This art icle was downloaded by: [ b- on: Bibliot eca do conheciment o online UP]

On: 27 May 2012, At : 15: 56

Publisher: Taylor & Francis
I nforma Lt d Regist ered in England and Wales Regist ered Number: 1072954 Regist ered office: Mort imer
House, 37- 41 Mort imer St reet , London W1T 3JH, UK
Journal of Busi ness & Economi c St at i st i cs
Publ i cat i on det ai l s, i ncl udi ng i nst r uct i ons f or aut hor s and subscr i pt i on i nf or mat i on:
ht t p: / / amst at . t andf onl i ne. com/ l oi / ubes20
The Fact orSpl i neGARCH Model f or Hi gh and Low
Frequency Correl at i ons
Jos Gonzal o Rangel
a
& Rober t F. Engl e
b
a
Gl obal Invest ment Resear ch, Gol dman Sachs Gr oup, Inc. , 200 West St r eet , New Yor k,
NY, 10282- 2198
b
Depar t ment of Fi nance, St er n School of Busi ness, New Yor k Uni ver si t y, 44 West Four t h
St r eet , Sui t e 9- 62, New Yor k, NY, 10012- 1126
Avai l abl e onl i ne: 22 Feb 2012
To ci t e t hi s art i cl e: Jos Gonzal o Rangel & Rober t F. Engl e ( 2012) : The Fact or Spl i neGARCH Model f or Hi gh and Low
Fr equency Cor r el at i ons, Jour nal of Busi ness & Economi c St at i st i cs, 30: 1, 109- 124
To l i nk t o t hi s art i cl e: ht t p: / / dx. doi . or g/ 10. 1080/ 07350015. 2012. 643132
PLEASE SCROLL DOWN FOR ARTI CLE
Full t erms and condit ions of use: ht t p: / / amst at . t andfonline. com/ page/ t erms- and- condit ions
This art icle may be used for research, t eaching, and privat e st udy purposes. Any subst ant ial or syst emat ic
reproduct ion, redist ribut ion, reselling, loan, sub- licensing, syst emat ic supply, or dist ribut ion in any form t o
anyone is expressly forbidden.
The publisher does not give any warrant y express or implied or make any represent at ion t hat t he cont ent s
will be complet e or accurat e or up t o dat e. The accuracy of any inst ruct ions, formulae, and drug doses
should be independent ly verified wit h primary sources. The publisher shall not be liable for any loss, act ions,
claims, proceedings, demand, or cost s or damages what soever or howsoever caused arising direct ly or
indirect ly in connect ion wit h or arising out of t he use of t his mat erial.
Supplementary materials for this article are available online. Please go to http://tandfonline.com/r/JBES
The FactorSplineGARCH Model for High
and Low Frequency Correlations
Jos e Gonzalo RANGEL
Global Investment Research, Goldman Sachs Group, Inc., 200 West Street, New York, NY 10282-2198
(josegonzalo.rangel@gs.com)
Robert F. ENGLE
Department of Finance, Stern School of Business, New York University, 44 West Fourth Street, Suite 9-62,
New York, NY 10012-1126 (rengle@stern.nyu.edu)
We propose a new approach to model high and low frequency components of equity correlations. Our
framework combines a factor asset pricing structure with other specications capturing dynamic properties
of volatilities and covariances between a single common factor and idiosyncratic returns. High frequency
correlations mean revert to slowly varying functions that characterize long-term correlation patterns. We
associate such term behavior with low frequency economic variables, including determinants of market
and idiosyncratic volatilities. Flexibility in the time-varying level of mean reversion improves both the
empirical t of equity correlations in the United States and correlation forecasts at long horizons.
KEY WORDS: DCC; Dynamic correlation; Factor models; Idiosyncratic volatility; Long-term forecast
evaluation; Time-varying betas.
1. INTRODUCTION
This article introduces a new approach to characterize and
forecast high and low frequency variation in equity correla-
tions. By separating short-term from long-term components,
our method not only facilitates the economic interpretation of
changes in the correlation structure but achieves improvements
over leading methods in terms of tting and forecasting equity
correlations.
A number of multivariate time series models have been pro-
posed over the last two decades to capture the dynamic proper-
ties in the comovements of asset returns. Multivariate versions
of the well-known univariate Generalized Autoregressive Con-
ditional Heteroscedasticity (GARCH) and Stochastic Volatility
(SV) models guided the initial specications (e.g., Bollerslev,
Engle, and Woodridge 1988 and Harvey, Ruiz, and Shephard
1994), but they had limitations because they were heavily param-
eterized and/or difcult to estimate. Simplied versions, such as
constant conditional correlation models (e.g., Bollerslev 1990;
Alexander 1998; Harvey et al. 1994), were also unattractive be-
cause they had problems in describing the empirical features of
the data. Bauwens, Laurent, and Rombouts (2006) and Shep-
hard (2004) presented extended surveys of this literature. Engle
(2002) introduced the Dynamic Conditional Correlation (DCC)
model as an alternative approach to achieve parsimony in the dy-
namics of conditional correlations, thus maintaining simplicity
in the estimation process. However, none of the aforementioned
models associate correlation dynamics with features of funda-
mental economic variables. Moreover, since they return to a
constant mean in the long run, their forecasting implications
for long horizons do not take into account changing economic
conditions. Thus, they produce the same long-term forecast at
any point in time.
Financial correlation models, on the other hand, have only
recently introduced dynamic patterns in correlations (e.g., Ang
and Bekaert 2002; Ang and Chen 2002; Bekaert, Hodrick, and
Zhang 2009). Although these models are linked to asset pricing
frameworks, which facilitate the association of correlation be-
havior with nancial and economic variables, a substantial part
of the variation in the correlation structure remains unexplained
and their implementation for forecasting correlations appears
difcult.
This article presents a new model that captures complex fea-
tures of the comovements of nancial returns and allows em-
pirical association of economic fundamentals with the dynamic
behavior of variances and covariances. Following recent devel-
opments in time series methods, this model incorporates, within
a single framework, the attractive features of the two approaches
mentioned above. Based on the correlation structure suggested
by a simple one-factor Capital Asset Pricing Model (CAPM)
specication, and short- and long-termdynamic features of mar-
ket and idiosyncratic volatilities, we derive a correlation model
that allows conditional (high frequency) correlations to mean
revert toward smooth time-varying functions, which proxy the
low frequency component of correlations. This property not
only represents a generalization of multivariate GARCH mod-
els that show mean reversion to a constant covariance matrix,
but also provides exibility to the long-term level of correla-
tions in adapting to the changing economic environment. (Factor
models in which second moments mean revert to constant levels
were presented by Engle, Ng, and Rothschild 1990; Diebold and
Nerlove 1989; and Chib, Nardari, and Shephard 2006, among
others.)
To achieve this goal, the semiparametric SplineGARCH ap-
proach of Engle and Rangel (2008) is used to model high and
2012 American Statistical Association
Journal of Business & Economic Statistics
January 2012, Vol. 30, No. 1
DOI: 10.1080/07350015.2012.643132
109
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

110 Journal of Business & Economic Statistics, January 2012
low frequency dynamic components of both systematic and
idiosyncratic volatilities. As a result, a slow-moving low fre-
quency correlation part is separated from the high frequency
part. Effects of time-varying betas and unspecied unobserv-
able factors are incorporated into the high frequency correlation
component by adding dynamic patterns to the correlations be-
tween the market factor and each idiosyncratic component and
between each pair of idiosyncratic risks. (A range of models
with time-varying betas have been studied by Bos and New-
bold 1984; Ferson and Harvey 1991, 1993; and Ghysels 1998;
among many others.) The high frequency patterns are modeled
using a DCC process. Therefore, the resulting FactorSpline
GARCH (FSG-DCC) model blends SplineGARCH volatility
dynamics with DCC correlation dynamics within a factor asset
pricing framework.
From an empirical perspective, this study analyzes high and
low frequency correlation patterns in the U.S. market by con-
sidering daily returns of stocks in the Dow Jones Industrial
Average (DJIA) during a period of 17 years. We nd that, in
addition to the recently documented economic variation in mar-
ket volatility at low frequencies (e.g., Engle and Rangel 2008;
Engle, Ghysels, and Sohn 2008), average idiosyncratic volatility
shows substantial variation in its long-term component, which
is highly correlated with low frequency economic variables in-
cluding an intersectoral Employment Dispersion Index given by
Lilien (1982). Since this variable measures the intensity of shifts
in product demand across sectors, we use it to proxy changes in
the intensity of idiosyncratic news. For instance, technological
change or any other driver of demand shifts can induce large
movements of productive factors fromdeclining to growing sec-
tors, and lead to increases in the intensity of rm-specic news.
Consistent with this intuition, we nd that this variable is pos-
itively related to idiosyncratic volatility. Since the intensity of
sectoral reallocation is associated with the same sources that
lead to variation in productivity (or protability) across rms
and sectors, our empirical ndings are also consistent with the
positive relationship between idiosyncratic volatility and the
volatility of rm protability suggested by Pastor and Veronesi
(2003).
We also investigate the forecast performance of the FSG-
DCC model focusing on long horizons (four to six months).
Based on an economic loss function and following the approach
of Engle and Colacito (2006), we perform a sequential out-
of-sample forecasting exercise that compares this model with
a broad set of competing models. We nd signicant evidence
that the FSG-DCCoutperforms its competitors at long horizons.
A complementary in-sample exercise presented in the online
Appendix A5 also supports this nding.
Overall, by combining SplineGARCH dynamics with DCC
dynamics within a factor framework that separates systematic
and idiosyncratic terms, our model provides exibility to the
level of mean reversion of asset correlations and captures eco-
nomic variation of its components. These features explain the
improvements in terms of empirical t and log-term correla-
tion forecasts shown in this paper and lead us to new empirical
results on the economic drivers of idiosyncratic volatility and
correlations.
The article is organized as follows: Section 2 provides a de-
scription of a number of correlation specications associated
with different assumptions in the factor setup. Section 3 in-
troduces the FSGDCC model. Section 4 presents an empirical
analysis of correlations in the U.S. market, empirical evidence of
economic variation in aggregate idiosyncratic volatility, and an
empirical evaluation of correlation specications derived from
different factor models. Section 5 examines the forecast perfor-
mance of the FSGDCC model, and Section 6 concludes.
2. A SINGLE-FACTOR MODEL AND RETURN
CORRELATIONS
In this section, we use a simple one-factor version of the arbi-
trage pricing theory (APT) asset pricing model of Ross (1976)
and describe howmodifying its underlying assumptions changes
the specication of the correlation structure of equity returns.
Let us assume that there is a single market factor that enters
linearly in the pricing equation as in the Sharpe (1964) CAPM
model. Under this specication, and measuring returns in ex-
cess of the risk-free rate, the excess return of asset i is generated
by
r
it
=
i
+
i
r
mt
+u
it
, (1)
where r
mt
denotes the market excess return. The rst term
characterizes asset is systematic risk and the second term de-
scribes its idiosyncratic component. Absence of arbitrage as-
sures
i
= 0 and E(r
it
) =
i
, where denotes the risk pre-
mium per unit of systematic risk. The standard APT structure
assumes constant betas and the following conditions:
E(r
mt
u
it
) = 0, i, (2)
E(u
it
u
jt
) = 0, i = j. (3)
The assumptions in the factor structure thus impose a restric-
tion in the covariance matrix of returns. Assuming that moment
restrictions in Equations (2) and (3) hold conditionally, we ob-
tain a specication for conditional correlations, which depends
exclusively on dynamic patterns in the conditional variances of
market and idiosyncratic risks. Furthermore, if the restriction in
Equation (2) holds conditionally, then the betas are constant and
correctly estimated fromsimple time series regressions of excess
returns on the market factor. Restriction (3) rules out correlation
between idiosyncratic innovations, which precludes the possi-
bility of missing pricing factors in the model. As suggested by
Engle (2007), these restrictions are empirically unappealing and
limit importantly the dynamic structure of correlations. Allow-
ing temporal deviations from Equations (2) and (3) increases
substantially the ability of the resulting correlation models to
capture empirical features of the data maintaining the economic
essence of the factor pricing structure. The following proposi-
tion characterizes these more exible correlation models.
Proposition 1. Consider the model specication in Equations
(1), (2), and (3), and let
t
denote the set of current and past
information available in the market.
(a) If E
t 1
(u
it
u
jt
) = 0, for some i = j, and E
t 1
(r
m,t
u
it
) =
0, i, then the conditional correlations will be given by

