes governing the
single risk-factor evolutions. Typically, correlatedBrownianmotions
are based on empirical covariance matrixes which have been esti-
mated fromhistorical time series for the governing risk factors. This
is aimed at ensuring the consistency and reasonableness of (real)
market scenarios generatedby the different stochastic processes (see
the next section (page 446 onwards) for a more in-depth discussion
of this subject).
Infact, the market factors may jumprather thanevolving continu-
ously; thus, we mayconsider usingjumpdiffusionor Levyprocesses
in certain circumstances. This is of particular importance also for
specifying the interplay between credit spread and rating migration
dynamics, since empirically measured ratings transition matrixes
do not reect the speed of increase in risk-neutrally measured credit
spreads. This particular reason necessitates the explicit modelling
of credit spreads based on market-implied-spread levels and addi-
tional consistency adjustments of the empirical migration matrixes
(Tang and Li 2007; Hopper et al 2008).
Methods for pricing nancial instruments
The models and methods used for pricing nancial instruments are
in principle the same as those for calculating market risk. But, due
to the problem of revaluing the whole OTC derivative or securi-
ties nancing portfolio at a large number of future simulation dates,
the possibility of running complex, computationally intensive and
time-consuming algorithms is rather limited. In contrast to the pric-
ing for P&L purposes, computationally intensive models have to be
replaced by analytical approximations or simplied models which
might be insignicantly less accurate but signicantly more efcient
(Gregory 2008).
444
By using a Brownian-bridge technique the so-called condi-
tional valuation approach by Lomibao and Zhu (2005) even path-
dependent exotics (of the rst generation as, for example, barrier
options, Asian options or swap-settled swaptions) can be revalued
in an efcient manner consistent with the paths generated under
the economic scenario generators and the plain vanilla derivatives
priced based on these scenarios.
Even modern techniques used for modelling the volatility sur-
face for market risk pricing purposes may exhibit a parameterisation
robust enough to be used to forecast longer time horizons (see, for
example, the empirical analysis performed by Rebonato et al (2009)
for a SABR/Libor market model). In this case, we are at the border-
line of consistent future market value generation and pricing based
directly in the model framework set up by the ESG.
SPECTRA OF MODEL RISKS IN CREDIT-RISK MEASUREMENT
SYSTEMS
As seen in the previous section we need to distinguish the three fun-
damental steps in generating counterparty exposure proles before
we can measure some kind of total model risk of a CRMS. According
to the above description of the measurement process, we dene the
following (in analogy with Martin (2010)).
Scenario-generation risks: those model risks (including all
aspects of specication, estimation and implementation risks
as dened in the well-known model risk theory (Derman 1990;
Gibson et al 1998; Rebonato 2005)) that occur by specifying the
evolution of risk factors and all relevant assumptions on the
choice of risk factors and time buckets, as well as generating
overall consistent future market scenarios.
Pricing risks: those model risks (including all aspects of spec-
ication, estimation and implementation risks) inherent to the
specic requirements on instrument model pricing routines
that are needed to evaluate the counterpartys positions in a
fast but adequate manner.
Aggregation risks: those model risks (including all aspects of
specication, estimation and implementation risks) occurring
in the process of aggregating the counterpartys exposures,
recognising any eligible netting agreements andmodelling the
445
MODEL RISK
margin process including collateral valuations, etc. We will
consider these risks in more detail in forthcoming publica-
tions and focus on pricing and scenario-generation risks in
this chapter (which broadly extends Martin (2010)).
It is important to note that any model risks inherent in bilateral
counterparty risk valuation models or credit-valuation-adjustment
(CVA) measurement routines
2
will not be covered by this chapter
and merit further research. The corresponding model risks might
be classied as specic model risks inherent to pricing and hedging
collateral risk on a single trade level for the time being
Sources of model risk in economic scenario generation
The overwhelming amount of model risk dominating the scenario-
generation risk is the specication risk of risk-factor evolution
dynamics in scenario-generation: including the choice of risk fac-
tors and adequate future simulation dates, the stochastic dynamics
used to model interest and FX rates, stock and index prices, credit
spreads, rating migrations, commodity futures, and so on, directly
inuences the overall level of the counterparty risk measure.
For example, using lognormal processes to describe the evolution
of interest rate or FXrates might lead to unrealistically large or small
future rates and, consequently, to an overestimation of counterparty
exposures resulting from instruments sensitive to these risk factors
(see, for example, Reimers and Zerbs 1999; Zerbs 1999).
On the other hand, modelling approaches that are too simplistic
might lead to an underestimation. Take the case, for example, when
a termstructure model wouldonly be able to generate future market
scenarios which are due to parallel shifts of todays term structure
or when an important (class of) risk factor is not even stochastically
modelled while being an important contributor to the total market
value and, hence, to the resulting counterparty risk. Consider the
European call option on the EuroStoxx50 from the examples in the
last section: while the scenarios were generated using an annualised
historic volatility, the revaluations were always performed using
the implied at-the-money volatility that was available at the trade
date. Assuming at least a deterministic term structure of volatil-
ity, by using an exponentially weighted time-to-maturity-adjusted
volatilityfor simplicity, yields a counterpartyexposure prole which
446
Figure 18.6 Stressing the at implied volatility assumption by a
deterministic time-dependence for illustrative purposes
0
500
1,000
1,500
2,000
2,500
0
.
0
0
0
.
3
3
0
.
6
7
1
.
0
0
1
.
3
3
1
.
6
7
2
.
0
0
2
.
3
3
2
.
6
7
3
.
0
0
3
.
3
3
3
.
6
7
4
.
0
0
4
.
3
3
4
.
6
7
5
.
0
0
5
.
3
3
E
x
p
o
s
u
r
e
Time buckets
shows not only a completely different shape but also a signicantly
higher exposure.
3
In reality, we have to consider the dynamics of the index volatil-
ity surface as an independent risk driver which has to be mod-
elled in a way correlated to the index dynamics. Clearly, this is one
of the most interesting challenges we face in the future develop-
ment of CRMSs. Nonetheless, a thorough validation process must
be set up to document and analyse all (simplifying) model assump-
tions on an ongoing basis (BCBS 2006). As an important example of
this point, it should be noted explicitly that a simplied modelling
might nonetheless also be considered as a starting point for further
research; for example, the use of a single stressed deterministic
future path of implied or compound-correlation skews of collat-
eralised debt obligation (CDO) index tranches avoids the difcul-
ties associated with correlating, eg, the normally distributed credit
spreads with the dynamics of the implied or compound correla-
tions skew. This enables a nancial institution to gain initial insight
into a complex product like CDO tranches, but should be transpar-
ently documentedandtestedin great detail in the validation process
because of a rather extreme model risk that the nancial institution
might face (see also the section below).
Martingale re-sampling methods proposed, for example, by Tang
andLi (2007) canbe of great importance for the generationof realistic
447
MODEL RISK
and consistent future market scenarios which often carry a further
portion of model risk. Given the idealised situation that we have
chosen perfectly the evolution dynamics for all relevant risk factors
as well as the simulation-time grid, we have to dene a way to gener-
ate a correlated scenario at each future simulation date. To do so, the
driving stochastic processes must be modelled in a correlated fash-
ion, which is an ambitious task because of the diversity of processes
needed to describe the risk-factor evolutions. The easiest way is to
assume a multivariate (log)normal distribution or normal copula
to specify this dependency, based on covariance matrixes derived
empirically. Both choices, a multivariate (log)normal distribution or
a normal copula alike, carry an enormous portion of model risk or,
more precisely, specication risk (see, for example, Frey et al (2001)).
But the consistency within a simulatedfuture market scenario has
to be checked further; for example, the well-known link between
credit spreads (or rating migrations) and stock prices implies that
a defaulting or close-to-defaulting corporate cannot have a high
equity share value, while the credit spread (or rating) indicates the
closeness-to-default and vice versa. Hence, consistency checks have
to be established to ensure that the simulated scenario makes sense
economically and, thus, to reduce the specication risks.
Unfortunately, this is still not the full picture for model risks that
have to be subsumed in scenario-generation risks: when it comes
to generating the actual market scenarios over a long time horizon,
it is essential to use continuous-time models (of path-dependent-
simulation or direct-jump-to-simulation-date type as proposed by
Pykhtin and Zhu (2006) and discussed in Pykhtin (2008)) or sophis-
ticated discretisation schemes to ensure the strong convergence of
the sampled scenario paths to the pre-specied risk-factor distribu-
tions (see Kloeden and Platen (1999) for an overview on different
methods). Clearly, these estimation risks will be of second order
compared with the overall noise in risk-factor evolutions on a long
simulation horizon for plain vanilla instruments, but it may become
of higher importance for more complex exotic deals with long-term
maturities.
Finally, any calibration of our ESGs (as well as any other market
risk model) based on historical data can certainly only reect that
information and the risk-factor behaviour observed in the past. For
example, during the only period of stressed compound correlations
448
and credit spreads prior to 2008 the so-called correlation crisis in
May 2005 compound correlation actually fell when credit spreads
rose quickly. On the other hand, during the 20089 credit crunch,
credit spreads as well as compound correlations increased substan-
tially, which is completely contradictory correlation behaviour not
previously observed in the markets. Therefore, we should always
think about the possibility of enriching the historical scenarios with
synthetic scenarios or hypothetical risk-factor correlations, at least
for stress-testing purposes.
Sources of model risk in pricing models
To a small extent the pricing risks are similar and in the same spirit
as in classical market risk modelling as far as the accuracy of pric-
ing models is considered as an isolated goal (see Bossy et al (2001),
Branger and Schlag (2004a,b), Hull and Suo (2001), Kerkhof et al
(2002), Rebonato (2005), Derman (1990), as well as Ammann and
Zimmermann(2002), whichis, tothe best of the authors knowledge,
the only study of pricing risk in the context of a CRMS).
Nevertheless, specic aspects of modelling the counterparty risks
should be discussed explicitly, since the choice of a certain instru-
ment model for counterparty risk measurement purposes is in
almost all cases a trade-off betweenaccuracyandcomputational ef-
ciency. This topic has not previously been treated in the academic
literature, so no measures are publicly available to support the risk
manager in their decision. Since the particular decision process has
to be taken into account to thoroughly quantify the model risks,
we can only provide a rather vague framework on the model risks
arising in this context.
In principle, a recalibration of each pricing model (or change to
the risk-neutral or pricing measure) would be necessary at every
future simulation date given the simulated real market scenario (ie,
the market scenario generated under the real or empirical measure).
Hence, a simple waytoassess the rst-order model riskis tocompare
the market-value distributions generatedunder the empirical or real
measure and under the risk-neutral measure (instead of performing
numerous recalibrations). The use of so-called stochastic deator
techniques which enable the pricing of nancial instruments under
the empirical measure is a relatively new development in actuarial
mathematics and needs further investigation for these purposes as
well.
449
MODEL RISK
Clearly, the adequacy of pricing models has also to be taken
into account. During the correlation crisis in May 2005 it turned
out that models for pricing CDO tranches as, for example, the
asymptotic analytic approximation pricing formula under a large
homogeneous pool assumption were too simplistic and no longer
able to capture the correlation skews (Patel 2005). During the
credit crunch, CDO tranche correlations approached and some-
times exceeded one which nally proved that the asymptotic ana-
lytic approximation under the large homogeneous pool assumption
should denitively not be used for risk, P&L or mark-to-market
calculations. Furthermore, the more involved models (eg, the ran-
dom factor loading model by Andersen and Sidenius (2004) or
stochastic factor models by Burtschell et al (2005a,b)) are currently
also unable to calibrate to the market under deterministic recov-
ery rates, as was typically done for simplicity prior to the Lehman
Brothers default (Li 2009). These drawbacks show that, in addi-
tion to accessibility and computational power, any (approximate)
pricing model used in CRMSs must still be able to t the market
prices over a sufciently long forecast period, which is usually also
reected in a robust calibration of the model parameters over that
period.
Another problem refers to the consistency of dynamics specied
in ESGs for the universe of tradeable plain vanilla instruments and
those stochastic evolution models used in pricing models, which
we can easily demonstrate by the following example. Assume that
a jump-diffusion dynamics is deemed necessary for the stochastic
evolution of a certain share. On the other hand, when it comes to
pricing an option on this share, a BlackScholes-type formula is
used which assumes a complete market and a lognormal distribu-
tion of the underlying asset. Hence, the evolutionary dynamics of
the option-pricing model and the simulated asset distribution of the
underlying are not consistent, which may lead to over- or underesti-
mation of risks (depending on the nancial institutions positions).
In this particular situation, even the Martingale re-sampling tech-
niques by Tang and Li (2007), which are very helpful elsewhere,
will not apply, so an empirical study on a set of benchmark trades
must be performed in order to gain a better idea of whether the
resulting effects have a conservative or progressive impact on the
counterparty exposure.
450
Aggregation of model risks
Overall, the two steps described above primarily describe the model
risks occurring in portfolios with uncollateralised trades only. Even
in this simplied case, the aggregation of scenario-generation risks
and pricing risks to an overall model risk is extremely difcult to
obtain.
A rather general but exible framework for assessing the model
risk of (general) pricing models was given by Cont (2006). In this
section we aim to adjust these requirements for a measure of model
uncertainty (which is equivalent to our term model risk used
throughout this chapter) for CRMSs.
(i) For liquidly traded plain vanilla instruments whose driving
risk factors are modelled stochastically in the CRMS, the price
is determined by the market within a bidask spread, ie,
there is no (market-risk) model uncertainty on the value of
a liquid instrument, which is a basic ingredient of simulation
of the future market scenarios. The modelling exercise there-
fore should emphasise the importance of these instruments,
and the adequacy of the corresponding risk-factor evolutions
which should be tested in-depth historically (in-sample and
out-of-sample).
(ii) Any measure of model uncertainty for CRMSs must take into
account the possibility of setting up (total or partial) hedging
strategies in a model-free way. If an instrument can be repli-
cated in a model-free way, then its value involves no model
uncertainty; if it can be partially hedged in a model-free way,
this should also reduce the model uncertainty on its value.
Hence, the CRMS should be designed to recognise these pos-
sibilities in order to reduce the computational burden and the
pricing model risks as well.
(iii) When some instruments (typically, European style call or put
options for short maturities and strikes near the money) are
available as liquidinstruments onthe market, they canbe used
as hedging instruments for more complex derivatives but also
carry important information on the risk-neutral counterparty
exposure of the underlying (as their market value is a measure
for the counterparty exposure of the plain vanilla underlying
for the time to maturity of the option). This information can
451
MODEL RISK
be used to challenge the corresponding risk-factor evolution
(in particular concerning the short-term at-the-money volatil-
ity assumptions for the driving risk factor(s)). This procedure
will ensure a consistent short-term prognosis, given todays
information, which is essential for prudent risk management
purposes.
(iv) If we intend to compare model uncertainty with other, more
common, measures of counterparty exposure risk of a port-
folio, the model uncertainty on the value of a portfolio today
as well as on a future simulation date should be expressed
in monetary units and normalised to make it comparable to
todays market value and the future exposure measure of the
portfolio.
(v) The consistency (mathematically and economically) of the
generated scenarios and techniques applied to revalue the
counterparty portfolio, netting sets or single trades must be
ensured by additional manual or automatic checks of the
simulated future market scenarios. This includes any estima-
tion techniques for assessing discrepancies in the dynamics of
risk-factor evolutions and pricing models.
Given these guidelines, a basic framework might be set up for
studying the model risk of counterparty risk measurement systems
following broadly the lines of Cont (2006). Athorough study of these
ideas has yet tobe carriedout, withsupportingempirical studies that
go far beyond the scope of this chapter.
Nevertheless, the model risk measurement process for counter-
party exposure measurement systems must be embedded into the
overall riskmanagement andriskcontrollingprocesses, whichare of
vital importance for the internal validation process of CRMSs. Cer-
tain additional processes should also be incorporated into the daily
market and counterparty risk management and measurement. For
example, we can compare the market values of all the instruments
which were delivered by daily P&Lcalculation processes with those
market values obtained from (possibly simplied) pricing models
in the CRMS. These differences could be used to set up deviation
limits, which might trigger an in-depth analysis of a single pricing
routine or the CRMS as a whole.
In market risk measurement systems the well-established back-
testing processes and methodologies usually also provide insights
452
intothe performance of pricingmodels andthe overall measurement
system. Until nowthere seems to have beenno comparable standard