i,j,t
=

i

j
V
t 1
(r
mt
)+E
t 1
(u
it
u
jt
)
_

2
i
V
t 1
(r
mt
)+V
t 1
(u
it
)
_

2
j
V
t 1
(r
mt
)+V
t 1
(u
jt
)
.
(4)
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

Rangel and Engle: FactorSplineGARCH Model 111
(b) If E
t 1
(u
it
u
jt
) = 0, for some i = j, and E
t 1
(r
m,t
u
it
) =
0, for some i, then the conditional correlations will be
given by

i,j,t
=
_

j
V
t 1
(r
mt
) +
j
E
t 1
(r
mt
u
it
) +
i
E
t 1
(r
mt
u
jt
)
+E
t 1
(u
it
u
jt
)
_

2
i
V
t 1
(r
mt
) +V
t 1
(u
it
) +2
i
E
t 1
(r
mt
u
it
)
_
1/2

2
j
V
t 1
(r
mt
) +V
t 1
(u
jt
) +2
j
E
t 1
(r
mt
u
jt
)
_
1/2
(5)
where
i
=
cov(r
it
,r
mt
)
V(r
mt
)
.
(c) Under the assumptions in b), an equivalent representation
incorporates time-varying conditional betas
(i)
r
it
=
it
r
mt
+ u
it
,
it

cov
t 1
(r
it
, r
mt
)
V
t 1
(r
mt
)
, i.
(6)
(ii) The conditional correlation can then be equiva-
lently expressed as

i,j,t
=

i,t

j,t
V
t 1
(r
mt
) +cov
t 1
( u
it
, u
jt
)
_
_

2
i,t
V
t 1
(r
mt
) +V
t 1
( u
it
)
_ _

2
j,t
V
t 1
(r
mt
) +V
t 1
( u
jt
)
_
.
(7)
The proof is given in Appendix A1. The constant terms are
omitted without loss of generality. Equation (4) describes a case
where the factor loadings are constant, but latent unobserved
factors (omitted fromthe specication and embedded in the error
term) can have temporal effects on conditional correlations.
Parts b) and c) of Proposition 1 consider the case of time-varying
betas. The representation in Equation (6) satises assumption
(2) (see Appendix A1) and can be rewritten as
r
it
=
i
r
mt
+u
it
,
i
=
cov (r
it
, r
mt
)
V (r
mt
)
,
u
it
= u
it
+(
it

i
) r
mt
, i, (8)
where this u
it
termalso satises Equation (2) and the equivalence
between the two representations follows.
Hence, Equation (5) provides a more general specication
for conditional correlations which, under some assumptions,
simultaneously incorporates the effects of time variation in the
betas and latent unobserved omitted factors. For this reason, it
is the basis of our econometric approach.
3. AN ECONOMETRIC MODEL FOR A FACTOR
CORRELATION STRUCTURE
3.1 The FactorSplineGARCH Model
Equity volatilities show different patterns at high and low
frequencies. Short-term volatilities are mainly determined by
fundamental news arrivals, which induce price changes at very
high frequencies. Longer term volatilities show patterns gov-
erned by slow-moving structural economic variables. Engle and
Rangel (2008) found economically and statistically signicant
variation in the low frequency market volatility in the United
States as well as in most developed and emerging countries
(Schwert 1989 and Engle et al. 2008 also examine the economic
variation of market volatility at low frequencies). We introduce
this effect in our factor correlation model by including an equa-
tion that describes the dynamic behavior of this low frequency
market volatility.
Incorporating the low frequency variation of idiosyncratic
volatilities into the correlation structure is also appealing be-
cause there is empirical evidence of long-term dynamics in
such volatilities. For instance, Campbell, Lettau, Malkiel, and
Xu (2001) found evidence of a positive trend in idiosyncratic
rm-level volatility during the 19621997 period and no trend in
market volatility, suggesting a declining long-termpattern in the
correlations of individual stocks over this period. Theoretical ex-
planations of the upward trend in idiosyncratic volatilities have
been associated with different rm features such as the variance
of total rm protability, uncertainty about average protabil-
ity, age, institutional ownership, and the level and variance of
growth options available to managers (see Pastor and Veronesi
2003; Wei and Zhang 2006; Cao, Simin, and Zhao 2008). At
the aggregate level, Guo and Savickas (2006) found high cor-
relation between idiosyncratic volatility and the Consumption
Wealth Ratio (CAY) proposed by Lettau and Ludvigson (2001).
Their results suggest that idiosyncratic volatility could measure
an omitted risk factor or dispersion of opinion. Overall, these
empirical and theoretical results motivate our approach of incor-
porating long-termpatterns of both systematic and idiosyncratic
volatilities into a model for correlation dynamics.
From an econometric standpoint, the SplineGARCH model
of Engle and Rangel (2008) provides a semiparametric frame-
work to separate high and low frequency components of volatil-
ities. Following this approach, we model the market factor in
Equation (1) as
r
mt
=
m
+

mt
g
mt

m
t
, where
m
t
|
t 1
(0, 1),
g
mt
=
_
1
m

m


m
2
_
+
m
(r
mt 1

m
)
2

mt 1
+
m
(r
mt 1

m
)
2

mt 1
I
r
mt 1
<0
+
m
g
mt 1
,

mt
= c
m
exp
_
w
m0
t +
k
m

i=1
w
mi
((t t
i1
)
+
)
2
_
, (9)
where g
mt
and
mt
characterize the high and low frequency
market volatility components, respectively, (t x)
+
= (t x)
if t > x, and, otherwise, is zero, and I
r
mt
<0
is an indicator
function of negative market returns, but different from Engle
and Rangel (2008), the high frequency component is mod-
eled as an asymmetric unit GARCH process following Glosten,
Jagannathan, and Runkle (1993). We thus capture the well-
documented leverage effect (see Black 1976; Christie 1982;
Campbell and Hentschel 1992), where bad news increase the fu-
ture high frequency volatility more than good news. The term
mt
approximates the unobserved low frequency market volatility
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

112 Journal of Business & Economic Statistics, January 2012
that responds to low frequency fundamental variables, such as
macroeconomic aggregates, which characterize the variation in
the economic environment. As in Engle and Rangel (2008),
this component is modeled using an exponential quadratic
spline with equally spaced knots that are selected according to
the Schwarz Criterion (BIC or Bayesian information criterion).
The term g
mt
describes transitory volatility behavior that, de-
spite its persistence, does not have long-term impacts on market
volatility.
Similarly, we model the idiosyncratic part of returns in (1) as
u
it
=

it
g
it

it
, where
it
|
t 1
(0, 1),
g
it
=
_
1
i

i


i
2
_
+
i
(r
it 1

i

i
r
mt 1
)
2

it 1
+
i
(r
it 1

i

i
r
mt 1
)
2

it 1
I
r
it 1
<0
+
i
g
it 1
,

it
= c
i
exp
_
w
i0
t +
k
i

r=1
w
ir
((t t
r1
)
+
)
2
_
, i. (10)
The g
it
s characterize the high frequency component of
idiosyncratic volatilities associated with transitory effects,
whereas the
it
s describe low frequency variation in idiosyn-
cratic volatilities. The I
r<0
s are indicator functions of neg-
ative returns that allow for rm-specic leverage effects. As
before, we take a multiplicative error model to describe interac-
tions between rm-specic news arrivals and low frequency
state variables measuring rm- and industry-specic condi-
tions. The intuition here is that a rm-specic news event will
have a larger effect, for example, when the rm is close to
bankruptcy or when a major technological change is affecting
the rms industry. In sum, the
it
s approximate nonparamet-
rically the unobserved long-term idiosyncratic volatilities that
are functions of low frequency economic variables, which af-
fect the magnitude and intensity of high frequency idiosyncratic
shocks.
Following the discussion in the previous section and
in Proposition 1, we incorporate time-varying correlations
between the market factor and idiosyncratic returns, and
among idiosyncratic terms themselves. Specically, we as-
sume that the vector of innovations in Equations (9) and (10),
(
m
t
,
1t
,
2t
, . . . ,
Nt
)

, follows the stationary DCC model of

Engle (2002). All elements in this vector have unit conditional
variance. Thus, from the second DCC estimation stage, we
have

i,j,t
=
q
i,j,t

q
i,i,t

q
j,j,t
,
q
i,j,t
=

i,j
+a
DCC
_

i,t1

j,t1

i,j
_
+b
DCC
_
q
i,j,t1

i,j
_
, i, j {1, . . . , N} ,
q
m,i,t
=

m,i
+a
DCC
_

m
t 1

i,t1

m,i
_
+b
DCC
_
q
m,i,t1

m,i
_
, i {m, 1, . . . , N} ,
(11)
where

i,j
= E(
it

jt
) and

i,i
= 1, for all i = 1, 2, . . . , N. We
add mean reverting features to the betas described in Propo-
sition 1 and assume that

m,i
= 0, for all i = 1, 2, . . . , N.
The other intercepts are estimated by correlation targeting
(see Engle 2002; Engle and Sheppard 2005a). The above
specications, together with the factor structure presented in
Section II, constitute the full FactorSplineGARCH (FSG-
DCC) model. The next proposition describes its correlation
structure.
Proposition 2. Given a vector of returns (r
1t
,
r
2t
, . . . , r
Nt
)

satisfying the factor structure in Equation

(1), let us assume that the common market factor r
mt
is
described by Equation (9), the idiosyncratic term u
it
follows the
process in Equation (10), for all i = 1, 2, . . . N, and the vector
of innovations (
m
t
,
1t
,
2t
, . . . ,
Nt
)

follows the DCC process

in Equation (11) and its assumptions. The high frequency
(conditional) correlation between r
it
and r
jt
is then given by

i,j,t
=
_

mt
g
mt
+
i

mt
g
mt

jt
g
jt

m,j,t
+
j

mt
g
mt

it
g
it

m,i,t
+

it
g
it

jt
g
jt

i,j,t
_

2
i

mt
g
mt
+
it
g
it
+2
i

mt
g
mt

it
g
it

m,i,t
_
1/2

2
j

mt
g
mt
+
jt
g
jt
+2
j

mt
g
mt

jt
g
jt

m,j,t
_
1/2
(12)
and the low frequency component of this correlation is time-
varying and takes the following form:

i,j,t
=

mt
+

it

jt

i,j
_

2
i

mt
+
it
_

2
j

mt
+
jt
. (13)
Assuming that E
t
(
k,t +h
) =
k,t
, h > 0, k = 1, 2, . . . , N,
(13) is then the long-term forecast of
i,j,t
lim
h

i,j,t +h|t
=
i,j,t
. (14)
The proof is given in Appendix A2. Note that Equation (12)
is a parameterized version of Equation (5) in Proposition 1.
Equation (13) approximates the slow-moving component of
correlations, which can be associated with long-term correla-
tion dynamics. Indeed, the high frequency correlation parsimo-
niously mean reverts toward this time-varying low frequency
term. This approximation may be improved by adding more
observable factors or allowing for time variation in the

i,j
s,
or both. The rst alternative can be easily implemented once
we have selected the new factors (see Engle and Rangel 2009).
The second extension would capture long-term effects of ex-
cluded factors, but it is methodologically challenging because it
requires exploring alternative functional forms (or restrictions)
to guarantee positive-deniteness of the covariance matrix. In
the empirical part of this article, we focus on the simplest case
given in Proposition 2.
Equation (13) also provides a simple approach to forecast
long-term correlations using economic variables. Specically,
we can use forecasts of low frequency market and idiosyn-
cratic volatilities, which can be obtained from models that
incorporate economic variables. For instance, the results of
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

Rangel and Engle: FactorSplineGARCH Model 113
Engle and Rangel (2008) allow construction of forecasts of
market volatility using macroeconomic and market information.
From a pure time series perspective, Equation (14) presents a
useful forecasting relation in which the time-varying low fre-
quency correlation can be interpreted as the long run correlation
forecast, assuming that the low frequency market and idiosyn-
cratic volatilities remain constant during the forecasting period.
Another interesting case embedded in Proposition 2 oc-
curs when low frequency variation is not allowed. This re-
stricted version corresponds to the Factor DCC (FG-DCC)
model of Engle (2007) and is derived by assuming
m,t
=
2
m
and
i,t
=
2
i
, i, in Equations (9) and (10). The correspond-
ing conditional variance specications for the factor and the
idiosyncrasies (h
mt
and h
it
,i) become standard mean reverting
asymmetric GARCH(1,1) processes. Thus, Equation (12) takes
the following form:

i,j,t
=
_

j
h
mt
+
j
_
h
mt
_
h
it

m,i,t
+
j
_
h
mt
_
h
it

m,j,t
+
_
h
it
_
h
jt

i,j,t
_

2
i
h
mt
+h
it
+2
i
_
h
mt
h
it

m,i,t
_
1/2

2
j
h
mt
+h
jt
+2
j
_
h
mt
h
jt

m,j,t
_
1/2
(15)
and the low frequency correlation, which represents the
long-run correlation forecast associated with the FG-DCC
model, is a constant given by the unconditional version of
Equation (15).
3.2 Estimation
To facilitate presenting our estimation approach, we rewrite
the FSG-DCCmodel using matrix notation. The systemof equa-
tions in the factor setup can be written as
r
t
= +Bu
t
, (16)
where r
t
= (r
mt
, r
1t
, . . . , r
Nt
)

is the vector of excess returns

at time t, u
t
contains both the market factor and the idiosyn-
crasies, u
t
= (r
mt
, u
1t
, u
2t
, . . . , u
Nt
)

, B =
_
1 0
1N
I
NN
_
, =
(
1
,
2
, . . . ,
N
)

, 0
1 N
is an N-dimensional row vector of ze-
ros, I
N N
is the N-dimensional identity matrix, and is a vector
of intercepts. The market factor is assumed to be weakly exoge-
nous (see Engle, Hendry, and Richard 1983). In this setup, the
covariance matrix of u
t
can be written as
H
u
t
= D
t
R
t
D
t
, (17)
where D
t
= diag{

kt
h
kt
}, for k = m,1, . . . , N, R
t
=
diag(Q
t
)
1/2
Q
t
diag(Q
t
)
1/2
is a correlation matrix, and the typ-
ical element of Q
t
is dened by Equation (11). The standard-
ized innovations in Proposition 2 are the elements of D
1
t
u
t
=
(
m
t
,
1t
,
2t
, . . . ,
Nt
)

. Going back to the original vector of re-

turns, we have that r
t
|
t 1
(, H
t
), where the full covariance
matrix is
var (r
t
) = H
t
= BD
t
R
t
D
t
B

, (18)
and the correlation matrix is R
t
= diag{BD
t
R
t
D
t
B

}
1/2
BD
t
R
t
D
t
B

diag{BD
t
R
t
D
t
B

}
1/2
. As in the standard DCC
model, the estimation problem can be formulated follow-
ing Newey and McFadden (1994) as a two-stage Gen-
eralized Method of Moments (GMM) problem. We can
write the vector of moment conditions as g(r
t
, , ) =
(m
1
(r
t
, ), m
2
(
t
, , ))

, where
t
D
1
t
u
t
is a vector of de-
volatized returns, is a vector of parameters containing the
alphas, the unconditional betas, and the volatility parameters
in Equations (9) and (10), and is a vector containing the
DCC parameters in Equation (11). The rst vector of moment
conditions, m
1
, has the scores of individual asymmetric Spline
GARCH models as its components. The second set of moment
conditions, m
2
, contains the functions that involve the corre-
lation parameters. If the rst optimization problem involving
the moment conditions in m
1
provides consistent estimates of
the volatility and mean parameters, the optimization of m
2
in the
second step will also provide consistent estimates.
In a Gaussian quasi-likelihood (QML) framework, assuming
multivariate normality leads to consistent estimates under mild
regularity conditions as long as the mean and the covariance
equations are correctly specied (see Bollerslev and Wooldridge
1992; Newey and Steigerwald 1997). Despite the convenience
of the Gaussian QML approach, choosing the number of knots
for the spline functions introduces an additional procedure in
the estimation process that does not necessarily provide QML
estimates of such quantities. Inaccurate choices of the number
of knots might introduce some biases due to misspecication of
volatilities. Since distributional assumptions might have an ef-
fect on this part of the estimation, we depart from normality and
use distributional assumptions that more realistically describe
the empirical features of returns. This can improve the accuracy
of the knot selection criterion and reduce the misspecication
problem. Therefore, for the rst stage of the estimation process,
we consider likelihoods from the Student-t distribution because
this distribution is better able to capture fat tails and the effect
of inuential outliers. The following log-likelihoods correspond
to the Asymmetric SplineGARCH model with Student-t inno-
vations (an extension of the Bollerslev 1987 model) and deter-
mine the moment conditions associated with the rst stage of
our GMM estimation:
L(
m
, v
m
) = log
_
((v
m
+1)/2)
(v
m
/2) ((v
m
2)g
mt

mt
)
1/2
_

v
m
+1
2
log
_
1 +
(r
mt

m
)
2

mt
g
mt
(v
m
2)
_
,
L(
i
, v
i
) = log
_
((v
i
+1)/2)
(v
i
/2) ((v
i
2)g
it

it
)
1/2
_

v
i
+1
2
log
_
1 +
(r
it

i

i
r
mt
)
2

it
g
it
(v
i
2)
_
,
i = 1, . . . , N, (19)
where denotes the gamma function and the vs refer to the
corresponding degrees of freedom. The resulting rst stage
estimates satisfy consistency under the identication conditions
of Newey and Steigerwald (1997).
Regarding the second stage, the estimation can be performed
as a joint estimation of the whole correlation matrix using QML
moment conditions. However, although DCC is very parsimo-
nious in its parameterization, it is biased and slow for large
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