for backtesting exposure measurement systems, which, again, are
harder to assess than market risk measurement models (due, in par-
ticular, to the extremely long forecast horizon, which in turn means
that the inuence of correlations and autocorrelations must not be
neglected).
CONCLUSION AND OUTLOOK
We have described a possible approach to measuring and managing
the model risks of CRMSs. In fact, this chapter provides a formal
but practical classication structure for identifying the sources of
model risks according to the process of generating future market-
value distributions and exposure measures.
This also points to several elds of potential future research. Until
now there has been no known aggregation scheme for measur-
ing the total effect of the different sources of model risks. Further-
more, empirical studies should be undertaken to tackle the distinct
sources in more detail. For example, the use of analytical approx-
imations in pricing of more complex OTC derivatives is computa-
tionally extremely attractive, although there is a trade-off between
pricing accuracy, computational efciency and the resulting model
risk which has to be analysed carefully (and should, in principle, be
part of the process of choosing a model).
Finally, the question of adequately modelling collateral processes
adds anadditional layer of complexity, andhence model risks, which
will be analysed and classied in forthcoming research.
The author thanks the anonymous referee for detailedinsights and
extremely valuable discussions on credit spread and credit index
correlation modelling.
1 For a discussionandconcise denitions of these counterparty exposure measures, see Pykhtin
and Zhu (2006) or Pykhtin (2005).
2 For an introduction to this work see Brigo and Capponi (2009), Brigo and Pallavicini (2008),
Brigo and Chourdakis (2008), Gregory (2009) or Crpey et al (2009).
3 Note that we used a rather extreme exponentially weighted deterministic volatility adjust-
ment. Nonetheless, there is a signicant effect on longer-dated options under real-world
conditions.
453
MODEL RISK
REFERENCES
Ammann, M., and H. Zimmermann, 2002, The Credit Model Risk of Interest Rate
Derivatives and Regulatory Implications, Working Paper, Swiss Institute of Banking and
Finance.
Andersen, L., and J. Sidenius, 2004, Extensions to the Gaussian Copula: Random
Recovery and Random Factor Loadings, The Journal of Credit Risk 1(1), pp. 2969.
BCBS, 2006, International Convergence of Capital Measurement and Capital Standards:
A Revised Framework, Report (Comprehensive Version), June, Basel Committee on
Banking Supervision.
Beck, T., 2010, Interest Rate Counterparty Credit Risk Modeling and Management,
Masters Thesis, University of Applied Sciences, Darmstadt.
Branger, N., and Ch. Schlag, 2004a, Model Risk: A Conceptual Framework for Risk
Measurement and Hedging, Working Paper, Goethe-Universitt Frankfurt, January.
Branger, N., and Ch. Schlag, 2004b, Can Tests Based on Option Hedging Errors Correctly
Identify Volatility Risk Premia?, Working Paper, Goethe-Universitt Frankfurt, May.
Brigo, D., and A. Capponi, 2009, Bilateral Counterparty Risk Valuation with Stochastic
Dynamical Models and Application to Credit Default Swaps, Working Paper, SSRN,
January.
Brigo, D., and K. Chourdakis, 2008, Counterparty Risk Valuation for Energy Commodi-
ties Swaps: Impact of Volatilities and Correlation, Quantitative Research Special Report,
June.
Brigo, D., and A. Pallavicini, 2008, Counterparty Risk and Contingent CDS Valuation
under Correlation between Interest Rates and Defaults, Working Paper, SSRN, March.
Burtschell, X., J. Gregory and J.-P. Laurent, 2005a, A Comparative Analysis of CDO
Pricing Models, Working Paper, BNP Paribas, April.
Burtschell, X., J. Gregory and J.-P. Laurent, 2005b, Beyond the Gaussian Copula:
Stochastic and Local Correlation, Working Paper, BNP Paribas, October.
Cont, R., 2006, Model Uncertainty and Its Impact on the Pricing of Derivatives
Instruments, Mathematical Finance 16(3), pp. 51947.
Crpey, S., M. Jeanblanc and B. Zargari, 2009, Counterparty Risk on a CDS in a Markov
Chain Copula Model with Joint Defaults, Working Paper, May.
Derman, E., 1990, Model Risk, Report, Goldman Sachs Publications.
Frey, R., A. J. McNeil and M. Nyfeler, 2001, Copulas and Credit Models Risk, October,
pp. 1114.
Gibson, R., F.-S. LHabitant, N. Pistre and D. Talay, 1998, Interest Rate Model Risk: What
Are We Talking About?, Working Paper, RiskLab, Zrich.
Gregory, J., 2008, Credit Risk Models and the Sub-Prime Crisis: Rating Agency Models,
Structured Credit and CDOPricing, Working Paper (presented at PRMIALondon Credit
Risk Symposium), February.
Gregory, J., 2009, Being Two-Faced over Counterparty Credit Risk, Risk, February.
Hegre, H., 2006, Interest Rate Modeling with Applications to Counterparty Risk,
Masters Thesis, Norwegian University of Science and Technology.
454
Hopper, G. P., M. R. Smith and J. Ovens, 2008, Measuring Counterparty Exposures:
Lessons Learned in Estimating Potential Future Exposures, Working Paper (presented
at International Association of Credit Portfolio Managers (IACPM) Annual Fall Meeting,
Toronto), November.
Hull, J., and W. Suo, 2001, AMethodology for Assessing Model Risk and Its Application
to the Implied Volatility Function Model, Journal of Financial and Quantitative Analysis
37(2), pp. 297318.
Kerkhof, J., B. Melenberg and H. Schumacher, 2002, Model Risk and Regulatory
Capital, Working Paper, Tilburg University, April.
Kloeden, P. E., and E. Platen, 1999, Numerical Solution of Stochastic Differential Equations,
Stochastic Modelling and Applied Probability, Volume 23, Third Edition (Springer).
Li, Y., 2009, A Dynamic Correlation Modelling Framework with Consistent Stochastic
Recovery, Working Paper, Barclays Capital, February.
Lomibao, D., and S. Zhu, 2005, AConditional Valuation Approach for Path-Dependent
Instruments, Working Paper, Capital Market Risk Management, Bank of America.
Martin, M. R. W., 2010, Model Risk in Counterparty Exposure Modeling, in Ch. Hoppe,
G. N. Gregoriou and C. S. Wehn (eds), Model Risk Evaluation Handbook (McGraw-Hill).
Patel, N., 2005, Crisis of Correlation Risk, June, pp. 468.
Pykhtin, M. (ed.), 2005, Counterparty Credit Risk Modelling (London: Risk Books).
Pykhtin, M., 2008, Pricing Counterparty Credit Risk for OTC Derivative Transactions,
Working Paper (presented at GARP Ninth Annual Risk Management Convention and
Exhibition, New York), February.
Pykhtin, M., andS. Zhu, 2006, MeasuringCounterpartyCredit Riskfor TradingProducts
under Basel II, Working Paper, Risk Architecture, Bank of America.
Rebonato, R., 2005, Theory and Practice of Model Risk Management, Report, Quan-
titative Research Centre (QUARC) of the Royal Bank of Scotland and Oxford Financial
Research Centre, URL: http://www.riccardorebonato.co.uk/papers/ModelRisk.pdf.
Rebonato, R., K. McKay and R. White, 2009, The SABR/LIBOR Market Model (London:
John Wiley & Sons).
Reimers, M., and M. Zerbs, 1999, A Multi-factor Statistical Model for Interest Rates,
Algo Research Quarterly 2(3).
Tang, Y., and B. Li, 2007, Quantitative Analysis, Derivatives Modeling, and Trading Strategies:
In the Presence of Counterparty Credit Risk for the Fixed Income Markets (Singapore: World
Scientic).
Wilde, T., 2005, Analytic Methods for Portfolio Counterparty Credit Risk, in M. Pykhtin
(ed.), Counterparty Credit Risk Modelling (London: Risk Books).
Zerbs, M., 1999, Mark-to-Future in Practice, Algo Research Quarterly 2(2).
455
19
Quantifying Systematic Risks in a
Portfolio of Collateralised Debt
Obligations
Martin Donhauser, Alfred Hamerle,
Kilian Plank
University of Regensburg
Structured credit instruments played a signicant role at the heart
of the 20089 credit crisis. It seemed as if both practitioners and aca-
demics largely disregardedthe specic risk prole of portfolio credit
derivatives like collateralised debt obligations (CDOs), especially
their extremely high sensitivity to systematic risks. In this chapter,
we aim at quantifying the systematic risk portion of CDOs based
on several risk measures. One of these is a bond representation that
admits straightforwardrisk comparisons of CDOs andconventional
bonds. Based on this, we examine common pooling and structur-
ing patterns and nd that they are mostly boosting systematic risk.
Finally, we briey address the diversication myth associated with
CDOs and show that although being name diversied CDOs imply
high factor concentration risk.
INTRODUCTION
The burst of the US housing bubble triggered an unprecedented
depreciation of the whole asset class of credit securitisations. In sev-
eral waves the major rating agencies downgraded loads of asset-
backed securities (ABSs) and CDOs. One major problem was that
even best-rated AAAsecurities were downgraded, sometimes mul-
tiple notches (Moodys Investors Service 2008). This ratingtransition
behaviour is totally uncommon with conventional bonds. Further-
more, rating agencies did often point out the historical stability of
ratings for structured nance products (Fitch Ratings 2006; Moodys
Investors Service 2007). A legitimate ensuing question is whether
structured nance products like CDOs are intrinsically dangerous
457
MODEL RISK
or whether they just differ from defaultable bonds. This pressing
questionwas partlyansweredbycredit ratingagencies, whoempha-
sised that CDO ratings do not have the same meaning as classical
bondratings. But what exactly is the difference? Both are defaultable
xed-income securities but theydiffer markedlyinterms of their risk
prole.
A consideration of what CDOs actually are immediately reveals
the difference. ACDOinvestment is a bet that the default rates in the
underlying pool will exceed the subordination (which is just a safety
buffer) with a certain probability. Now, the type of default clustering
which is necessary for tranches with higher subordination to be hit
is only possible through systematic effects. CDO tranche risks are
primarily driven by systematic risks, while conventional bonds are
driven by both idiosyncratic and systematic components. Although
this fact is frequently stated in the literature (see, for example, Dufe
2008; Fender et al 2008), there are only few detailed analyses avail-
able. Specically, onlylittle knowledge exists onhowcertainpooling
and structuring patterns imply an increase of systematic risks, such
as the choice of collateral or the structure of issued tranches. This
issue was largely disregarded because of a business model called
arbitrage CDO. Arbitrage CDOs are vehicles earning money by
relatively cheap hedging of a pool of credit risk investments. As
we argue later, this excess spread may have been increased by
arbitrage of systematic risks.
Coval et al (2009a,b) and Brennan et al (2009) analyse aspects of
a possible mispricing of CDO tranches, caused by their increased
sensitivity to systematic risks. These authors argue that the price
investors pay for structured instruments is too high if their invest-
ment decision solely relies on the rating. According to Coval et al
(2009a) this is especially true for senior tranches, which default
just when the economy is in a very bad state. In contrast, the
results obtained by Brennan et al (2009) indicate that the AAA-rated
tranches are only marginally mispriced, and that the highest prots
can be gained with junior tranches. Eckner (2008) comes to similar
conclusions.
In this chapter we try to quantify CDO sensitivity to system-
atic risks. As mentioned in Hamerle et al (2008) and Krahnen and
Wilde (2009), the risk properties of ABSs and CDOs differ signif-
icantly from those of corporate bonds. We base our analyses on
458
QUANTIFYING SYSTEMATIC RISKS IN A PORTFOLIO OF COLLATERALISED DEBT OBLIGATIONS
conditional expected loss (ie, expected loss conditional upon the
market factor). Moving from benign to adverse states of the mar-
ket factor this measure clearly reveals that tranches have a much
higher sensitivity with respect to systematic factors than bonds
with a comparable rating. We estimate the asset correlation asso-
ciated with a CDO tranche treating the structured instrument as
a single-name credit instrument (ie, a loan equivalent). Yahalom
et al (2008) from Moodys point out that this tractable approach
requires appropriate parameterisation to achieve a reasonable
description of the cross-correlation between the structured instru-
ment and the rest of the portfolio. They provide an approach dif-
ferent fromours. For the determination of the correlation parameter
they estimate the joint default correlation of two CDOs loading on
the same factor and back out the asset correlation that is consis-
tent with this default correlation in the Gaussian single-risk-factor
model. The approach requires loss given default to be modelled
separately.
In this chapter we examine the systematic risk arising from typ-
ical pooling and tranching patterns. We provide a comprehensive
default risk assessment of structured credit products and show that
rst moments are insufcient to reect risk concentrations and sys-
tematic risk sensitivity. To measure the systematic risk and to com-
pare different xed-income products, we explain how to calibrate
the Gaussian single-risk-factor model to replicate the risk prole of
a tranche. It turns out that the resulting asset correlations of loan
equivalents of multi-name derivatives are substantially higher than
those of single-name instruments. Furthermore, as mentioned ear-
lier, buying assets with lower systematic risk and funding them
by issuance of assets with higher systematic risk offers arbitrage
gains. We showhowarrangers may utilise pooling and tranching to
increase arbitrage gains.
The chapter is organised as follows. In the next section we intro-
duce our model foundation. The section that follows presents clas-
sical risk measures for the centre and tails of a loss distribution.
After that, we describe the aforementioned bond representation of
CDO tranches. We then examine common pooling and structuring
practices against the background of systematic risk. In the penulti-
mate section we briey discuss the role of diversication with CDO
investments. The nal section gives concluding remarks.
459
MODEL RISK
SET-UP
This section focuses on presenting the model set-up that is used for
analysing the risk characteristics of CDOs.
CDO model
A CDO is structured like a balance sheet. In a typical cash CDO a
pool of assets (collateral pool) is fundedbyissuance of debt securities
(tranches). In a synthetic transaction a pool of long credit risk pro-
tection is hedged by issuance of short credit risk tranches. We shall
disregard any principal cashows here and focus on synthetic trans-
actions. Note that this does not imply any restrictions for our results.
We map the collateral pool of defaultable names by means of the
de facto standard model in practice: the Gaussian single-risk-factor
model.
Consider a portfolio comprising i = 1, . . . , N credit-risky names.
Each reference position is represented by the terminal asset value R
i
,
which is constructed as
R
i
=
M+
_
1
i
(19.1)
M represents the market factor common to all obligors, while
i
is
an individual innovation. The parameter is usually called asset
correlation. It controls the relative relevance of the two factors and
thus turns out to be the correlation of any pair (R
i
, R
i
), i i
. may
thus be interpreted as the degree of dependence of R
i
on M. All
names depend on M by the same level of magnitude, ie,
i
= for
all i. The factors M and
i
are standard normally distributed and
independent. Thus, R
i
is also standard normally distributed.
The default of obligor i is modelled as a threshold event, ie, when
R
i
falls short of a specic threshold c
i
the borrower is in default. This
is modelled using the default indicator
D
i
:= 1
R
i
c
i
(19.2)
where D
i
is unity if R
i
c
i
and zero otherwise. For later analyses
it is useful to note that is a measure of dependence of R
i
on M.
In connection with the threshold model its meaning is a little bit
different. With respect to D
i
we may interpret as a measure of
sensitivity of D
i
concerning M. This notion will be relevant in later
sections.
460
Now, given the unconditional probability of default
i
:= Pr(D
i
= 1) = Pr(R
i
c
i
)
this threshold can be backed out. So
c
i
=
1
(
i
) (19.3)
where () denotes the standard normal cumulative distribution
function.
The structure of R
i
establishes independence if conditioned on M.
This may be useful in computational terms. For example, given M,
the portfolio default rate is an N-fold convolution of independent
random variables with (conditional) probabilities of default
i
(M) = Pr(R
i
c
i
M) =
_
c
i
M
_
1
_
(19.4)
This conditional probability of default (PD) will be relevant below.
Note that, as with D
i
, also measures the degree of sensitivity of
(M) concerning M.
To obtain losses L from default counts or rates, we additionally
need information regarding the loss given default (LGD) or equiv-
alently the recovery rate RR
i
= 1 LGD
i
. We dene LGD
i
as the
fractional loss of obligor is exposure at default, EAD
i
. It is assumed
to be non-stochastic.
Based on this, the loss rate in the collateral pool can be calculated
as
L =
N
_
i=1
w
i
LGD
i
D
i
(19.5)
where exposure weights are given by w
i
= EAD
i
/ EAD with total
exposure EAD =
N
i=1
EAD
i
.
Base case CDO
Next, we introduce a base case CDO conguration.
Collateral pool
The asset side of the base case CDO has a maturity of T = 5 years.
It comprises N = 100 corporate bonds rated BBB (
i
= = 3. 25%
for all i) with equal notional weight w
i
= w = 1/100. As assumed
above, all names haveuniformfactor dependency
i
= = 10%. The
recovery rate is also assumed to be homogeneous among obligors
and is set to a common level of RR
i
= RR = 40% so that loss given
default LGD = 60%. Table 19.1 summarises this conguration.
461
MODEL RISK
Table 19.1 Asset pool conguration
Parameter T N w RR LGD
Level 5 100 3.25% 10% 1/100 40% 60%
Table 19.2 Structure of liabilities
No. Tranche a (%) b (%) Rating
1 Equity 0.0 4.0
2 Junior mezzanine 4.0 6.5 B
3 Senior mezzanine 6.5 11.5 BBB
4 Senior 11.5 100.0 AAA
5 Pool 0.0 100.0
Liability structure
The liability side of the transaction is structured into four tranches.
Each tranche (tr) is specied by attachment point a and detachment
point b, respectively, with0 a < b 1. Table 19.2 shows the details.
The structure presented in Table 19.2 is based on the hitting prob-
abilities of the tranches (for a denition see page 463). Ratings of
tranches canbe relatedtohittingprobabilities, as ratings of corporate
bonds are related to default probabilities.
RISK MEASURES FOR COLLATERALISED DEBT
OBLIGATIONS
Our major interest in this chapter revolves around measuring the
objective (ie, real-world) risks of CDOs. In this section we analyse
rst- and second-order measures such as expected and unexpected
loss. As descriptive statistics always represent just a summary of
certain distributional characteristics, we have to consider several
measures to get a full picture. In the next three subsections we dene
these risk measures and then apply them to our sample CDO in the
fourth subsection.
Rating-based risk measures for CDOs
We start with the most relevant rst order measures, probability of
default (PD) and expected loss (EL). These are the foundations of
agency ratings (eg, of Moodys, Standard & Poors and Fitch).
462
Hitting probability
The probability of default of a tranche is often referred to as its hit-
ting probability. It is denedas the probability that a tranche incurs
losses
tr
= Pr(L > a) (19.6)
For example, Standard & Poors and Fitch determine ratings based
on this measure.
Tranche loss
Givencollateral pool loss rate L, the loss incurredbya certaintranche
(tr) = (a, b) with 0 a < b 1 is calculated as
L
tr
=
1
b a
[(L a) 1
a<Lb
+(b a) 1
L>b
]
=
_
_
0, L a
L a
b a
, a < L b
1, L > b
(19.7)
In other words, a tranche (a, b) absorbs only pool losses in excess of
a with limit b a. Thus, expected tranche loss is the expectation of
the tranche loss rate
E(L
tr
) = P(L > b) +
1
b a