114 Journal of Business & Economic Statistics, January 2012
covariance matrices. Alpha is biased downward and may ap-
proach zero (see Engle and Sheppard 2005b). Since our empir-
ical analysis includes a moderately large number of assets, we
consider two approaches that have recently been suggested in
the literature to reduce the bias problem: the MacGyver method
of Engle (2007), and the Composite Likelihood (CL) method
of Engle, Shephard, and Sheppard (2008). The rst approach is
appealing because it is easy to implement, although it has the
drawback of making it difcult to conduct inference. The second
method addresses this issue by dening a CLfunction as the sum
of quasi-likelihoods of pairs of assets. In the following section,
we present results using the standard multivariate method and
these two novel strategies. In Section 5, we performa sequential
out-of-sample forecasting exercise using the CL method.
4. HIGH AND LOW FREQUENCY CORRELATIONS IN
THE U.S. MARKET
4.1 Data
We use daily returns on DJIA stocks from December 1988
to December 2006. The data are obtained from the Center for
Research in Security Prices (CRSP). The index underwent some
changes during this period, including additions, deletions, and
mergers. We include all stocks in the 2006 index and those in
the 1988 index that could be followed over the sample period
(Chevron (CVX), Goodyear (GT), and International Paper (IP)),
obtaining a sample of 33 stocks. Regarding the market factor,
we use daily returns on the S&P 500 and the one-month T-bill
rate as the time-varying risk-free rate. Stock names and tickers
are described in the online Appendix A3.
4.2 Description of Estimation Results
We estimate the FSG-DCC model following the two-stage
GMM approach described in Section 3 and present the results
in the online Appendix A3 (Table A3-2). The parameters esti-
mated in the rst step are those in Equations (9), (10), and (19).
For reasons of space, we only present the optimal number of
knots selected for the spline terms. The second step involves
the DCC parameters in Equation (11). The rst column of this
table shows the alphas which, in general, are not signicantly
different from zero, as suggested by the CAPM framework (the
only exceptions are Altria Group (MO), General Motors (GM),
and Goodyear (GT)). The second column reports the estimated
unconditional market betas associated with each stock. All are
highly signicant. The third column presents the estimates of the
ARCHvolatility coefcients. In general, they are signicant, ex-
cept for the market factor and IBM, and their median is 0.05. The
fourth column presents the estimates of the GARCH volatility
effects. They are all signicant and their median is 0.84. The fth
column shows estimates of the leverage effects, which exhibit a
median of 0.03. They are statistically signicant in about half of
the cases and also positive. These results are consistent with the
leverage theory of Black (1976) and Christie (1982). However,
this effect is substantially higher and signicant for the market
factor, which provides stronger support for the volatility feed-
back hypothesis. The sixth column shows the degrees of freedom
of the univariate Student-t distributions, which uctuate between
4 and 10, and their median is 6. These values are in line with the
traditional evidence of nonnormality and excess of kurtosis in
equity returns. The last column of Table A3-2 shows the optimal
number of knots in the spline functions according to the BIC.
These numbers reect changes in the curvature of the long-term
trend of idiosyncratic volatilities. Examples of such patterns for
individual stocks (Intel (INTC), Exxon Mobil (XOM), and In-
ternational Paper (IP)) are shown in the rst column of graphs in
Figure 1. The bottom part of Table A3-2 presents the estima-
tion results of the second stage. The rst column reports the
standard DCC estimates from the traditional approach and their
standard errors. The second and third columns present the Mac-
Gyver DCC and the CL estimates, respectively. The estimator
of a
DCC
increases from 0.0027, using the standard multivariate
method, up to 0.004 (or 0.005) when using the CL (or the Mac-
Gyver) method. This suggests that both methods deliver a bias
correction of the traditional DCC estimator.
The third column of graphs in Figure 1 illustrates the time
series properties of the FSG-DCC, with a number of exam-
ples for individual stocks that show the high and low frequency
correlation components and the model-free (100-day) rolling
correlations. The high frequency component mean reverts to-
ward the slow-moving low frequency component. It is visually
clear by looking at the model-free rolling correlations that the
FSG-DCC model characterizes fairly well the trend behavior in
correlations. The second column of graphs in Figure 1 illustrates
with some examples the models ability to capture time variation
in the betas. The graphs show the conditional betas implied by
the FSG-DCC model, which are computed using Equation (6)
and the model-based volatilities and correlations in Equations
(9), (10), and (11) that allow us to obtain the relevant condi-
tional covariance terms. To further illustrate the exibility of
the conditional betas, the graphs also present model-free (100-
day) rolling betas computed from rolling regressions using the
specication in Equation (1).
4.3 Aggregate Volatility Components and Average
Correlations
The most distinguishing feature of the FSG-DCC model is its
ability to characterize dynamic long-term correlation behavior
by exploiting the structure of a factor asset pricing model and
the low frequency variation in systematic and idiosyncratic
volatilities. Engle and Rangel (2008) and Engle et al. (2008)
provide evidence that the low frequency market volatility
responds to changes in the slow-moving macroeconomic
environment. This subsection illustrates the effects of the low
frequency volatility components on aggregate correlations and
presents evidence that the aggregate low frequency component
of idiosyncratic volatility systematically varies with economic
variables. This evidence also holds at the sectoral level (we
present an analysis based on the 48 equally weighted industry
portfolios of Fama and French (1997) in Appendix A4).
Cross-sectional aggregation facilitates illustrating the effect
of our volatility components on correlations. We construct
aggregates by taking the cross-sectional average of our dynamic
components at each point in time. Hence, the aggregate average
correlation associated with model m for a specic period p is
dened as:
m
T
p
=
1
N

N
i=1
{
1
T
p
1
(N1)

j=i

T
p
t =1

m
i,j,t
}, where T
p
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

Rangel and Engle: FactorSplineGARCH Model 115
Figure 1. Idiosyncratic volatilities, betas, and correlations. NOTES: The estimation uses daily returns on the DJIA stocks from December
1988 to December 2006. The data are obtained from CRSP. Company names are referred to by their tickers (INTC = Intel, XOM = Exxon
Mobil, and IP = International Paper). Volatilities column: HFV stands for High frequency idiosyncratic volatility (see the second equation
in Equation (10)) and LFV refers to Low frequency idiosyncratic volatility (see the third equation in Equation (10)). The last number in the
series labels denotes the optimal number of knots. Betas column: The solid line plots the FSG-DCC conditional beta and the dashed line plots
the rolling betas based on 100-day rolling regressions. Correlations column: LFC refers to Low frequency correlation, HFC stands for High
frequency correlation, and CR refers to 100-day Rolling correlations.
denotes the number of daily observations in such a period, and

m
i,j,t
is a time-varying pairwise correlation (from model m). The
average low frequency idiosyncratic volatility for period p is
dened as Ivol
T
p
=
1
T
p
1
N

T
p
t =1

N
i=1

1/2
it
. Figure 2 shows av-
erages for biannual subperiods (from the entire sample) of low
frequency market and idiosyncratic volatilities. Before 1997
1998, while the average low frequency idiosyncratic volatility
showed an increasing pattern, the low frequency market volatil-
ity declined. This is consistent with the ndings of Campbell
et al. (2001) and suggests a decline in correlations, which is con-
rmed by both the aggregate model-free rolling correlations and
the aggregate FSG-DCC correlations in Figure 3. After 1997,
market and idiosyncratic volatilities seem to move in a similar
fashion with opposite effects on correlations which, at the aggre-
gate level, follow a nonmonotonic pattern. An interesting effect
is observed in the last period where, although market volatility
reached historical lows, correlations decreased only moderately
due to the low levels of idiosyncratic volatility.
We have illustrated that low frequency variation in idiosyn-
cratic volatilities is not negligible and can have important
effects on the level of correlations. We now analyze the eco-
nomic sources of this variation, keeping the analysis at the ag-
gregate level. As mentioned in Section 3, the trend behavior
in idiosyncratic volatility has gained relevance in the literature
following the results of Campbell et al. (2001). At the micro
level, the theoretical framework of Pastor and Veronesi (2003)
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

116 Journal of Business & Economic Statistics, January 2012
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
89-90 91-92 93-94 95-96 97-98 99-00 01-02 03-04 05-06
V
o
l
a
t
i
l
i
t
i
e
s
Period
Average Idiosyncratic Volatility
Average Market Volatility
Figure 2. Average low frequency market and idiosyncratic volatilities for two-year periods. NOTES: The average low frequency volatility for
period p is dened as follows: Average Annualized Idiosyncratic Volatility of Asset i =
1
Tp

Tp
t =1

1/2
it
, where
1/2
it
is the annualized low frequency
volatility for asset i at time t, i = m, 1, 2, . . . , N, and T
p
is the number of daily observations in period p.
suggests a positive relationship between idiosyncratic volatility
and both the variance of rm protability and uncertainty about
the average level of rm protability. In this regard, changes in
the prots of a rm are associated with two main causes. The
rst cause is changes in the rms productivity, which can be
explained by supply-side factors such as technological change
(e.g., Jovanovic 1982) and/or demand-side factors such as prod-
uct substitutability (e.g., Syverson 2004). The other cause is
price variations (of products and factors), which are also related
to interactions between idiosyncratic demand shocks and the
level of competition within the relevant industries. Therefore,
uctuations in demand (which could be related to changes in
taste, technological changes, new trade liberalization policies,
and newregulations, among others) can have an impact on these
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
89-90 91-92 93-94 95-96 97-98 99-00 01-02 03-04 05-06
C
o
r
r
e
l
a
t
i
o
n
s
Period
Average Correlations
Rolling Correlations
FSG-DCC Correlations
Figure 3. Average correlations. NOTES: The average rolling correlation for period p is dened as follows:
Rolling
Tp
=
1
N

N
i=1
{
1
Tp
1
(N1)

j=i

Tp
t =1

Rolling
i,j,t
}, where T
p
denotes the number of daily observations in period p, and
Rolling
i,j,t
is the rolling correlation
between assets i and j at time t. Similarly, the average FSG-DCC correlations are constructed by using
i,j,t
in Proposition 2.
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

Rangel and Engle: FactorSplineGARCH Model 117
.1
.2
.3
.4
.5
.0004
.0006
.0008
.0010
.0012
.0014
.0016
90 92 94 96 98 00 02 04 06
Average Rolling Idiosyncratic Volatility
Average Low frequency Idiosyncratic Volatility
Employment Dispersion Index (EDI)
Figure 4. Employment dispersion index and aggregate idiosyncratic
volatilities. NOTES: The graph shows quarterly aggregates of the fol-
lowing variables: (1) Average Rolling Idiosyncratic Volatility, dened
as ARIV
t
=
1
N

N
i=1
(
1
100

100
k=1
u
2
i(t k)
)
1/2
, where u
it
is the daily id-
iosyncratic return from Equation (1). (2) Average Low Frequency
Volatility dened as Ivol
t
=
1
N

N
i=1

1/2
it
, where the s are daily
low frequency idiosyncratic volatilities (see Equation (10)). (3) Em-
ployment Dispersion Index, which follows the denition as in Lilien
(1982): EDI
k
= {

11
i=1
x
ik
X
k
(log x
ik
log X
k
)
2
}
1/2
, where x
ik
is em-
ployment in industry i (among 11 industry sectors) at quarter k, and X
k
is aggregate employment.
two fundamentals, as well as on their volatility and on the in-
tensity of both rm- and industry-specic news.
At a more aggregate level, the theory of sectoral reallocation
that followed the work of Lilien (1982) explains how random
uctuations in sectoral demand associate with sectoral shifts
in the labor market. The Employment Dispersion Index (EDI)
suggested by Lilien (1982) proxies the intensity of these sec-
toral shifts and can therefore be used as an indicator of id-
iosyncratic news intensity. For example, a technological change
can induce important movements of labor and other production
factors from declining to growing sectors. This sectoral reallo-
cation of resources can be accompanied by a higher intensity of
idiosyncratic shocks and therefore by increases in rm-specic
volatility. Following this intuition, we associate the measure
of Lilien with low frequency variation of aggregate idiosyn-
cratic volatility. We also control for the economic variables
that Guo and Savickas (2006) relate to the aggregate behav-
ior of idiosyncratic volatility, such as the CAY of Lettau and
Ludvigson (2001), the market volatility, and the market liquid-
ity. As in this study, all variables are aggregated at a quarterly
frequency. The S&P 500 excess-returns volatility is used to
proxy the market volatility. As in Chordia, Roll, and Subrah-
manyam (2001), we use the average quoted spread (QSPR) to
measure market liquidity.
We use two model-based measures of aggregate low fre-
quency idiosyncratic volatility: 1) the cross-sectional average
of low frequency idiosyncratic volatilities, aggregated at a quar-
terly level, and 2) the cross-sectional average of rolling moving
averages (based on a 100-day window) of squared idiosyncratic
returns from Equation (10), aggregated at a quarterly level.
Figure 4 shows a graph of these two measures of aggregate
idiosyncratic volatility and Liliens EDI from1990 to 2006. The
visual high correlation is conrmed in Table 1, which reports the
sample Pearsons correlation across our idiosyncratic volatility
measures and the economic explanatory variables mentioned
above. As expected, idiosyncratic volatility is positively corre-
lated with the EDI and the market volatility. The rolling measure
shows correlation coefcients of 0.48 and 0.72, respectively,
whereas the spline measure shows correlations of 0.34 and 0.64,
respectively. Moreover, idiosyncratic volatility is negatively cor-
related with both CAY and market liquidity. Overall, the signs
of these correlations are consistent with the results of Guo and
Savickas (2006), and with the expected effects of sectoral real-
location on the volatility of rms fundamentals. To explore fur-
ther these relationships, we project separately the two measures
of idiosyncratic volatility on the explanatory variables over our
sample period using a linear regression framework. Due to the
nature of the idiosyncratic volatility aggregates, especially the
spline measure, the regressions will be affected by a severe se-
rial correlation problem in the residuals. To lessen this problem
and address endogeneity issues associated with simultaneous
causality, we use the GMM with robust Newey and West (1987)
standard errors, and four lags of the explanatory variables as in-
struments. Table 2 reports the estimated coefcients associated
Table 1. Sample correlations: Idiosyncratic volatilities and economic
variables
Average idiosyncratic
volatilities (quarterly
frequency)
a
Economic variables
d
Average
Rolling
Idiosyncratic
Volatility
b
Average Low
Frequency
Idiosyncratic
Volatility
c
Employment Dispersion Index (EDI) 0.48 0.34
ConsumptionWealth Ratio (CAY) 0.40 0.39
Market factor volatility 0.72 0.64
Illiquidity (QSPR) 0.30 0.23
NOTES: a) Average idiosyncratic volatilities are based on daily DJIA stock returns.
Measures aggregated at a quarterly frequency for the 19902006 period.
b) The Average Rolling Idiosyncratic Volatility is dened as
ARIV
t
=
1
N
N

i=1
_
1
100
100

k=1
u
2
i(t k)
_1/2
,
where u
it
is the daily idiosyncratic return from Equation (1).
c) The Average Low Frequency Idiosyncratic Volatility is dened as
Ivol
t
=
1
N
N

i=1

1/2
it
,
where the s are the daily low frequency volatilities estimated from Equation (10).
d) The EDI follows the denition of Lilien (1982):
EDI
k
=

11

i=1
x
ik
X
k
(log x
ik
log X
k
)
2

1/2
,
where x
ik
is employment in industry i (among 11 industry sectors) at quarter k, and X
k
is
aggregate employment; the CAY is dened as in Lettau and Ludvigson (2001); the market
factor volatility is estimated according to Equation (9); and, the average quoted spread
(QSPR) follows the denition as in Chordia et al. (2001).
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

118 Journal of Business & Economic Statistics, January 2012
Table 2. Idiosyncratic volatility GMM regressions
Average Rolling
Idiosyncratic Volatility
Average Low
Frequency
Idiosyncratic Volatility
Constant 0.082 0.080
(0.037) (0.076)
EDI 104.581 117.298
(53.314) (67.166)
CAY 0.154 0.031
(0.562) (0.699)
Market volatility 0.615 0.643
(0.233) (0.218)
QSPR 0.083 0.002
(0.073) (0.087)
R
2
0.533 0.365
J-statistic 0.051 0.049
NOTES: The Average Rolling Idiosyncratic Volatility is dened as
ARIV
t
=
1
N
N

i=1
_
1
100
100

k=1
u
2
i(t k)
_1/2
,
where u
it
is the daily idiosyncratic return from Equation (1). The Average Low Frequency
Idiosyncratic Volatility is dened as
Ivol
t
=
1
N
N

i=1

1/2
it
,
where the s are the daily low frequency volatilities estimated from Equation (10). These
averages are based on daily DJIA stock returns. Aggregating the variables at a quarterly
frequency from 1989 to 2006, the two measures of idiosyncratic volatility are regressed on
the following set of variables: the EDI of Lilien (1982), the CAY of Lettau and Ludvigson
(2001), the market factor volatility estimated from Equation (9), and the average quoted
spread (QSPR) of NYSE stocks (as dened in Chordia et al. (2001)). The regressions are
estimated using the Generalized Method of Moments (GMM) with NeweyWest standard
errors and four lags of the regressors as instruments (Standard errors are reported in
parentheses, bold font indicates statistical signicance at the 10% level).
with the two linear projections. The two regressions suggest the
same effects and, as in the sample correlation analysis, the es-
timated coefcients show the expected sign. Nevertheless, only
the EDI and the market volatility are statistically signicant. It is
important to mention that the EDI series is highly noisy before
the mid-1980s. Structural breaks should therefore be taken into
account for analyses incorporating longer sample periods.
4.4 Empirical Fit of Factor Correlation Models
This subsection evaluates a range of one-factor models with
varying dynamic components in terms of their empirical t using
our sample of 33 Dow stocks. The process follows a simple-to-
general strategy. We start from the simplest case, labeled FC-C,
where factor and idiosyncratic volatilities are constant over the
sample period (at high and low frequencies), and restrictions in
Equations (2) and (3) apply. We estimate the return correlations
fromthis model and use a Gaussian metric to compute the quasi-
likelihood. We then consider subsequent models that relax one
or more assumptions of the initial model, and obtain their quasi-
likelihoods. The last step involves comparing the empirical t
of this range of factor correlation models based on this Gaussian
metric. This allows us to assess which restrictions in the factor
structure are the most important to describe correlation behavior.
The assumptions to be weakened are: 1) constant volatility of
the market factor, 2) constant volatility of the idiosyncratic com-
ponents of returns, 3) no omitted factors, and 4) constant betas.
By adding high frequency variation to the volatility of the
factor through a standard GARCH process, and keeping the id-
iosyncratic volatilities constant, we obtain a specication called
FG-C. Similarly, when high and low frequency splineGARCH
dynamics are added to the market factor volatility holding the
idiosyncratic volatilities constant, we obtain the FSG-C model.
These models and their correlations, along with a range of spec-
ications derived from adding dynamics to the previous as-
sumptions, are described in Table 3. Their quasi-likelihoods are
constructed fromthe general factor structure in Equation (16) as-
suming that r
t
|
t 1
N(0, H
t
). A mapping of each correlation
specication in Table 3 with a specic covariance matrix pro-
vides the inputs to compute the quasi-likelihood of each model.
We estimate the models in the rst panel of Table 3 using QML.
For the other models, we follow a two-stage GMM approach
with Gaussian moment conditions.
Table 4 presents the log quasi-likelihoods of the factor mod-
els in Table 3, along with likelihood ratio tests that compare
each model with the biggest FSG-DCC model. The results in-
dicate that the FSG-DCC model dominates the other speci-
cations. Close to this model is the FSG-IDCC model, which
assumes constant betas, but allows for aggregate effects of un-
specied unobservable factors and both high and low frequency
variation in market and idiosyncratic volatilities. The FG-DCC
model follows in the list. The three biggest models show the
best empirical t even when they are penalized by the BIC (see
Table 4). In contrast, the specication with poorest empirical
t is the constant correlation model (FC-C). Thus, we nd that
models with low frequency dynamics dominate those with only
high frequency dynamics. The results also show large improve-
ments in the quasi-likelihoods when we relax the assumptions
of constant idiosyncratic volatilities and no omitted factors.
This conrms that, besides the importance of modeling market
behavior, adding dynamic features to the second moments of id-
iosyncratic terms largely improves the empirical t of this class
of one-factor CAPM models.
5. FORECAST PERFORMANCE OF THE
FACTORSPLINEGARCH MODEL
Our forecast comparison follows the approach of Engle and
Colacito (2006) by using an economic loss function to assess
the performance of each model. However, different from them,
we use forward portfolios based on forecasts of the covariance
matrix associated with each of the models to be compared.
Specically, we focus on a portfolio problem where an investor
wants to optimize today his/her forward asset allocation given
a forward conditional covariance matrix. In the classical mean
variance setup, this problem can be formulated as
min
w
t +k|t
w

t +k|t
H
t +k|t
w
t +k|t
s.t. w

t +k|t
=
0
, (20)
where is the vector of expected excess returns,
0
is the
required return, w
t +k|t
denotes a portfolio at time t + k that was
formed using the information at time t, and H
t +k|t
is a k-step
ahead forecast (at time t) of the conditional covariance matrix
of excess returns. The solution to Equation (20) therefore is
w
t +k|t
=
H
1
t +k|t

H
1
t +k|t

0
, (21)
and represents optimal forward portfolio weights given the in-
formation at time t. Each covariance forecast H
t +k|t
implies a
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

Rangel and Engle: FactorSplineGARCH Model 119
Table 3. Correlation models from factor assumptions
Panel 1: Dynamic volatility components
a
Constant volatilities
Components with high and low frequency dynamics
Factor component Factor and idiosyncratic components
FC-C: Constant factor and
idiosyncratic volatilities
FG-C: GARCH factor volatility FG-G: GARCH factor and
idiosyncratic volatilities

FGC
i,j,t
=

i

j
hmt

2
i
hmt +
2
i
_

2
j
hmt +
2
j

FGG
i,j,t
=

i

j
hmt

2
i
hmt +h
it
_

2
j
hmt +h
jt

FCC
i,j
=

i

2
m

2
i

2
m
+
2
i
_

2
j

2
m
+
2
j
FSG-C: SplineGARCH factor volatility FSG-SG: SplineGARCH factor
and idiosyncratic volatilities

FSGC
i,j,t
=

i

j
mt gmt

2
i
mt gmt +
2
i
_

2
j
mt gmt +
2
j

FSGSG
i,j,t
=

i

j
mt gmt

2
i
mt gmt +
it
g
it
_

2
j
mt gmt +
jt
g
jt
Panel 2: Other dynamic components
b
Idiosyncratic correlations (latent unobserved factors in the error term) All components
FG-IDCC: FG-G with latent omitted factors FG-DCC

FGIDCC
i,j,t
=

i

j
hmt +

h
it

h
jt

u
i,j,t

2
i
hmt +h
it
_

2
j
hmt +h
jt
See Equation (15)
FSG-IDCC: FSG-SG with latent omitted factors FSG-DCC