_
b
a
(l a) dF
L
(l) (19.8)
where F
L
(l) denotes the cumulative distribution function of L.
Moodys determines ratings based on this measure.
Tranche loss severity
The loss given default of a tranche is the random variable
LGD
tr
= L
tr
L > a (19.9)
Its expectation is a common risk parameter
E(LGD
tr
) = E(L
tr
L > a) (19.10)
Tail measures
There is a large number of tail or downside risk measures like semi-
variance, value-at-risk or expected shortfall. For most of these mea-
sures a decomposition is available which allows the attribution of
the contribution of a single constituent to the overall measure.
463
MODEL RISK
Value-at-risk
The most widely used tail measure is value-at-risk (VaR). The VaRat
condence level (1 ) [0, 1], VaR
1
(L), is the (1 )-quantile
of the loss distribution:
VaR
1
(L) = minl [0, 1] : Pr(L l) 1 (19.11)
Losses of VaR
1
(L) or higher are dened to occur in 100%of all
loss scenarios.
Expected shortfall
Since VaR has several theoretical decits, the literature prefers
expected shortfall ES
1
(L). Expected shortfall at condence level
(1 ) is the expectation of loss above VaR
1
(L). It is given as
ES
1
(L) = E(L L > VaR
1
(L)) (19.12)
Analysis of variance
Our factor model with a common systematic variable M lends itself
to the analysis of variance and thus admits closer insights into the
relevance of systematic and idiosyncratic risk
V(L) = V[E(L M)] +E[V(L M)] (19.13)
That is, total risk as measured by variance can be decomposed into
the variance of conditional expectation and the expectation of con-
ditional variance. The former is an absolute measure of systematic
risk.
The relative importance of the two variance components can be
seen by division by V(L), ie
V[E(L M)]
V(L)
. , .
V
[E(LM)]
+
E[V(L M)]
V(L)
= 100% (19.14)
Later, we report the rst term (the systematic component) as well
as the total variance.
Results
Now we want to apply the risk measures introduced so far to the
sample base case CDO(see page 461) andcompare the tranches with
the whole collateral pool. The latter can be considered as a tranche
with attachment points (a, b) = (0, 1). The tranche structure was
464
Table 19.3 Results: CDO risk measures
Tranche 1 2 3 4 Pool
a 0 0.04 0.065 0.115 0
b 0.04 0.065 0.115 1 1
tr
0.8545250 0.1282040 0.0314680 0.0013850 0.8545250
E(L
tr
) 0.4354460 0.0643830 0.0091760 0.0000310 0.0195130
VaR
0.99
(L
tr
) 1 1 0.38 0 0.084
ES
0.99
(L
tr
) 1 1 0.7213280 0.0220020 0.1045280
V(L
tr
) 0.1087647 0.0448334 0.0051819 0.0000012 0.0003360
V[E(L
tr
M)] 0.0655638 0.0245024 0.0027281 0.0000006 0.0002247
V
[E(L
tr
M)] 0.6028040 0.5465210 0.5264680 0.4579700 0.6686360
introduced in Table 19.2 and the collateral pool conguration was
shown in Table 19.1. The results are given in Table 19.3.
The upper part of the table includes hitting probability, expected
loss, VaR and expected shortfall for each tranche and the whole
pool. The lower part contains the variances and their systematic
components.
Let us start our interpretation with hitting probability. Hitting
probabilities necessarily decrease when we move up the capital
structure. The equity tranche incurs losses in 85% of cases and the
pool as a whole must necessarily have the same value. The other
tranches have a signicantly lower rst loss risk, eg, the senior
tranche has only 13bp. This does not carry over to expected loss.
Clearly, the equity tranche always bears the highest level of expected
loss while portfolio EL is signicantly lower. We see that expected
loss decreases as we move upthe capital structure. The expectedloss
of our senior tranche is extremely low (3. 1 10
5
).
With respect to tail measures, we observe that VaR and ES are
unity for the two lowest tranches, which is due to a signicant mass
singularity at L
tr
= 100%. In other words, these two tranches have
a high wipe-out probability. For higher tranches both measures
become increasingly smaller. Note that for the senior tranche VaR
equals zero since the hitting probability is 0.13% and 0.13% < 1% =
. Here, we recognise a decit of quantile-based tail measures, as
they may be insensitive to the level of condence.
To summarise, tranches may have signicant tail risk as measured
by VaRand ES. There is a large likelihood of total loss for mezzanine
465
MODEL RISK
tranches in particular. Furthermore, we observe a leverage effect
betweenpool andtranches: junior tranches have signicantly higher
tail risks in comparison with the whole pool, while senior tranches
may have clearly lower tail risks. Note, however, that both VaR and
ES do not reect risk concentrations. For example, both the equity
tranche andthe junior mezzanine tranche seemto carry the same tail
risk, althoughthey differ signicantly interms of expectedloss. Sim-
ilarly, as we will outline later, although both tranches have different
sensitivity to the systematic risk factor M, both have the same VaR
and ES. Thus, both measures fail to reect systematic risk sensitivity.
Our last categoryof riskmeasures is variancebased. Total variance
V[L
tr
] and systematic variance V[E[L
tr
M]] are difcult to inter-
pret in terms of their absolute value. Nevertheless, they are appro-
priate for risk comparisons, eg, of different securities with identical
expected loss (see page 481).
SYSTEMATIC RISK CHARACTERISTICS OF
COLLATERALISED DEBT OBLIGATIONS
In the previous section we studied several risk measures. We found
that the tranches differ clearly in terms of their risk prole both from
each other and from the pool as a whole. This suggests that bonds
and tranches have intrinsically different risk proles. In this section,
we show how to replicate the risk prole of a tranche by means of a
standard bond model.
On the one hand, this offers additional insight into the systematic
risk characteristics of CDOs. It is often stated that classical bonds
and CDO tranches are not comparable and that the rating of the
latter may hide important risk components. A bond representation
provides an opportunity to compare these two asset classes directly
andadmits straightforwardintegrationof CDOtranches into a static
portfolio model.
Hitting-probability prole
In the next two sections we study the default behaviour of CDO
tranches. First, we calculate the conditional hitting probability of
a tranche given different values of M. We call this functional
dependence a hitting-probability prole (HP prole).
Figure 19.1 shows conditional hitting probabilities. Black points
relate to the mezzanine tranche of our base case and grey points
466
Figure 19.1 Hitting-probability proles of the BBB mezzanine tranche
and a BBB bond
6 4 2 0 2 4 6
0.0
0.2
0.4
0.6
0.8
1.0
Common factor
H
i
t
t
i
n
g
/