FSGIDCC
i,j,t
=

i

j
mt gmt +

it
g
it

jt
g
jt

u
i,j,t

2
i
mt gmt +
it
g
it
_

2
j
mt gmt +
jt
g
jt
See Equation (12)
NOTES: a) denotes constant (C) volatilities, h and g refer to GARCH (G) and SplineGARCH (SG) variances, respectively.
b) These models are parameterizations of Equations (4) and (5) in Proposition 1.
particular forward portfolio w
t +k|t
given a vector of expected
returns. However, an important issue arises when the true vec-
tor of expected returns is unknown. Engle and Colacito (2006)
mentioned that a direct comparison of optimal portfolio volatil-
ities can be misleading if a particular estimate of the expected
excess returns, such as their realized mean, is used to com-
pute such volatilities. Their framework isolates the effect of
covariance information by using a wide range of alternatives
for the vector of expected excess returns and the asymptotic
properties of sample standard deviations of optimized portfo-
lio returns. We follow their approach and consider different
vectors of expected excess returns associated with a variety of
multivariate hedges that are constructed by holding one asset
for return and using the other assets for hedge. We then com-
pare the standard deviations of returns on long-term forward
hedge portfolios formed with each models covariance matrix
forecast.
5.1 Out-of-Sample Evaluation: A Sequential
Forecasting Exercise
We examine the out-of-sample forecast performance of the
FSG-DCC model using our group of 33 Dow stocks. Our anal-
ysis implements a sequential exercise that is described in Fig-
ure 5 and can be characterized by iterations of the following
procedure: 1) a set of models is estimated based on an initial
Table 4. Evaluation of factor correlation models
Assumption
a
Model Quasi-log- likelihood Parameters Quasi-likelihood ratio
c
BIC
FC-C 426550 67 45980

186.96
1 FG-C 427110 69 44860

187.20
FSG-C 427160 75 44760

187.21
2 FG-G 440670 135 17740

193.03
FSG-SG 442820 383 13440

193.51
3 FG-IDCC 446930 663 5220

194.80
FSG-IDCC 449100 911 880

195.29
4
b
FG-DCC 447330 696 4420

194.91
FSG-DCC 449540 944 195.42
NOTES: a) This column corresponds to the assumptions that are weakened in the factor specications:
1. Constant volatility of the market factor.
2. Constant volatility of the idiosyncratic components.
3. No omitted factors.
4. Constant betas.
The models are described in Table 3 and the sample includes the 33 DJIA stocks described in Appendix A3 (Table A3-1).
b) The last step is associated with the full FG-DCC and FSG-DCC models.
c) The quasi-likelihood ratios compare the models on each row with the FSG-DCC model (see last row).

) Indicates that the quasi-likelihood ratios are above the 1% critical value of a chi-square distribution with degrees of freedom given by the difference between 944 and the number of
parameters associated with each model.
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

120 Journal of Business & Economic Statistics, January 2012
Figure 5. Sequential forecasting exercise. NOTES: This gure describes the iterations of an out-of-sample sequential forecasting exercise.
In the rst iteration, a number of covariance models are estimated using data from 12/01/1988 to 12/31/1995. Then, multistep out-of-sample
forecasts are constructed for each model during a six-month period (from 01/01/1996 to 06/30/1996). In iteration 2, the estimation period is
extended six months (from 12/01/1988 to 06/30/1996) and a new set of multistep out-of-sample forecasts are generated for the following six
months. At each iteration, the estimation period incorporates the previous iterations forecasting period and new out-of-sample forecasts are
generated. Twenty-two iterations are completed, ending at 12/31/2006. The forecasting periods do not overlap.
period with T
0
daily observations, and multihorizon covariance
forecasts from1 to 126 days ahead are computed out of the sam-
ple, 2) the 126-day forecasting period is incorporated into a new
estimation period with T
1
= T
0
+126 daily observations, the
models are re-estimated and, nally, a new set of out-of-sample
forecasts is constructed for the following six months. We iter-
ate these two steps several times starting with a sample period
from December 1988 to June 1995. The process is repeated 22
times (up to December 2006) and none of the out-of-sample
forecast blocks overlap (see Figure 5). The exercise consid-
ers the FSG-DCC specication and ve competing models: 1)
the stationary DCC, 2) the FG-DCC, 3) the sample covariance
(SCOV), 4) the static one-factor beta covariance (BCOV), and
5) the optimal shrinkage covariance of Ledoit and Wolf (2003)
(LCOV). The last three models are analyzed in Jagannathan and
Ma (2003). All models are re-estimated at each iteration. For
the FSG-DCC model, we extrapolate the spline functions by
keeping them constant at their last values during the forecasting
periods and restricting them to have a zero-slope in the last ob-
servation of the sample. By providing a conservative approach
that underestimates the behavior of the series near the end of
the sample, the zero-slope condition helps to reduce potential
anomalies caused by outliers near the right boundary (e.g., Sil-
verman 1984 and Nychka 1995 illustrate the boundary issues
associated with smoothing splines).
We focus on long-horizon covariance forecasts to form op-
timized forward hedge portfolios according to (21), using a
variety of vectors of expected returns associated with different
hedges and a required return normalized to 1. From the multi-
horizon forecasts generated at each iteration, we only consider
the last 40 days (i.e., out-of-sample forecast from87 to 126 days
ahead). We compute standard deviations of returns on these out-
of-sample portfolios as

(j)
p,OS
=

22
i=1

126
k=87
_
w
(j)
p,T
i
+k|T
i
(r
T
i
+k
r)
_
2
22 40
,
j = FSG, FG-DCC, DCC, SCOV,
BCOV, LCOV,
p = 1, 2, . . . , 33,
(22)
where T
i
is the last day of the estimation period associated with
iteration i, r denotes the sample mean of daily excess returns,
r
T
i
+k
is the vector of one-day excess returns at day T
i
+k, and
w
(j)
p,T
i
+k|T
i
corresponds to a T
i
+k-day forward hedge portfolio
constructed from covariance model j, where asset p is hedged
against the other assets.
Table 5 reports these standard deviations for each model and
hedge portfolio. The FSG-DCC specication produces smallest
volatilities for 15 hedges; the sample covariance is preferred for
ve hedges; the shrinkage covariance and the FG-DCC models
dominate in four cases each; the DCC model is preferred in
three cases; and, the static beta covariance model dominates in
only one case. These results suggest that the FSG-DCC model
is the preferred model in most of the cases. Furthermore, the
average across all hedges (shown in the last row of the table)
also indicates that the best performer is the FSG-DCCspecica-
tion. We also performed an in-sample forecasting exercise and
obtained the same conclusions (see online Appendix A5). Since
the differences across models appear small, we examine further
the signicance of these results by performing joint Diebold
and Mariano (1995) style tests following Engle and Colacito
(2006). The purpose is to test the equality of the FSG-DCC
model with respect to each of its competitors. This approach is
based on statistical inference about the mean of the difference
between square returns on optimized portfolios generated by
the FSG-DCC specication and a competitor model m. For each
iteration i (with last observation = T
i
), a vector of difference
series associated with hedge p is dened as
u
p,m
T
i
=
_
_
w
FSGDCC
p,T
i
+k|T
i
(r
T
i
+k
r)
_
2

_
w
m
f h,T
i
+k|T
i
(r
T
i
+k
r)
_
2
, k = 87, . . . , 126
_
, (23)
where m = FG-DCC, DCC, SCOV, LCOV, BCOV, and p =
1, 2, . . . , 33. Using these hedge difference vectors, we construct
a joint difference vector that stacks all of them as follows:
U
FSGDCC,m
T
i
=
_
u
1,m
T
i
, u
2,m
T
i
, . . . , u
33,m
T
i
_
, i = 1, 2, . . . , 22.
(24)
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

Rangel and Engle: FactorSplineGARCH Model 121
Table 5. Out-of-sample evaluation: Standard deviations of optimized
forward hedge
Portfolios at multiple long horizons
Forward
hedge
portfolios FSG-DCC FG-DCC DCC BCOV SCOV LCOV