d
e
f
a
u
l
t

p
r
o
b
a
b
i
l
i
t
y
BBB bond
Senior mezzanine
tranche
relate to any of the collateral pool bonds. Note that although both
have a BBB rating, their risk proles differ signicantly.
Moving frompositive (good) factor levels to negative (bad) factor
levels (ie, from right to left) the curve of the tranche increases much
faster than that of the bond. Apositive economic environment (large
values of M) almost completely precludes a tranche hit, while for the
bond some default risk remains. In such a positive scenario, small
changes of Mhave little impact on a tranches hitting probability but
have a modest inuence on a bonds probability of default. Thus,
CDOs seem to be more stable against macroeconomic changes.
However, this conclusion is misleading. If the systematic risk fac-
tor M falls into a certain critical region (in our example above, this
region is at about M = 2), CDOs turn out to be extremely sensitive
to an economic downturn. Because the steepness of the sensitivity
curve is visibly greater in the critical region, even small changes in
M may lead to tremendous deterioration of credit quality. In even
worse economic conditions, a hitting event is almost certain.
Actually, ratings for structured nance assets are thought to be
more stable thancorporate bondratings. However, it is also acknow-
ledged that if rating migrations do occur, these changes are of a
greater order of magnitude with tranches (Jobst and de Servigny
467
MODEL RISK
Figure 19.2 EL proles of the BBB mezzanine tranche and a BBB bond
6 4 2 0 2 4 6
0.0
0.2
0.4
0.6
0.8
1.0
Common factor
E
x
p
e
c
t
e
d

l
o
s
s
BBB bond
Senior mezzanine
tranche
2007). For instance, while rating changes of corporate bonds occur
more frequently but only by one or two notches, rating changes of
CDO tranches seldom occur, but if the do so they are by multiple
notches (Moodys Investors Service 2008). In our opinion, the differ-
ent sensitivity regarding systematic risks can, at least to some extent,
explain this phenomenon which we have seen, for example, in the
20089 nancial crisis.
But the HP prole does not reect all relevant risk characteristics.
The LGD of a CDO is always a random variable that also depends
on the systematic risk factor M. Furthermore, loss distribution and
sensitivity to systematic risk depend heavily on the thickness of the
tranche. Both are disregarded when using an HP prole. An alter-
native representation comprising these aspects is the expected-loss
prole E(L
tr
M).
Expected tranche loss prole
The expected-loss prole (EL prole) is determined similarly to
the HP prole. We simulate a number of factor realisations and
corresponding losses. Then we collect losses with similar factor
realisations and calculate the mean.
Figure 19.2 shows the EL proles of the BBB mezzanine tranche
and also that of a BBB bond. Again the differences of both curves are
468
rather obvious. As with HP prole, the pace of transition from zero
to total loss is a measure of sensitivity to systematic risk. Moving
from right to left the tranche prole rises much more steeply than
the corporate bond prole.
Note that in our homogeneous case the EL prole of the whole
collateral pool is identical to that of any bond in the collateral pool.
This means that, in absolute value, the collateral pool and the bond
contain exactly the same systematic risk. However, both differ in
terms of idiosyncratic risk. While each individual bond still carries a
signicant amount of idiosyncratic risk, the pool as a whole is rather
name diversied, ie, its variance is mainly due to systematic factor
movements.
Note that the EL prole is also benecial for pricing considera-
tions. Among others, the price of a (defaultable) security depends
on its expected loss as well as on the amount of systematic risk.
The EL prole reects both. Two credit products with identical
EL proles are exposed to identical default and systematic risks
and therefore should realise the same market price (Hamerle et al
2009).
Bond representation
In a bond representation the CDO tranche is treated as a single-
name credit instrument (ie, a loan equivalent). The loan equivalent
approach requires appropriate parameterisation to achieve a rea-
sonable approximation of the tranches risk prole. We consider the
tranche as a virtual borrower for which the single-factor model
holds. Our rst objective is to estimate the virtual asset correla-
tion
tr
of the CDO tranche. Since the tranches risk prole is given
by the ELprole, our second objective is to approximate the ELpro-
le of the tranche (obtained via simulation) by the corresponding
EL prole of the virtual bond as accurately as possible. In general
we have the decomposition
E(L
tr
M) =
tr
(M) E(LGD
tr
M) (19.15)
In a rst approach, the HP prole
tr
(M) can be used in the bond
representation. Considering a CDO tranche as a virtual bond,
the conditional hitting probability is expressed in the single-factor
model as a function of M by
tr
(M) =
_
c
tr
_

tr
M
_
1
tr
_
(19.16)
469
MODEL RISK
The default threshold c
tr
can be calculated using the unconditional
hitting probability
tr
given in Equation 19.6.
The only remaining parameter in Equation 19.16 is the virtual
asset correlation
tr
. This parameter can nowbe determined in such
a way that the function assigned in Equation 19.6 approximates
the simulated HP prole as accurately as possible (Donhauser 2010;
Hamerle et al 2008).
Another approach stems from the rating agency Moodys (Ya-
halom et al 2008). It is based on the assumption of two CDO
tranches with collateral pools containing different assets having the
same characteristics concerning number and risks. Furthermore, it
is assumed that both tranches can be modelled using a single-factor
model with identical risk parameters. The virtual asset correla-
tion can then be determined using the simulated joint default prob-
ability and the bivariate normal distribution from the single-factor
model. Bothapproaches provide almost identical results (Donhauser
2010).
The main characteristic of both approaches is the goal of nding
(the risk parameters of) a virtual bond whose conditional hitting
probability matches the simulated HP prole as closely as possible.
To copy the EL prole using the above approaches we need the
conditional expectedLGDof the tranche (see Equation 19.15), which
can be calculated via stochastic simulation. An approximation using
a common function is not possible without further complexity and
the assumption of a constant expected LGDprovides an insufcient
approximationof the ELprole. Thenagain, the bondrepresentation
aims to be as basic as possible and comparable to approaches that
are used for modelling traditional single-name products.
Therefore, we introduce another approach. We start directly from
the EL prole on the left-hand side of Equation 19.15. Assuming a
constant

LGD
tr
, we look for a virtual bond whose conditional hit-
ting probability approximates the simulated ELprole as accurately
as possible.
Using Equation 19.15 the implied hitting probability of the
virtual bond conditional on the systematic risk factor is then
determined from
tr
(M) =
E(L
tr
M)
LGD
tr
(19.17)
470
The tranches implied unconditional hitting probability is given
by
tr
=
E(L
tr
)
LGD
tr
(19.18)
The unconditional expected tranche loss E(L
tr
) is also calculated in
the course of in the simulation and the default threshold c
tr
is
c
tr
=
1
(
tr
) (19.19)
In the next step, the tranche LGD,

LGD
tr
, is determined as the
maximum loss of the tranche.
The maximum loss of the collateral pool is given by
L
max
=
N
_
i=1
w
i
(1 RR
i
) (19.20)
whichreduces toL
max
= 1RR = LGDfor a homogeneous portfolio.
Based on this, the maximum tranche loss of a tranche with
attachment point a and detachment point b is
LGD
tr
=
_
_
min(L
max
, b) a
b a
if L
max
> a
0 otherwise
(19.21)
In general,

LGD
tr
= 1 for all tranches except senior or super-senior
tranches, ie,

LGD
tr
< 1 onlyfor the tranche withthe highest seniority
(b = 1). For the non-senior tranches, the implied HP prole of the
virtual bondis equal to the ELprole, while the ELprole is scaled
up for the senior tranche with