AA
0.2573 0.2569 0.2569 0.2787 0.2549 0.2550

AIG
0.1914 0.1947 0.1917 0.1949 0.1925 0.1923

AXP
0.2130 0.2138 0.2137 0.2293 0.2151 0.2152

BA
0.2690 0.2702 0.2697 0.2744 0.2715 0.2714

C
0.1979 0.1992 0.1985 0.2182 0.1979 0.1983

CAT
0.2445 0.2461 0.2452 0.2532 0.2455 0.2451

CVX
0.1593 0.1586 0.1575 0.2065 0.1584 0.1592

DD
0.1997 0.2009 0.2013 0.2119 0.1981 0.1980

DIS
0.2718 0.2695 0.2708 0.2708 0.2726 0.2723

GE
0.1921 0.1947 0.1918 0.1874 0.1913 0.1909

GM
0.3035 0.3061 0.3055 0.3069 0.3043 0.3040

GT
0.3509 0.3504 0.3508 0.3645 0.3547 0.3546

HD
0.2704 0.2707 0.2705 0.2837 0.2703 0.2704

HON
0.2714 0.2756 0.2719 0.2812 0.2718 0.2718

HPQ
0.3688 0.3735 0.3677 0.3925 0.3665 0.3668

IBM
0.2363 0.2428 0.2367 0.2506 0.2388 0.2386

INTC
0.3325 0.3394 0.3347 0.3794 0.3351 0.3357

IP
0.2292 0.2296 0.2287 0.2531 0.2279 0.2283

JNJ
0.1792 0.1808 0.1797 0.1972 0.1803 0.1797

JPM
0.2169 0.2183 0.2190 0.2427 0.2187 0.2189

KO
0.2161 0.2163 0.2149 0.2232 0.2150 0.2150

MCD
0.2522 0.2541 0.2525 0.2531 0.2525 0.2523

MMM
0.1937 0.1965 0.1938 0.1977 0.1925 0.1924

MO
0.3172 0.3204 0.3174 0.3210 0.3176 0.3176

MRK
0.2156 0.2162 0.2156 0.2417 0.2142 0.2141

MSFT
0.2474 0.2510 0.2485 0.2711 0.2485 0.2481

PFE
0.2439 0.2482 0.2438 0.2710 0.2440 0.2439

PG
0.1827 0.1826 0.1824 0.1991 0.1820 0.1820

T
0.2139 0.2161 0.2145 0.2522 0.2158 0.2168

UTX
0.2158 0.2204 0.2166 0.2195 0.2157 0.2154

VZ
0.1924 0.1953 0.1913 0.2237 0.1901 0.1904

WMT
0.2351 0.2370 0.2349 0.2500 0.2364 0.2366

XOM
0.1517 0.1506 0.1513 0.1966 0.1518 0.1528
All portfolios 0.2374 0.23757 0.2393 0.2545 0.23764 0.2377
NOTES: Sample standard deviations of returns on optimized forward hedge portfolios
subject to a required return of 1 and based on 22 iterations of out-of-sample covariance
forecasts at horizons from 87 to 126 days ahead. The forecasts are constructed from
FSG-DCC, FG-DCC, DCC, BCOV (static one-factor beta covariance), SCOV (sample
covariance), and LCOV (optimal shrinkage covariance of Ledoit and Wolf (2003)) models,
respectively. The 22 sequential sample periods are described in Figure 5 and include the 33
DJIA stocks described in Appendix A3 (Table A3-1). The stock in the corresponding row
is hedged against the rest of the stocks.
The null of equality of covariance models tests that the
mean of U
FSGDCC,m
t
equals zero. For each comparison, the
test is performed by running a regression of U
FSGDCC,m
t
on
a constant. The regressions are estimated by GMM using
robust heteroscedasticity and autocorrelation consistent (HAC)
covariance matrices. Table 6 reports the t-statistics for the
DieboldMariano tests (each column names a competing
model). A negative value suggests that the FSG-DCC model
dominates over the column model. At a 5% condence level,
the results indicate once more that the FSG-DCC dominates its
competitors at long horizons.
Table 6. Joint DieboldMariano tests to compare the long-term
forecast performance of the FSG-DCC model relative to competing
models
Competing models
FSG-DCC
vs. column
model DCC FG-DCC LCOV SCOV BCOV
t-statistics 3.66 2.96 2.14 2.07 5.72
NOTES: This table reports t-statistics for joint DieboldMariano tests that evaluate the
forecast performance of the FSG-DCC model in relation to the following competitors:
DCC, FG-DCC, BCOV (static one-factor beta covariance), SCOV (sample covariance),
and LCOV (optimal shrinkage covariance of Ledoit and Wolf (2003)). Each t-statistic is
derived from estimating a regression of a vector of differences of squared realized returns
of two models on a constant. The vector of differences is constructed using squared realized
forward hedge portfolio returns associated with the FSG-DCC model and the competing
column model. The forward hedge portfolios are constructed using the sample periods
described in Figure 5 and the 33 DJIA stocks in Appendix A3 (Table A3-1). We include
only long-term forward hedge portfolios (87126 days forward). The vector regressions
are estimated using the GMM with a HAC covariance matrix. A negative value indicates a
better performance of the FSG-DCC model.
6. CONCLUDING REMARKS
This article develops a new model for asset correlations that
characterizes dynamic patterns at high and low frequencies
in the correlation structure of equity returns. By exploiting
a factor asset pricing structure and the dynamic properties
of low frequency volatilities associated with systematic and
idiosyncratic terms, we introduce a slow-moving correlation
component that proxies low frequency changes in correlations.
Our semiparametric approach generalizes DCC and other
multivariate GARCH models by allowing the high frequency
correlation component to mean revert toward a time-varying
low frequency component. This framework allows the level
to which conditional correlations mean revert to adapt to the
varying economic conditions.
At high frequencies, our model incorporates dynamic effects
that arise from relaxing assumptions in the standard one-factor
CAPM model. These effects account for time-varying betas
and missing pricing factors. At low frequencies, the long-term
trends of the market and the idiosyncratic volatilities govern the
dynamics of low frequency correlations. We provide evidence
that, in addition to the recently documented economic variation
of market volatility at low frequencies, average idiosyncratic
volatility exhibits substantial variation in its long-termtrend. We
nd that this variation is highly correlated with low frequency
economic variables including an intersectoral employment dis-
persion index that proxies the intensity of sectoral reallocation
of resources in the economy and, since these movements are
mainly driven by shocks that are specic to either individual
rms or sectors, it also serves as an indicator of idiosyncratic
news intensity.
The ability of our correlation model to incorporate non-
parametrically low frequency features improves not only the
empirical t of equity correlations and their association with
economic conditions, but the forecasting of correlations as well.
Indeed, in-sample and out-of-sample forecasting experiments
indicate that, at long horizons, this newmodel with time-varying
long-term trends outperforms standard models that mean revert
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

122 Journal of Business & Economic Statistics, January 2012
to xed levels. This result is explained by the models exibility
in adjusting the level of mean reversion to the varying economic
conditions.
APPENDIX A1
Proof of Proposition 1. Part a): Consider the specication in
Equation (1), and assumptions (2) and (3). If E
t 1
(r
m,t
u
it
) = 0,
i, but E
t 1
(u
it
u
jt
) = 0, for some i = j, then cov
t 1
(r
i
, r
j
) =

j
V
t 1
(r
mt
) +E
t 1
(u
it
u
jt
), and Equation (4) follows from the
denition of conditional correlation.
Part b): Given that E
t 1
(u
it
u
jt
) = 0, for some i = j,and
E
t 1
(r
m,t
u
it
) = 0, for some i, then the relevant covariance and
variance terms are
cov
t 1
(r
i
, r
j
) =
i

j
V
t 1
(r
mt
) +
i
E
t 1
(r
mt
u
jt
)
+
j
E
t 1
(r
mt
u
it
) +E
t 1
(u
it
u
jt
), (A.1)
and
V
t 1
(r
kt
) =
2
k
V
t 1
(r
mt
) +V
t 1
(u
kt
)
+2
k
E
t 1
(r
mt
u
kt
), k = i, j. (A.2)
Equation (5) is thus obtained by combining Equations (A.1)
and (A.2) with the denition of conditional correlation.
Part c): Under the representation in Equation (6), the condi-
tional covariance between market returns and the idiosyncratic
residual will be zero. In fact, using the denition of conditional
beta:
cov
t 1
(r
it
, r
mt
) =
it
V
t 1
(r
mt
), (A.3)
then, substituting the return equation in Equation (6) into this
expression gives
cov
t 1
(
it
r
mt
+ u
it
, r
mt
) =
it
V
t 1
(r
mt
) +cov
t 1
( u
it
, r
mt
).
(A.4)
Thus, from equalizing Equations (A.3) and (A.4), we have
cov
t 1
( u
it
, r
mt
) = 0, (A.5)
and Equation (2) is satised. Now, rewriting Equation (6) as
r
it
=
i
r
mt
+u
it
,
i
=
cov (r
it
, r
mt
)
V (r
mt
)
,
u
it
= u
it
+(
it

i
) r
mt
, i, (A.6)
and using Equations (A.3) and (A.5), it is clear that the newerror
term in Equation (A.6) also satises assumption (2). Moreover,
combining Equations (A.5) and (A.6) yields the following con-
ditional moments:
cov
t 1
(u
it
, r
mt
) = (
it

i
)V
t 1
(r
mt
), i, (A.7)
cov
t 1
(u
it
, u
jt
) = cov
t 1
( u
it
, u
jt
)
+(
it

i
)(
jt

j
)V
t 1
(r
mt
). (A.8)
The numerator in (5) can be rewritten as
cov
t 1
(r
i
, r
j
) =
i

j
V
t 1
(r
mt
) +
i
cov
t 1
(r
mt
, u
jt
)
+
j
cov
t 1
(r
mt
, u
it
) +cov
t 1
(u
it
, u
jt
), (A.9)
and by substituting Equations (A.7) and (A.8) into this equation,
the numerator in Equation (7) is obtained.
Similarly, the expressions in the denominator of Equation (5)
can be rewritten as
V
t 1
(r
kt
) =
2
k
V
t 1
(r
mt
) +V
t 1
(u
kt
)
+2
k
cov
t 1
(r
mt
, u
kt
), k = i, j. (A.10)
Now, from Equations (A.5) and (A.6), V
t 1
(u
kt
) =
V
t 1
( u
kt
) +(
kt

k
)
2
V
t 1
(r
mt
), k = i, j. Hence, plugging
these expressions and Equation (A.7) into Equation (A.10) leads
to the denominator in (7).
APPENDIX A2
Proof of Proposition 2. Consider the following vectors of re-
turns, factor loadings, and innovations: r
t
= (r
1t
, r
2t
, . . . , r
Nt
)

,
= (
1t
,
2t
, . . . ,
Nt
)

, and u
t
= (u
1t
, u
2t
, . . . , u
Nt
)

. Given a
vector F
t
of common factor(s) and omitting the constant terms,
without loss of generality, we can rewrite the model in Equation
(1) as
r
t
= F
t
+u
t
. (A.11)
Given the t1 information set
t 1
, the conditional covari-
ance matrix therefore is
E
t 1
(r
t
r

t
) = E
t 1
(F
t
F

t
)

+E
t 1
(F
t
u

t
)
+E
t 1
(u
t
F

t
)

+E
t 1
(u
t
u

t
). (A.12)
In particular, for the one-factor CAPM case, F
t
= r
mt
and
Equation (A.12) takes the following form:
E
t 1
(r
t
r

t
) = V
t 1
(r
mt
)

+E
t 1
(r
mt
u

t
)
+E
t 1
(u
t
r
mt
)

+E
t 1
(u
t
u

t
). (A.13)
From Equation (9), the typical (i, j) element of the rst term
on the right-hand side (RHS) of Equation (A.13) is

j
V
t 1
(r
mt
) =
i

mt
g
mt
. (A.14)
Similarly, from Equations (9)(11), the typical (i, j) element
of the second term is

i
E
t 1
(r
mt
u
jt
) =
i

mt
g
mt

jt
g
jt

m,j,t
, (A.15)
the typical (i,j) element of the third term is

j
E
t 1
(r
mt
u
it
) =
j

mt
g
mt

it
g
it

m,i,t
, (A.16)
and, the typical (i,j) element of the last term is
E
t 1
(u
it
u
jt
) =

it
g
it

jt
g
jt

i,j,t
. (A.17)
Equation (12) follows by plugging these conditional expec-
tations and V
t 1
(u
kt
) =
kt
g
kt
, k = i, j, into Equation (6).
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

Rangel and Engle: FactorSplineGARCH Model 123
The unconditional version of Equation (5) is then used to
derive the low frequency correlation

i,j
=
_

j
V(r
mt
) +
j
E(r
mt
u
it
) +
i
E(r
mt
u
jt
) +E(u
it
u
jt
)
_

2
i
V(r
mt
) +V(u
it
) +2
i
E(r
mt
u
it
)
_
1/2

2
j
V(r
mt
) +V(u
jt
) +2
j
E(r
mt
u
jt
)
_
1/2
Under the assumption that both the factor(s) and the idiosyn-
crasies are unconditionally uncorrelated, we have

i,j
=

i

j
V (r
mt
)+E(u
it
u
jt
)
_

2
i
V(r
mt
)+V(u
it
)
_

2
j
V(r
mt
)+V(u
jt
)
, (A.18)
and, from (9), (10), and the law of iterated expectations (LIE):
V(r
mt
) =
mt
E(g
mt
) =
mt
,
and
V(u
it
) =
it
E(g
it
) =
it
, i = 1, 2, . . . , N.
Also,
E(u
it
u
jt
)

it

jt
=
E
_
g
1/2
it

it
g
1/2
jt

jt
_

E(g
it
)
_
E(g
jt
)
= corr
_
g
1/2
it

it
, g
1/2
jt

jt
_


i,j
.
Note that

i,j,t
=

i,j,t
, t , thus we approximate

i,j
with the
sample correlation,

i,j
, from Equation (11). Plugging the pre-
vious expressions into Equation (A.18), we obtain the time-
varying low frequency correlation in Equation (13).
Moreover, if we assume E
t
(
k,t +h
) =
k,t
, h > 0, k =
1, 2, . . . , N, then the long-horizon forecast of Equation (12)
can be constructed using the mean reversion properties of the
GARCHandDCCequations. Indeed, the GARCHdynamics im-
ply lim
h
g
k,t +h|h
= 1, k = m, 1, 2, . . . , N. The long-horizon
correlation forecasts associated with the vector of innovations
are also given by the terms targeting correlations (see Equation
(11)):
lim
h