LGD
tr
< 1.
Finally, the virtual asset correlation
tr
is estimated by means
of optimisation
arg min

tr
_
K
_
k=1
[
tr
(m
k
)

tr
(m
k
)]
2

tr
[0, 1]
_
(19.22)
where
tr
(m
k
) =
_
1
(
tr
)
_

tr
m
k
_
1
tr
_
(19.23)
(m
k
)
K
k=1
is a sufciently accurate discretisation of the support of M
and

tr
(m
k
) aresimulatedimpliedconditional hittingprobabilities
evaluated at m
k
.
In summary, a CDO tranche is approximated by a virtual bond
in a single-factor model according to Equation 19.1, with virtual
471
MODEL RISK
probability of default

tr
, virtual asset correlation
tr
and

LGD
tr
as virtual LGD of the bond representation. The approach ensures
that the EL prole of the virtual bond resembles that of the sim-
ulated EL prole of the CDO tranche. As outlined above, this bond
representation based on the ELprole is a more appropriate approx-
imation of the default behaviour and risk prole of the CDOtranche
than an approximation based on the conditional hitting probability
(with xed LGD).
Results
Applying the procedure described in the last subsection to our
sample conguration yields the results presented in Table 19.4.
Several points are worthmentioning. First, the difference between
tr
and
tr
as reported in Table 19.3 derives from the fact that we t
to

tr
(M) instead of
tr
(M). Assuming the constant

LGD
tr
implies
that the hitting probability of the bond representation has to be
lower than the original hitting probability. For instance, the equity
tranche has
equity
= 0. 854 but

equity
= 0. 435. Next, we see that the
calibrated asset correlations of all tranches are higher than for the
pool as a whole. The highest correlation leverage can be observed
with mezzanine tranches, while it is more moderate for equity and
senior tranches. The reasonfor the latter is that equitytranches suffer
the majority of expected losses (which occur in both good and bad
times). On the other end of the capital spectrum, the senior tranche
carries the end of the pool-loss tail and is thus driven by unexpected
losses. Nevertheless, due to its usually large notional share (in our
case 100% 11. 5% = 88. 5%) its calibrated
tr
is lower than that of
the mezzanine tranches.
Now, howgoodis our approximation? To answer this we measure
the goodness of t by means of mean-squared error dened by
MSE =
1
K
K
_
k=1
[
tr
(m
k
)

tr
(m
k
)]
2
(19.24)
The results in Table 19.4 show that mezzanine tranches can be
better approximated than the lowest and highest tranche. This is
purely due to a lack of functional exibility in the Gaussian copula
model. Beingpoint symmetric about its inectionpoint the Gaussian
conditional PD model is perfectly able to reproduce the proles of
mezzanine tranches. Because of their position at the ends of the cap-
ital structure the equity as well as the senior tranche prole do not
472
Q
U
A
N
T
I
F
Y
I
N
G
S
Y
S
T
E
M
A
T
I
C
R
I
S
K
S
I
N
A
P
O
R
T
F
O
L
I
O
O
F
C
O
L
L
A
T
E
R
A
L
I
S
E
D
D
E
B
T
O
B
L
I
G
A
T
I
O
N
S
Table 19.4 Approximation results for the bond representation
Tranche a b

tr

LGD
tr

tr
tr MSE
Equity 0 0.04 0.435135 1 0.421919 0.005697 0.000231
Junior mezzanine 0.04 0.065 0.064186 1 0.739621 0.000802 0.000008
Senior mezzanine 0.065 0.115 0.009139 1 0.753105 0.000520 0.000006
Senior 0.115 1 0.000056 0.548023 0.321983 0.003523 0.002821
Pool 0 1 0.032490 0.6 0.100016 5.9010
6
3.2310
8
4
7
3
MODEL RISK
Figure 19.3 Goodness of t of approximated conditional expected loss
15 10 5
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Common factor
5
Common factor
I
m
p
l
i
e
d

h
i
t
t
i
n
g

p
r
o
b
a
b
i
l
i
t
y
I
m
p
l
i
e
d

h
i
t
t
i
n
g

p
r
o
b
a
b
i
l
i
t
y
0
15 10 5 10 5 0
(a) (b)
Circles, real conditional expected loss; lines, tted conditional expected loss (vir-
tual bonds and pool approximation). (a) Black circles, junior mezzanine tranche;
dark grey circles, senior mezzanine tranche; light grey circles, pool. (b) Black
circles, equity tranche; grey circles, senior tranche, lines denote virtual bonds.
fulll the symmetry criterion. Therefore, the goodness of t of these
tranches is less satisfactory. The pool is particularly well approxi-
mated. Figure 19.3 shows real (ie, collateral pool simulation based)
proles andttedproles. The graphs fully reect the results of MSE
calculation.
STRUCTURES BOOSTING SYSTEMATIC RISKS
In the last section we showed the leverage effect of tranching on sys-
tematic risk and provided a bond representation of CDO tranches.
In this section we examine the consequences on past pooling and
structuring practice.
Systematic risk and CDO pricing
CDOs transform collateral pool-loss distributions into new and dif-
ferent loss distributions. Risks are completely reallocated. The hit-
ting probability and expected loss of a large portion of the capi-
tal structure are distinctly lower than for an average collateral pool
asset. It is known that many investors based their investment deci-
sions mainly on the ratings of the CDOtranches. The ratings rely on
assessment either of hitting probabilities or of expectedlosses due to
default. However, measures of default probability or expected loss
do not take account of the states of the economy in which the losses
occur. But it is also well known that systematic risk is price relevant.
474
Therefore, depending on their exposure to systematic risks, securi-
ties with identical credit ratings can trade at very different prices
and command a wide range of yield spreads.
In the previous section we discussed the leverage effect of tranch-
ing on systematic risk and showed that the systematic risk of all
tranches rises dramatically. This indicates that investors should
receive much higher spreads for their investments in CDOs as a risk
premium than those which they are paid on corporate bonds with
identical credit ratings. However, this was not the case in the years
prior to inception of the nancial crisis (Brennan et al 2009). Due
to rising demand for structured credit products the spread differ-
ences between CDOs and corporate bonds with comparable ratings
narrowed. In some cases the spreads of corporate bonds were even
higher than the spreads of corresponding tranches. If investors are
guided solely by the tranches ratings in their valuation process and
ignore the increased systematic risk, there is a general way to CDO
arbitrage. In this case, it is advantageous for the CDO arranger to
put together the collateral pool in such a way that the tranches cre-
ate securities with high systematic risk. If the tranches can be sold
at prices similar to those of corporate bonds with same ratings, a
maximumprot potential for the arranger is implied (for details see
Hamerle et al 2009).
In the following we describe some possibilities of generating
tranches with high systematic risk. It comes as no surprise that pre-
cisely these types of transactions can be found in many CDOs issued
prior to the outbreak of the 20089 nancial crisis.
Collateral pool diversication
Afrequentlystatedbenet of CDOs orABSs is the fact that the invest-
ment is already diversied (Fitch Ratings 2008a). Pooling reduces
idiosyncratic loss variance and becomes more effective with increas-
ing pool size and increasing homogeneity of the pool constituents.
These are standard results from portfolio theory. What is question-
able, however, is whether a diversied investment is really a benet
to the investor. Our previous results suggest that the systematic risk
of tranches increases as pool diversication increases.
We want to examine this now. To this end, we compare tranche
risk measures of the N = 100 pool with those of an N = 1,000 pool as
well as withaninnitelyne-grainedpool (N ). Since increasing
475
M
O
D
E
L
R
I
S
K
Table 19.5 Risk measures for different portfolio sizes
N = 1,000
, ..
a 0.000 0.036 0.059 0.101 0.000
b 0.036 0.059 0.101 1.000 1.000
tr
0.9984310 0.1274960 0.0272250 0.0017100 0.9984310
E(L
tr
) 0.4923910 0.0643760 0.0091550 0.0000310 0.0195030
VaR
0.99
(L
tr
) 1 1 0.3604350 0 0.0738000
ES
0.99
(L
tr
) 1 1 0.6686230 0.0179260 0.0896710
V(L
tr
) 0.0829043 0.0315945 0.0028030 0.0000005 0.0002364
V[E(L
tr
M)] 0.0777922 0.0288771 0.0025666 0.0000004 0.0002252
V
[E(L
tr
M)] 0.9383375 0.9139929 0.9156945 0.8963110 0.9526432
tr
0.4924453 0.0643278 0.0091739 0.0000559 0.0325076

tr
0.4896021 0.9195403 0.8969803 0.3296629 0.0999921
LGD
tr
1 1 1 0.5551323 0.6000000
4
7
6
Q
U
A
N
T
I
F
Y
I
N
G
S
Y
S
T
E
M
A
T
I
C
R
I
S
K
S
I
N
A
P
O
R
T
F
O
L
I
O
O
F
C
O
L
L
A
T
E
R
A
L
I
S
E
D
D
E
B
T
O
B
L
I
G
A
T
I
O
N
S
Table 19.5 (Cont.) Risk measures for different portfolio sizes
N
, ..
a 0.000 0.035 0.058 0.099 0.000
b 0.035 0.058 0.099 1.000 1.000
tr
1 0.1270540 0.0270495 0.0017655 1
E(L
tr
) 0.5001338 0.0643764 0.0091725 0.0000306 0.0195000
VaR
0.99
(L
tr
) 1 1 0.3568718 0 0.0726457
ES
0.99
(L
tr
) 1 1 0.6843768 0.00330547 0.0890376
V(L
tr
) 0.0795542 0.0299154 0.0025724 0.0000004 0.0002247
V[E(L
tr
M)] 0.0795542 0.0299154 0.0025724 0.0000004 0.0002247
V
[E(L
tr
M)] 1 1 1 1 1
tr
0.5001338 0.0643764 0.0091725 0.0000550 0.0325

tr
0.4995689 0.9422924 0.9152650 0.3312429 0.1000
LGD
tr
1 1 1 0.5559412 0.6000
4
7
7
MODEL RISK
the number of names changes PDas well as EL(and thus the rating)
of each tranche, we have to modify the capital structure in order
to maintain comparability. The idea is that tranche comparisons are
only useful when they bear the same level of expected loss. Thus,
starting from the highest tranche, we change attachment points in
such a way that each tranche obtains the same expected loss E(L
tr
)
as the corresponding tranche in the base case with N = 100. The
equity tranche cannot be compared, since its attachment point is
xed (a = 0). The innitely ne-grained case is known as Vasicek
approximation and often renders computation of risk measures
signicantly easier. Table 19.5 shows the resulting risk measures.
First, the estimates of
tr
show a clear increase in comparison
with the base case, which means that tranche sensitivity is higher for
N = 1,000 and still higher for N . Our relative variance measure
additionally signals that all tranches have virtually no idiosyncratic
risk for N = 1,000. Almost the whole tranche loss variance is driven
by the systematic factor. For N , the variance is purely system-
atic. This, however, does not implicate a reduction of the investors
portfolio variance. Instead, it means higher risk concentration. In
other words, diversication in the pools underlying a CDO tranche
implies concentration risk in the investors portfolio. We refer the
reader to the next section (see page 481) as well as Hamerle and
Plank (2009) for a more detailed analysis of diversication and risk
concentration with CDOs.
ABS CDOs
In our previous analysis we have shown that CDO tranches bear
signicant systematic risk. Thus, we may hypothesise that CDOs
including tranches in the collateral pool have still higher systematic
risk exposure. Such double layer structures are known as CDOs of
ABSs, ABS CDOs or structured nance CDOs (SF CDOs).
We shall investigate this hypothesis now. Exact calculation of the
loss distribution and risk measures of ABS CDOs is usually burden-
some. Our bond approximation permits a simplied solution. First,
given a pool conguration we determine

tr
and
tr
of our tranches.
These are the so called inner CDOs that formthe collateral pool of
an outer CDO. Thus, our second step is to simulate a portfolio of
bonds with

tr
and
tr
. Based on this, we determine risk measures
for the capital structure of the outer CDO (the ABS CDO tranches).
478
Table 19.6 ABS CDO collateral pool composition
Parameter Mezzanine RMBS BBB Bond
Absolute amount 70 30
Notional share 70% 30%
Rating BBB BBB
PD 0.9139% 3.25%
75.31% 10%
LGD 100% 60%
Table 19.7 Outer CDO structure based on expected tranche loss
No. Tranche a (%) b (%)
1 Equity 0.0 2.0
2 Mezzanine 2.0 63.2
3 Senior 63.2 100.0
4 Pool 0.0 100.0
For our ABS CDO we chose a collateral pool with the following
composition as shown in Table 19.6.
We donot choose a pure RMBS pool, as market practice was tomix
bonds andRMBS(Bankfor International Settlements 2008). The asset
pools underlying the RMBS comprise 100 BBB bonds as shown in
the right-hand column. Thus, these bond types are used as collateral
pool for the inner CDOs and also as 30% of the collateral pool of the
outer CDO. However, bonds in the collateral pool of the outer CDO
are assumedtobe drivenbya separate, uncorrelatedfactor

M. Bonds
in the inner CDOs collateral pools are driven by M.
As above, the capital structure of the outer CDO is chosen top
down, so that expected tranche losses of senior and mezzanine
tranche equal those of the base case (Table 19.4). Thus, all tranches
except for the equity tranche of ABS CDO and base case CDO are
comparable. The structure is given in Table 19.7.
We omit a junior mezzanine tranche here. Interestingly, the senior
tranche demands much higher subordination to achieve base case
levels of expected loss. This in turn suggests extreme tails of the ABS
CDOloss distribution. As a result of such high senior subordination,
the mezzanine tranche is extremely large.
The resulting risk measures are shown in Table 19.8.
479
MODEL RISK
Table 19.8 Risk measures for the ABS CDO
Tranche Equity Mezzanine Senior Pool
a 0 0.02 0.632 0
b 0.02 0.632 1 1
tr
0.5892880 0.1026500 0.0003280 0.5892880
E(L
tr
) 0.3332090 0.0090900 0.0000330 0.0122400
VaR
0.99
(L
tr
) 1.0000000 0.2614380 0.0000000 0.1800000
ES
0.99
(L
tr
) 1.0000000 0.5038230 0.1005930 0.3295660
V(L
tr
) 0.1173862 0.0038131 0.0000155 0.0016823
V[E(L
tr
M)] 0.0581641 0.0036870 0.0000134 0.0016052
V
[E(L
tr
M)] 0.4954930 0.9669400 0.8652660 0.9542100
tr
0.3336270 0.0090386 0.0000490 0.0138822
LGD
tr
1.0000000 1.0000000 0.6739130 0.8800000

tr
0.4661735 0.8271813 0.5554627 0.5547004
We see clearly increased bond correlation estimates
tr
. Tail risk
measures of the senior tranche and pool are much higher than for
the base case. For instance, for the collateral pool ES is 0.10 for the
base case and 0.32 for the ABS CDO. For the mezzanine tranche ES
and VaRare lower, which is certainly due to the comparatively large
size. Finally, from V
[E(L
tr
M)] we see that both mezzanine and
senior tranche are almost exclusively driven by systematic risk.
As recently pointed out by Fitch Ratings (2008b), we may sum-
marise that ABS CDOs imply an even higher level of systematic risk
sensitivity. We ndthis fact intail measures andbondapproximation
parameters as well as variance measures.
Thin tranches
As we sawinthe last sections, systematic risk onthe asset side canbe
increased by both increasing the pool size and choosing assets with
higher systematic risk. In this subsection we turn to the liability side.
Donhauser (2010) shows the effect of subordination and tranche
width on their sensitivity to systematic risks in detail. In summary,
the subordination only affects a tranches probability to be hit
by losses. The longer the tranche is protected against losses in the
collateral pool (the higher the subordination of the tranche), the
smaller the resulting hitting probability will be. Reducing the attach-
ment point without changing the tranche width shifts the ELprole
480
towards better (more positive) realisations of the systematic risk fac-
tor. Thereby, the steepness of the prole is almost unaffected. In con-
trast to the degree of subordination, the tranche widthdoes affect the
systematic factor sensitivity. The smaller the tranche width, the big-
ger the contribution of a single hitting event becomes to the tranche
loss rate. So the slope of the prole increases with diminishing
tranche width.
In practice, senior tranches are very wide andmezzanine tranches
are usually thin. Very thin tranches (so-called tranchelets) have
become increasingly popular in recent times. As described by
Tavakoli (2008), sometimes a thinAAAtranche was cut at the bottom
of the senior tranche, making the superordinate tranche even more
safe. That is why it is called super senior. We are interested in how
these structurings affect the risk measures of the resulting tranches.
To this end we split the mezzanine tranche (a, b) = (0. 065, 0. 115)
into ve tranchelets with 1% width. Furthermore, we form an
additional 5% tranche at the bottom of the senior tranche.
Tables 19.9 and 19.10 show the risk measures for the original
tranches as well as the new (thinner) tranches.
Most importantly, from
tr
we see that the systematic sensitiv-
ity of the new tranches is signicantly higher. For the mezzanine
tranches it rises from 0.75 for the original mezzanine tranche to 0.8
on average for the tranchelets. The same holds true for the senior
tranches, where the original senior tranche has
tr
= 0. 32, while
the new senior tranche shows an asset correlation of
tr
= 0. 81. A
similar result is reected by our relative variance measure.
To summarise, we nd that systematic risk factor sensitivity
decreases with tranche width. This represents potential for a wider
funding gap.
DIVERSIFICATION AND CONCENTRATION OF RISK
So far, we have found that CDO tranches generally carry high sys-
tematic risk. We hypothesised that although the underlying pool of
a CDOmay be highly diversied, this does not apply to the investor
portfolio the CDO is part of. Indications for this proposition were
seen in the high asset correlation of the CDOs bond representation.
We want to elaborate on this further. To this end, we compare
four different homogeneous investment portfolios respectively con-
taining 100 and 200 similar securities: a pure bond portfolio, a pool
481
M
O
D
E
L
R
I
S
K
Table 19.9 Risk measures for thin mezzanine tranches
Original
mezzanine
Parameter tranche Mezzanine tranchelets
a 0.065 0.065 0.075 0.085 0.095 0.105
b 0.115 0.075 0.085 0.095 0.105 0.115
tr
0.0315800 0.0315800 0.0157320 0.0078520 0.0055700 0.0028320
E(L
tr
) 0.0091080 0.0210190 0.0120310 0.0066170 0.0037190 0.0021570
VaR
0.99
(L
tr
) 0.38 1 0.9 0 0 0
ES
0.99
(L
tr
) 0.7222750 1 1 0.8546870 0.6758710 0.7686790
V(L
tr
) 0.0051585 0.0183948 0.0106270 0.0060093 0.0033259 0.0019329
V[E(L
tr
M)] 0.0027287 0.0083803 0.0046377 0.0024539 0.0013101 0.0007258
V
[E(L
tr
M)] 0.5289730 0.4555780 0.4364090 0.4083530 0.3939260 0.3755060
tr
0.0092140 0.0212780 0.0121400 0.0067110 0.0037650 0.0021770
LGD
tr
1 1 1 1 1 1

tr
0.7517390 0.7966750 0.8084510 0.8184150 0.8247900 0.8313230
4
8
2
Table 19.10 Risk measures for thin senior tranche and super senior
tranche
Original New Super
senior senior senior
Parameter tranche tranche tranche
a 0.115 0.115 0.165
b 1 0.165 1
tr
0.0013810 0.0013810 0.0000940
E(L
tr
) 0.0000310 0.0005160 0.0000020
VaR
0.99
(L
tr
) 0.0000000 0.0000000 0.0000000
ES
0.99
(L
tr
) 0.0226460 0.3721510 0.0252260
V(L
tr
) 0.0000012 0.0002974 0.0000001
V[E(L
tr
M)] 0.0000006 0.0001264 <0.0000001
V
[E(L
tr
M)] 0.4579000 0.4251280 0.3825020
tr
0.0000570 0.0005140 0.0000050

tr
0.3209300 0.8083430 0.3867380
LGD
tr
0.5480230 1.0000000 0.5209580
of pro-rata bond-portfolio investments, a portfolio of CDO tranches
and a portfolio of ABS CDO tranches. We do this for two different
levels of expected loss, each linked to a rating of either BBB or AAA.
The attachment and detachment points are set to match the desired
level expected loss. Note that the rst four alternatives have equal
expected losses, as do the second four alternatives. The CDOs are
backed by BBB bonds and the ABS CDOs are backed by a pool of
mezzanine CDO tranches. Both CDOs andABS CDOs are modelled
as loanequivalents. Altogether, we compare eight different invest-
ment alternatives. Inthe followingall the investment alternatives are
described in detail.
1. The rst portfolio comprises BBB-rated corporate bonds. As
in the preceding sections, this goes along with a probability of
default of 3.25%, a loss severity that is set to LGD = 60% and
an assumed asset correlation of = 10%. Thus, the expected
loss is E(L) = 1. 95%.
2. The secondcase is a pro-rata investment intoa pool of homoge-
neous bond-portfolios. Each portfolio comprises 100 corporate
bonds withriskparameters as describedinalternative 1. So the
expected loss is E(L) = 1. 95% too.
483
MODEL RISK
3. The third alternative is an investment into a pool of mezzanine
CDO tranches with a = 6% and b = 8. 6%. The collateral pool
to which the tranche is related is composed identically to alter-
native 1. The tranches are modelled as loan equivalents, with
parameters being calibrated as shown in the approximation
section (see page 469). This leads to a virtual asset correlation
of
tr
= 77. 59%. By choosing the attachment point a = 6%, this
alternative has a virtual default probability of 1.95%. In con-
nection with the detachment point b = 8. 6% and the virtual
LGD of 100% the expected loss equals 1.95% as well.
4. The investment alternative number four is an investment into
a pool of BBB-rated ABS CDO tranches. Each of these second-
layer tranches is backed by 100 mezzanine CDO tranches as
presented in alternative 3. For this case we nd a virtual asset
correlation of
tr
= 97. 57%. To achieve expected loss neutral-
ity we choose a = 17. 5%, which provides an implied hitting
probability of

tr
= 1. 95%. With b = 46. 7%and

LGD
tr
= 100%
the expected loss is once more set to E(L) = 1. 95%.
5. The portfolio of investment alternative number ve comprises
AAA-rated corporate bonds. This is linked to a default proba-
bility of 0.15%. The LGDis set to LGD = 60%and the asset cor-
relationis assumedtobe = 10%. Thus, the resultingexpected
loss of holding this portfolio is E(L) = 0. 09%.
6. Equally, as for alternatives 1 and 2, in the sixth case we con-
sider anpro-ratainvestment intoapool of homogeneous bond-
portfolios (each including 100 corporate bond positions) that
are identically composed to those presented in alternative 5.
So the expected loss is E(L) = 0. 09% too.
7. The portfolio of investment alternative seven consists of AAA-
rated senior CDO tranches. The tranches are linked to the
collateral pool introduced in case 3. The implied parameters
for simulating this investment are a = 5. 52%, b = 100%,

tr
= 21. 34%,

tr
= 0. 156% and

LGD
tr
= 57. 66%. So the
expected loss of 0.09% is equal to those of alternatives 5 and 6.
Note that the tranches of cases 3 and 7 are linked to identical
collateral pools. The resulting structures are set just to gain the
desiredlevels of expectedloss. The overlapof the two tranches
484
therefore does not bother us because we do not look at two
tranches of the same capital structure here.
8. The last alternative is a multi-layer investment into a portfolio
of AAA-ratedABS CDOtranches, each backed by a pool of 100
mezzanine CDO tranches, as presented in alternative 3. The
expected loss of 0.09% is accomplished by setting a = 87. 1%,
b = 100%,
tr
= 95. 8%,

tr
= 0. 09% and

LGD
tr
= 100%.
We aim to compare alternatives 14 with alternatives 58. The
resulting risk measures of the portfolio alternatives canbe compared
in Table 19.11. We show the two different portfolio sizes, N = 100
and N = 200.
As the most important result, we nd that tail risk measures are
signicantly higher for CDOtranche portfolios than for bond pools.
For example, the VaR of alternatives 1 and 2 is about 8% but that for
alternatives 3and4is 49%and93%, respectively. The same holds true
for the high-grade alternatives (58). However, for alternative 8 and
N = 100 we again see the insensitivity issue as discussed earlier (ie,
the condence level is too low). Expected shortfall is more reliable
here and shows clearly increased values for the tranche portfolios.
Our variance measures agree with these ndings. They are consis-
tently higher for tranche portfolios. In addition, relative systematic
variance is very high for all portfolios except for the pure bond port-
folios 1 and 5. It is important to note that although these portfolios
seem highly diversied since V
[E(L M)] is close to unity for

alternatives 24 and 68, they bear different levels of systematic risk.
From the tail risk measures in connection with the relative variance
measure we see that alternatives 2 and 6, the pro-rata bond pool
investments, are name diversied as well as with moderate system-
atic risk, while 3, 4 as well as 7, 8 are name diversied but still carry
high systematic risk.
To summarise risk concentrations in tranche portfolios are signif-
icantly higher than in normal bond portfolios. Thus, CDO invest-
ments implicate that idiosyncratic risk is highly diversied but con-
centration risk is built up. A more detailed discussion of this topic
can be found in Hamerle and Plank (2009).
CONCLUSION
Inthis chapter we extensivelyexaminedsystematic risks withCDOs.
We compared risk measures of bonds and CDOs and showed that
485
M
O
D
E
L
R
I
S
K
Table 19.11 Risk measures of investment alternatives
N = 100
Alternative
, ..
1 2 3 4 5 6 7 8
E(L) (%) 1.95 1.95 1.95 1.95 0.09 0.09 0.09 0.09
VaR
0.99
(L) (%) 8.4 7.26 49 93 1.2 0.56 1.15 0
ES
0.99
(L) (%) 10.52 8.80 70.1 99 1.98 0.78 2.21 21.69
V(L) (%) 0.034 0.022 0.74 1.48 0.0007 0.0001 0.0010 0.056
V[E(L M)] (%) 0.022 0.022 0.73 1.48 0.0001 0.0001 0.0005 0.056
V
[E(L M)] (%) 66.90 100 98.41 99.72 19.89 100 47.19 99.39
N = 200
Alternative
, ..
1 2 3 4 5 6 7 8
E(L) (%) 1.95 1.95 1.95 1.95 0.09 0.09 0.09 0.09
VaR
0.99
(L) (%) 7.8 7.26 48.5 93.5 0.9 0.56 1.15 0
ES
0.99
(L) (%) 9.61 8.80 69.20 99.1 1.39 0.78 1.99 19.59
V(L) (%) 0.028 0.022 0.74 1.49 0.0004 0.0001 0.0007 0.056
V[E(L M)] (%) 0.022 0.022 0.73 1.49 0.0001 0.0001 0.0004 0.056
V
[E(L M)] (%) 80.19 100 99.19 99.86 33.11 100 64.05 99.69
4
8
6
tranching produces securities with high systematic risk. Further-
more, we investigated drivers of systematic risk with CDOs. On
the asset side, increasing the number of names and choosing assets
with high systematic risk increases the systematic risk of resulting
tranches. In particular, CDOs based on collateral pools comprising
CDO tranches are very sensitive to systematic factors. On the part
of the liability structure we showed that smaller tranches are more
systematic-risk sensitive. Finally, we briey addressed the myth
of the benet of CDO diversication. CDO investments are name
diversied but contribute signicantly to risk factor concentrations.
REFERENCES
Bankfor International Settlements, 2008, TheJoint ForumCredit RiskTransfer: Develop-
ments from2005to2007, Basel Committee onBankingSupervision, Bankfor International
Settlements.
Brennan, M. J., J. HeinandS.-H. Poon, 2009, Tranching andRating, Working Paper 567,
Manchester Business School.
Coval, J. D., J. W. Jurek and E. Stafford, 2009a, Economic Catastrophe Bonds, American
Economic Review 99(3), pp. 62866.
Coval, J. D., J. W. Jurek and E. Stafford, 2009b, The Economics of Structured Finance,
Journal of Economic Perspectives 23(1), pp. 325.
Donhauser, M., 2010, Risikoanalyse strukturierter Kreditprodukte, PhD Thesis, Uni-
versitt Regensburg.
Dufe, D., 2008, Innovations inCredit RiskTransfer: Implications for Financial Stability,
BIS Working Paper 255, Bank for International Settlements.
Eckner, A., 2008, Risk Premia in Structured Credit Derivatives, Working Paper, Stanford
University.
Fender, I., N. Tarashev and H. Zhu, 2008, Credit Fundamentals, Ratings and Value-At-
Risk: CDOs versus Corporate Exposures, BIS Quarterly Review 1, pp. 87101.
Fitch Ratings, 2006, Fitch Ratings 19912005 US Structured Finance Transition Study,
Fitch Ratings Credit Market Research (Structured Finance).
Fitch Ratings, 2008a, Global Rating Criteria for Corporate CDOs, Fitch Ratings
Structured Finance.
Fitch Ratings, 2008b, Global Rating Criteria for Structured Finance CDOs, Fitch Ratings
Structured Credit.
Hamerle, A., R. Jobst and H.-J. Schropp, 2008, CDOs versus Anleihen: Risikoprole im
Vergleich, Risiko-Manager 22, pp. 814.
Hamerle, A., T. Liebig and H.-J. Schropp, 2009, Systematic Risk of CDOs and CDO
Arbitrage, Deutsche Bundesbank Discussion Paper Series 2: Banking and Financial
Studies.
Hamerle, A., and K. Plank, 2009, Is Diversication Possible with CDOs? Promises and
Falacies of an Investment Class, Discussion Paper, Universitt Regensburg.
487
MODEL RISK
Jobst, N., and A. de Servigny, 2007, The Handbook of Structured Finance (McGraw-Hill).
Krahnen, J. P., and C. Wilde, 2009, CDOs and Systematic Risk: Why Bond Ratings Are
Inadequate, Discussion Paper, Goethe University Frankfurt.
Moodys Investors Service, 2007, Structured Finance Rating Transitions: 19832006,
Special Comment.
Moodys Investors Service, 2008, Structured Finance Rating Transitions: 19832007,
Special Comment.
Tavakoli, J. M., 2008, Structured Finance and Collateral Debt Obligations (Chichester: John
Wiley & Sons).
Yahalom, T., A. Levy and A. S. Kaplin, 2008, Modeling Correlation of Structured
Instruments in a Portfolio Setting, Moodys KMV.
488
Epilogue
The US mortgage crisis of 20079 occurred across all risk categories:
credit, operational, market andliquidity. Model failures were amajor
contributor in every category.
The seeds of this crisis were plantedin2005. As housingprices rose
and interest rates ticked higher, consumer appetite shifted. Agreater
percentage of scally conservative consumers withdrew from the
market, leaving mainly the nancially adventurous. Credit scores
cannot measure consumer psychology, just past performance. Con-
sequently, the deterioration in the quality of new loan originations
in 2005 and 2006 went largely unnoticed. This is a classic example
of adverse selection.
The credit-risk problems of the 2005 and 2006 vintages turned
into the operational-risk problems of 2007. Many experienced man-
agement teams overruled established lending guidelines, booking
any loan available on the mistaken premise that the lender would
bear no risk, because all the loans could be securitised. Indeed,
the securitisation market frenzy in 2007 reected this misguided
assumption. Investors withscarce informationandpoor models mis-
priced the residential mortgage-backed securities (RMBSs) pools.
When the market bubble collapsed in 2008, a dramatic liquidity cri-
sis occurred, sending the general US economy into the recession of
20089. The recession spread globally due not only to the domino
effect of the slowdown in the US economy, but in some cases to sim-
ilar originations practices and risk management problems around
the world.
Clearly many models failed and the risk in those models was not
understood.
The adverse selection in the 2005 and 2006 vintage went unno-
ticed by scores because scoring models are designed to rank order
the risk of being bad, not predict the probability. Traditional logistic
regression scores are blind to level shifts in risk. Similarly, the most
common loss-forecasting models in retail lending are roll-rate mod-
els. With these models, forecasts are usually made at the portfolio
level using a simple extrapolation of historic roll rates. When large
volumes of poor quality loans were being booked in 2005, the roll
489
MODEL RISK
rates actually went down. New loans are lower risk because of life-
cycle effects, so the roll-rate models were deceived just when they
were needed most.
However, in every phase of this crisis failure was not necessary.
Models were available that did not fail.
Those teams who usedrobust vintage-basedmodels suchas those
described in Reinventing Retail Lending Analytics, (Joseph L. Bree-
den, 2010) were able to detect the quality deterioration and produce
alarming forecasts. Unfortunately, even some organisations with
usable forecasts chose to ignore them, because portfolio managers
made the same mistaken assumptions that fuelled the poor origina-
tions: housing prices never fall, so the loan can always be recovered
and credit risk does not matter because all loans can be securitised.
No model can prevent poor judgement.
The most common defence of these portfolio failures was that no
one could have foreseen a nationwide collapse in housing prices.
Certainly, few economists were predicting such a collapse. Never-
theless, stress test models showed as early as 2005 that a simple
attening of housing prices would precipitate a mortgage crisis.
Deep economic insight was not required to see the problems in the
increasingly lax lending practices.
Modelling RMBSs pools was worse than modelling the portfolios
they came from, because investors have historically hadvery limited
information. Basic information about the current bureau score, the
product mix and recent delinquency rates were the extent of the
information. Consequently, whole loans were modelled either with
a simple score-to-odds calibration or with a roll-rate model. Neither
of these approaches is truly predictive, as is now clear.
The modellingof collateraliseddebt obligations andcredit default
swaps failedthe most spectacularly. Byrelyingonanefcient market
hypothesis that those trading whole loans know the true prices, the
CDO and CDS markets traded themselves into a dramatic bubble.
Copula models are now blamed for these failures, but a copula is
just a means of aggregating risk. In this case, the problem was not
the model but rather its implementation and interpretation.
Theensuingliquiditycrisis was theoneareawherefewgoodmod-
ellingalternatives exist. Noone foresawthe extent of the correlations
between mortgage lending, securitisation, liquidity and economic
collapse, although in hindsight the linkages are apparent. Modelling
490
EPILOGUE
liquidity risk and restructuring the nancial system regulations to
reduce those risks will take more research.
Thesurvivinglenders recognisethat scores androll rateextrapola-
tions are insufcient forecasting tools. Mortgage lenders are recog-
nising that having a loan backed by a home does not guarantee
against losses. Many companies are rushing to ll the information
void faced by those who tried to price RMBSs in the past.
Practitioners assume that model risk is something that can be
qualied and priced for and have capital held against. In fact, model
risk can be quantied in effective ways, but only if the models in
question fundamentally work. As we have seen, we cannot price
or hold capital against risk when the model measuring that risk is
fundamentally awed.
Furthermore, we must be careful about the methods and time
periods used to quantify model risk. From2003 through 2005, a roll-
rate model with simple moving average extrapolation would have
appearedto provide reasonable accuracy. Unfortunately, this was an
illusion. Roll-rate models andscore-odds calibrations cannot predict
turning points, so their accuracy during calm periods says nothing
about the likelihood or severity of their failures during transitions.
Any model that did not include explicit measures of the product life
cycle and environmental impacts could not have succeeded through
the recent crisis, so any prior estimates of their accuracy were illu-
sory. Models that are vintage or account level, that include life-cycle
effects and that accept explicit scenarios for the future environment
can be tested effectively during calm or volatile periods to quantify
accuracy and model risk.
Computing model risk for retail lending credit risk is feasible
whenthe appropriate models are inuse. For operational risk, market
risk, liquidity risk and other product categories, each model must
rst be examined structurally to determine if it is even possible for
it to capture turning points. Only after the model has been proven
to be structurally viable can we realistically quantify the risk within
the model.
Risk management has always been viewed as a cost. When times
are tough, risk management teams are the rst to be downsized. As
we have seen, under-resourced risk management teams implement
the cheap and easy models even when those models are known to
be awed. This paradigmmust be changed. Risk management must
491
MODEL RISK
be viewed as insurance against future disasters. Consumers with
mortgages are required to maintain insurance on their property to
protect against loss, but lenders often operate without insurance by
failing to maintain useful risk management models and practices.
Joseph Breeden
President, Strategic Analytics Inc.
492
Index
(page numbers in italic type relate to tables or gures)
A
Alternative Investment
Management Association
(AIMA) 4245
asset portfolio modelling,
diversied 15581, 157
dissimilar assets and,
combining loss
distributions for 1704
loss distributions and 1704
conditional versus
unconditional 172
outputs, reduction and
providing context for
1749
portfolio model versus meta
model, 175
risk sources and mitigants
15581
commercial exposures
15868
general overview 1568
loss distributions for
similar assets, combining
1704
mark-to-par algorithm
and, 166
model, internal alternative
to 1689
model, key features and
settings for 1601
model, recasting 1612
model, vended, operating
within constraints of 1636
recovery and 1668
value to loss 1623
asset return dynamics, general
equilibrium effect,
monetary policy and
93130
CCL (Chang et al) 2008 and,
econometric set-up of
10211
empirical results 10611
identication of regimes
105
stationarity of Markov
regime-switching model
1056
two-state Markov process
1045
CCL (Chang et al) 2009, focus
of 11125, 113, 114, 117
data 11315
econometric procedures
and 11517
estimation results 117
in-sample forecasting
11718, 118, 129
out-of-sample forecasting
11825, 119, 120, 121, 122,
123, 124, 125, 130
two asset returns,
necessity of forecasting
1258, 126, 127, 128
data, the economic model
and 1002
regime-switching SVAR
models and 97100
Australia, hedge fund industry
in 42132, 422, 431
descriptive statistics of 430
legal structure and risk
disclosure of 4246, 425
see also hedge fund
performance
B
behavioural scores 217, 2201,
222, 223, 226, 230
comparison of credit-risk
models for portfolios of
MODEL RISK
retail loans based on
20931
dynamics of, Markov chain
models 2239, 2279
explained 21214
proportional hazard models
21820, 2201, 222, 223
reputational structural model
21418
Buiter, Professor Willem 934
business cycles, credit risk and
47, 5
Business Week 93
C
Caymans Corporation 424
CCL (Chang et al) 2008,
10211
identication of regimes 105
1056
1045
CCL (Chang et al) 2009, focus of
11125, 113, 114, 117
data 11315
econometric procedures and
11517
in-sample forecasting 11718,
118, 129
11825, 119, 120, 121, 122,
123, 124, 125, 130
two asset returns, necessity of
forecasting 1258, 126, 127,
128
collateralised debt obligations
(CDOs) 447, 450, 45787
ABS 47880, 479, 480
asset-pool conguration and
462
base-case conguration of
4612
boosting systematic risks
with 45787, 4767
CDO pricing and 4745
collateral pool
diversication and 4758
diversication and
concentration of risk and
4815
liability structure and 462, 462
model 4601
outer structure 479
pricing, systematic risk and
4745
risk measures for 4626
analysis of variance and
464
rating-based 4623
results 4646, 465
tail 4634
set-up to analyse risk
characteristics of 4602
model 4601
systematic risk characteristics
of 46674
bond representation
46972
expected tranche-loss
prole 468, 4689
hitting-probability prole
4668, 467
results 4724, 473
thin tranches and 4801, 482,
483
tranche loss and 463, 479
counterparty credit risk
measurement systems 438,
439, 440, 442
aggregation of 4513
basic ingredients 43845
economic scenario
generation and 4434,
4469
nancial instruments and,
methods for pricing 4445
identication and
classication of 43753
spectra of model risks in
44553
494
INDEX
credit derivatives 4, 9, 9, 11, 11,
170, 277, 31214, 457
credit-portfolio loss
distributions 8, 12, 45
credit-portfolio risk 79, 1516
correlations and 7
models 153294
credit rating systems, a latent
variable approach to
validating 27794
co-ratings, distribution of 285
empirical analysis 28493
consensus rating 2913
data description 2845
model selection and
parameter estimates
28691
residual analysis 2913,
292
results 28593
PD estimation and, 2804
rating bias 288, 289
rating errors 289, 290
credit risk, business cycles and
47, 5
credit-risk models, comparison
of, for portfolios based on
behavioural scores 20931
dynamics of, Markov chain
models 2239, 2279
proportional hazard models
21820, 2201, 222, 223
reputational structural model
21418
scoring explained 21214
D
delinquency rates 4, 45, 490
derivatives cashows in
liquidity risk models
299335
exercise probabilities 306
non-path-dependent
derivatives and, expected
cashows for 299308
American options 31012
barrier options 30810
credit derivatives 31214
path-dependent derivatives
and, expected cashows
for 30814
asymmetric 3049
notation and basics from
derivatives-pricing theory
299301
symmetric 3024
positive cashow at expiry,
probabilities for 310
E
economic and regularity capital
under stress 13749
general framework 13842
data description and 1425
economic capital and
1412
framework differences 142
regulatory capital and
13941
risk-factor model 1389
excess returns, distribution of
2933, 324, 34, 40
over time 3341
F
factor models, calibration and
validation of 239
calibration problem 23930
with equity data 24041
with historical default
information 241
nancial crisis:
another global view on 318
increasing volatilities
dominate during 335
issues of credit-risk
management raised by 137
number of hedge fund
failures increase during
432
operational versus credit risk
as cause of 360, 387
reputation of economics and
nance threatened by 93
unreliability of market-risk
models highlighted by 315
495
MODEL RISK
nancial instruments, methods
for pricing 4445
Finland 183206
see also macro shocks to loan
losses
G
Gaussian latent risk factors and
model risk 4852
basic single-risk-factor model
4850
homogeneous
single-risk-factor model
5052
general equilibrium effect,
monetary policy, asset
return dynamics and
93130
CCL (Chang et al) 2008 and,
10211
identication of regimes
105
1056
1045
CCL (Chang et al) 2009, focus
of 11125, 113, 114, 117
data 11315
econometric procedures
and 11517
two asset returns,
necessity of forecasting
1258, 126, 127, 128
in-sample forecasting
11718, 118, 129
11825, 119, 120, 121, 122,
123, 124, 125, 130
data, the economic model
and 1002
regime-switching SVAR
models and 97100
H
hedge fund performance 422,
430, 431
Australian example 42132
descriptive statistics of 430
legal structure and risk
disclosure of 4246, 425
data and empirical results
42932, 431
methodology and hypotheses
4269
operational risk and 42132
homogenisation and renement
of regulation 1617
I
impairment rates 13, 14, 14
imperfect information, investor
behaviour and 1924
information, pooling and
distribution of 17
International Swaps and
Derivatives Association
(ISDA) 1011, 142, 143
investor behaviour:
imperfect information and
1924
price dynamic and 249
L
liquidity risk models, modelling
derivatives cashows in
299335
exercise probabilities 306
non-path-dependent
derivatives and, expected
cashows for 299308
asymmetric 3048
notation and basics from
derivatives-pricing theory
299301
symmetric 3024
path-dependent derivatives
and, expected cashows
for 30814
American options 31012
barrier options 30810
496
INDEX
credit derivatives 31214
positive cashow at expiry,
probabilities for 310
Long-Term Capital Management
421, 423
M
macro shocks to loan losses,
transmission of 183206,
184
industry-specic default-rate
model estimation:
comparison 193
diagnostic tests 191
procedure 18792
results, 1926, 18890
loan-loss distribution and,
2007 versus pre-1990s
crisis 2045
loan-loss model and,
empirical t of 2034
simulation of loan losses and
196203
summary of 198
market risk 31555
ageing, interest-rate and
credit-spread levels,
impact of 325
bond and portfolio values
and 322
bond values at issuance,
interest rates and
credit-spread levels
inuence on 322
bootstrapping model,
potential future
credit-spread VaR and
actual credit-spread VaR
evolution under 333
bootstrapping model,
potential future
credit-spread VaR and
actual credit-spread VaR
evolution, with stressed
scenarios 335
delta-normal VaR approach
as measure of 31627
rst-kind model risk and
315, 31718
generic 31619
second-kind model risk
and 315, 31718, 319, 320,
3257
two-factor, example of
31925
factors, time series of 319
historical bootstrapping,
interest-rate and
credit-spread paths
obtained by 332
model risk and, backtesting
results used to deduce,
3545
modelling, assessing
adequacy of 33955
a posteriori backtesting
34452
a priori validation 3434
backtesting results as
guide to model risk 3545
building blocks 3401
embedded validation and
backtesting procedures
3524, 353
market risk model, results
from model risk in 3423
typical modelling errors
3412
portfolio sensitivity and 324
potential future, concept of
32735
application of PVaR 3315
denition and motivation
3279
VaR, calculation of 32931
sensitivity and covariance
parameters, impact of, on
portfolio VaR 326
simulated sensitivities,
quantiles of 332
simulated volatilities,
quantiles of 333
total risk in interest-rate risk
and credit-spread risk,
decomposition of 324
497
MODEL RISK
volatilities and correlations,
historical values of 320
Markov chain models 2239,
2279
model provider, concentration of
15
model risk:
aggregation of 4513
in counterparty credit risk
measurement systems 438,
439, 440, 442
aggregation of 4513
basic ingredients 43845
economic scenario
generation and 4434,
4469
nancial instruments and,
methods for pricing 4445
identication and
classication of 43753
spectra of model risks in
44553
credit derivatives and 4, 9, 11,
277, 31214
average maturity of 11
transactions 9
dening 467
downturn 318
correlations and 7
credit portfolio risks and
79, 1516
credit risk, business cycles
and 47, 5
lessons 1518
securitisations and 914
in Garch-type nancial time
series 7588, 83, 84, 85, 86
simulation procedure 812
simulation results 827
simulation study 817
stylised facts concerning
769, 812, 87, 88
volatility modelling and
7981
Gaussian latent risk factors
and 4852
basic single-risk-factor
model 4850
homogeneous
5052
knowledge transfer and,
facilitation of 18
measures of 478
average value-at-risk 46,
47
models, evaluation and
stress-testing of 17
non-Gaussian latent risk
factors and 4572
generalised homogeneous
559
generalised
535
mathematical appendix
6572
risk measures and 5964
extremal risk-factor
distributions 612
VaR, AVaR and model risk
sets 5961
worst-case VaR and AVaR
624
sources of, in economic
scenario generation 4469
sources of, in pricing models
44950
sources and mitigants of
15581
N
non-Gaussian latent risk factors
and model risk 4572
559
generalised single-risk-factor
model 535
mathematical appendix 6572
non-Gaussian latent risk factors
in model risk 4572
559
498
INDEX
generalised single-risk-factor
model 535
mathematical appendix 6572
normalised returns:
distribution of 312, 32
frequency of 35, 36, 37, 38, 39
O
operational value-at-risk,
estimation of 359417
data, application to 36983
backtesting 37983
expected loss, VaR and
CVaR 3789
operational frequency
distributions 3712
operational loss
distributions 3735
operational loss
distributions,
goodness-of-t tests for
3758
purpose and description
36971
external-type losses 372,
374, 375, 376, 377, 3801,
384, 386, 387, 3889, 390
frequency functions and 392
human-type losses 3936,
399, 403, 4089, 415
operational losses 372, 374,
376, 386, 392401
operational risk and,
compound model for
3629
capital charge,
implications of data
misspecication on 3659
Poisson process 3624
unbiased estimation of
severity and frequency
distributions 3645
process-type losses 3936,
400, 404, 41011, 416
relationship-type losses
3936, 398, 402, 4067, 414
robust approach to 3837
technology-type losses
3936, 401, 405, 41213,
417
P
point-in-time (PIT) credit-risk 5,
56, 7, 8, 13, 16, 257
probability of default (PD) 8,
163, 164, 165, 167, 176, 185,
2779, 281, 286, 2867, 293,
294, 312, 461, 4623
asset-return correlation
estimator, summary
statistics of 273
asymmetric asset correlation
and 26374
asymmetric correlation effect
and the estimation of
LRPD 26973
asymmetric correlation
models of credit risk 268
true parameters of 270
asymptotic single-risk factor
implications on the
estimation of 26374
latent variable model for
2804
long-run (LRPD) 2658,
26973, 272
model and LRPD, overview
of 2668
R
renement of regulation,
homogenisation and 1617
regime-switching SVAR models
97100
resecuritisations 1415
risk-factor models 45, 138, 142
basic single 4850
Gaussian single 459, 460
single 559, 65
generalised single 535, 65
homogeneous single 502
for portfolio credit risk 1389
risk measures and model risk
5964
499
MODEL RISK
extremal risk-factor
distributions 612
VaR, AVaR and model risk
sets 5961
worst-case VaR and AVaR
624
S
stress testing 17, 137
empirical results of 1458
scenario design 1456
scenarios 147
structural credit portfolio
models:
default data, estimators
applied to 2425
homogeneous cohorts 242
maximum-likelihood
estimator 2445
moment estimator 2424
default data, model
validation based on
24550
for default risk 2359
counting defaults 2389
default and asset
correlations 2368
model denition 2356
for default and migration risk
2502
factor models, calibration and
validation of 239
calibration problem
23930
with equity data 23041
with historical default
information 241
industry cohorts 247, 24850
intra-correlations:
over time 251
using rating data 255
variation of 256
model extensions 2568
modelling the credit cycle
2578
relaxing the normal
distribution assumption
2567
rating cohorts, correlations of
2459, 246
validating 23359
validation based on rating
data 2506
counting rating
migrations 2523
maximum-likelihood
estimator for 2535
model for default and
migration risk 2502
model validation 2556
structural vector
auto-regressive (SVAR)
106, 107, 108, 109,
regime-shifting 97100,
112
stylised facts 75, 769, 812,
87, 88, 96, 178, 340, 342, 343
T
through-the-cycle (TTC) model
49 passim, 13, 14, 257
V
validating credit rating systems,
a latent variable approach
to 27794
co-ratings, distribution of 285
empirical analysis 28493
consensus rating 2913
data description 2845
model selection and
parameter estimates
28691
residual analysis 2913,
292
results 28593
PD estimation and, 2804
rating bias 288, 289
rating errors 289, 290
volatility models 7981, 305
500

Model Risk

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Model Risk

Uploaded by

Copyright:

Available Formats

M

/K) and x() 312

(u) values for the external-type

(u) values for the external-type

(u) values for the relationship,

dened by Equation 3.14 coincides

. For portfolios with a large number of obligors, this fact

given in Equation 3.17 follows a common Vasicek

] in Equation 3.19 specialises to

] = /(2). The random

denote the distribution function of a standardised gen-

(x) +(1 c)(x), 0 < c < 1

(, , ) sothat the dis-

(, , ) coincides withthe distribution

(, , ) in an interval [a, b]. It further coincides

(, , ) have nite means

in the GHSRF model.

have? More particularly, what are the worst-case value-at-risk

be dened by Equation 3.17. Then the distribution of

depends on the default probability and on the correlation .

have? More particularly, what are the worst-case value-at-risk

(cf Equation 3.14) and upper quantile function

be dened as in Lemma 3.15.

dened in Equation 3.22. The subset G species

] giveninEquations 3.18 and3.19, respec-

is dened by Equation 3.22.

1 < < 1, then

is dened in Equation 3.17.

1 . Since the random vari-

are bounded by the unit interval, the risk measures

] have the upper

1 < < 1, then the worst-case value-at-risk and the

reach the maximum value 1, ie

This proves the rst inequality in Equation 3.20. The secondinequal-

and 0 < < 1. The discrete random variable

) = , which are derived in the proof of Lemma 3.18.

the worst-case loss distribution

(2) 5.2253 0.0571

(3) 5.2456 0.1121 5.1525

(4) 5.2352 0.3082 5.2012 0.0000

(5) 5.2451 0.0053

(6) 5.2458 0.0045

(7) 5.2443 0.0051

(8) 5.2450 0.0044

5.1620 0.3012 5.1138 0.3012

(11) 5.2211 0.2990 5.1954 0.0000

(12) 5.2155 0.0293

5.3919 0.8287 5.3958 0.0000

(2) 5.3650 0.0092

5.3911 0.2961 5.3950 0.0000

(3) 5.4411 0.0003

(4) 5.4290 0.0004

(5) 5.4434 0.0012

5.4060 0.1078 5.4063 0.0000

(6) 5.4404 0.0012

(7) 5.4124 0.0010

(8) 5.4115 0.0010

(9) 5.3641 0.0151

5.3897 0.1876 5.3938 0.0000

(11) 5.3641 0.0151

(12) 5.3621 0.0110

1.1664 0.3849 1.1715 0.0000

(2) 1.1399 0.0060

(3) 1.1208 0.0050