i,j,t +h
= lim
h
q
i,j,t +h|t

q
i,i,t +h|t

q
j,j,t +h|t
=

i,j
_

i,i
_

j,j
=

i,j
, (A.19)
i, j {1, 2, . . . , N}. Also, from our assumption that the id-
iosyncrasies are unconditionally uncorrelated with the fac-
tor,

m,i
= 0 and lim
h

m,i,t +h|t
= 0, i = 1, . . . , N. Hence, by
substituting the long-run forecasts of each termin Equation (12),
we obtain
lim
h

i,j,t +h|t
=

mt
+

it

jt

i,j
_

2
i

mt
+
it
_

2
j

mt
+
jt
, (A.20)
which coincides with the low frequency correlation.
SUPPLEMENTAL MATERIALS
This appendix expands on the material presented in three sec-
tions of the paper: 1) it provides a more detailed description of
the stocks used in our empirical analyses and the FSG-DCC es-
timates described in Subsection 4.2, 2) it presents an empirical
exercise that extends the results in Subsection 4.3 by exploring
the association between sectoral idiosyncratic volatilities and
economic variables, and 3) it presents results from an in-sample
forecasting evaluation that complements the out-of-sample re-
sults presented in Subsection 5.1.
ACKNOWLEDGMENTS
We thank Siddhartha Chib, Stephen Figlewski, Eric Ghy-
sels, Massimo Guidolin, Andrew Karolyi, Andrew Patton, Jeff
Russell, Kevin Sheppard, and seminar/conference participants
at the Federal Reserve Bank of New York, LSE, NYU Stern,
Ohio State University, the University of Arizona, Cambridge,
Chicago, Manchester, Warwick, the 2007 LAMES conference,
the 2007 Multivariate Volatility Models Conference, the 2008
Frank Batten Young Scholar Conference in Finance, the 2008
Stanford University SITEConference, and the 2009 SoFiECon-
ference for comments and suggestions. All errors are the respon-
sibility of the authors. The nal draft of this paper was written
while Jos e Gonzalo Rangel was working at the Central Bank of
Mexico. The opinions expressed in this article are solely those
of the authors and do not necessarily reect the views of the
Central Bank of Mexico or Goldman Sachs Group, Inc.
[Received May 2009. Accepted May 2011.]
REFERENCES
Alexander, C. O. (1998), Volatility and Correlation: Methods, Models and
Applications, in Risk Management and Analysis: Measuring and Mod-
elling Financial Risk, ed. C. O. Alexander, New York: Wiley, pp. 125
172. [109]
Ang, A., and Bekaert, G. (2002), International Asset Allocation With Regime
Shifts, Review of Financial Studies, 15, 11371187. [109]
Ang, A., and Chen, J. (2002), Asymmetric Correlations of Equity Portfolios,
Journal of Financial Economics, 63, 443494. [109]
Bauwens, L., Laurent, S., and Rombouts, J. (2006), Multivariate GARCH
Models: A Survey, Journal of Applied Econometrics, 21, 79109. [109]
Bekaert, G., Hodrick, R. J., and Zhang, X. (2009), International Stock Returns
Comovements, Journal of Finance, 64, 25912626. [109]
Black, F. (1976), Studies of Stock Price Volatility Changes, Proceedings
of the 1976 Meetings of the American Statistical Association, Business
and Economics Statistics Section, Alexandria, VA: American Statistical
Association, pp. 177181. [111,114]
Bollerslev, T. (1987), A Conditionally Heteroskedastic Time Series Model for
Speculative Prices and Rates of Return, Reviewof Economics and Statistics,
69, 542547. [113]
(1990), Modelling the Coherence in Short-Run Nominal Exchange
Rates: A Multivariate Generalized ARCH Model, Review of Economics
and Statistics, 72, 498505. [109]
Bollerslev, T., Engle, R., and Wooldridge, J. M. (1988), ACapital Asset Pricing
Model With Time Varying Covariances, Journal of Political Economy, 96,
116131. [109]
Bollerslev, T., and Wooldridge, J. (1992), Quasi-Maximum Likelihood Esti-
mation and Inference in Dynamic Models With Time-Varying Covariances,
Econometric Reviews, 11, 143172. [113]
Bos, T., and Newbold, P. (1984), An Empirical Investigation of the Possibility
of Systematic Risk in the Market Model, Journal of Business, 57, 3541.
[110]
Campbell, J., and Hentschel, L. (1992), No News Is Good News: An Asym-
metric Model of Changing Volatility in Stock Returns, Journal of Financial
Economics, 31, 281318. [111]
Campbell, J., Lettau, M., Malkiel, B., and Xu, Y. (2001), Have Individual
Stocks Become More Volatile? An Empirical Exploration of Idiosyncratic
Risk, Journal of Finance, 56, 143. [111,115,115]
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2

124 Journal of Business & Economic Statistics, January 2012
Cao, C., Simin, T., and Zhao, J. (2008), Can Growth Options Explain the
Trend in Idiosyncratic Risk?, Review of Financial Studies, 21, 2599
2633. [111]
Chib, S., Nardari, F., and Shephard, N. (2006), Analysis of High Dimensional
Multivariate Stochastic Volatility Models, Journal of Econometrics, 134,
341371. [109]
Chordia, T., Roll, R., and Subrahmanyam, A. (2001), Market Liquidity and
Trading Activity, Journal of Finance, 56, 501530. [117]
Christie, A. (1982), The Stochastic Behavior of Common Stock Variances:
Value, Leverage and Interest Rate Effects, Journal of Financial Economics,
10, 407432. [111,114]
Diebold, F., and Mariano, R.(1995), Comparing Predictive Accuracy, Journal
of Business and Economic Statistics, 13, 253263. [120]
Diebold, F., and Nerlove, M. (1989), The Dynamics of Exchange Rate Volatil-
ity: AMultivariate Latent-Factor ARCHModel, Journal of Applied Econo-
metrics, 4, 122. [109]
Engle, R. (2002), Dynamic Conditional CorrelationA Simple Class of Mul-
tivariate GARCH Models, Journal of Business and Economic Statistics,
17, 239250. [109,112]
(2007), High Dimension Dynamic Correlations, Prepared
for a Festschrift for David Hendry, Oxford: Oxford University.
[110,113]
Engle, R., and Colacito R. (2006), Testing and Valuing Dynamic Correlations
for Asset Allocation, Journal of Business and Economic Statistics, 24,
238253. [110,118,120]
Engle, R., Ghysels, E., and Sohn, B. (2008), On the Economic Sources of Stock
Market Volatility, Working Paper, New York University. [111,114]
Engle, R., Hendry, D., and Richard J. (1983), Exogeneity, Econometrica, 51,
277304. [113]
Engle, R., Ng, V., and Rothschild M. (1990), Asset Pricing With a Factor ARCH
Covariance Structure: Empirical Estimates for Treasury Bills, Journal of
Econometrics, 45, 213238. [109]
Engle, R., and Rangel, J. G. (2008), The SplineGARCH Model for Low
Frequency Volatility and Its Global Macroeconomic Causes, Review of
Financial Studies, 21, 11871222. [109,110,111,113,114]
Engle, R., and Rangel, J. G. (2009), High and Low Frequency Correlations in
Global Equity Markets, Working Paper, Banco de M exico. [112]
Engle, R., and Sheppard, K. (2005a), Evaluating the Specications of Covari-
ance Models for Large Portfolios, Working Paper, University of Oxford.
[112]
(2005b), Theoretical Properties of Dynamic Conditional Correlation
Multivariate GARCH, Working Paper, University of Oxford. [114]
Engle, R., Shephard, N., and Sheppard, K. (2008), Fitting and Testing Vast
Dimensional Time-Varying Covariance Models, unpublished manuscript.
[114]
Fama, E., and French, K. (1997), Industry Costs of Equity, Journal of Finan-
cial Economics, 43, 153193. [114]
Ferson, W., and Harvey, C. (1991), The Variation of Economic Risk Premium,
Journal of Political Economy, 99, 385415. [110]
(1993), The Risk and Predictability of International Equity Returns,
Review of Financial Studies, 6, 527566. [110]
Ghysels, E. (1998), On Stable Factor Structures in the Pricing of Risk: Do
Time Varying Betas Help or Hurt?, Journal of Finance, 53, 549573. [110]
Glosten, L., Jagannathan, R., and Runkle, D. (1993), On the Relation between
the Expected Value and the Volatility of the Nominal Excess Return on
Stocks, Journal of Finance, 48, 17791801. [111]
Guo, H., and Savickas, R. (2006), Idiosyncratic Volatility, Stock Market Volatil-
ity, and Expected Stock Returns, Journal of Business and Economic Statis-
tics, 24, 4356 [111,117]
Harvey, A., Ruiz, E., and Shephard, N. (1994), Multivariate Stochastic Variance
Models, Review of Economic Studies, 61, 247264. [109]
Jagannathan, R., and Ma, T. (2003), Risk Reduction in Large Portfolios: Why
Imposing the Wrong Constraints Helps, Journal of Finance, 54, 1651
1684. [120]
Jovanovic, B. (1982), Selection and Evolution of Industry, Econometrica, 50,
649670. [116]
Ledoit, O., and Wolf, M. (2003), Improved Estimation of the Covariance Matrix
of Stock Returns With an Application to Portfolio Selection, Journal of
Empirical Finance, 10, 603621. [120,121,121]
Lettau, M., and Ludvigson, S. (2001), Consumption, Aggregate Wealth, and
Expected Stock Returns, Journal of Finance, 56, 815849. [111,117]
Lilien, D. (1982), Sectoral Shifts and Cyclical Unemployment, Journal of
Political Economy, 90, 777793. [110,117]
Newey, W., and McFadden, D. (1994), Large Sample Estimation and Hypoth-
esis Testing, Chapter 36 in Handbook of Econometrics (Vol. IV), eds. R. F.
Engle and D. McFadden, Amsterdam: Elsevier Science B.V., pp. 21132245.
[113]
Newey, W., and Steigerwald, D. (1997), Asymptotic Bias for Quasi-Maximum
Likelihood Estimators in Conditional Heteroskedastic Models, Economet-
rica, 65, 587599. [113]
Newey, W., and West, K. (1987), A Simple, Positive Semi-Denite, Het-
eroskedasticity and Autocorrelation Consistent Covariance Matrix, Econo-
metrica, 55, 703708. [117]
Nychka, D. (1995), Splines as Local Smoothers, The Annals of Statistics, 23,
11751197. [120]
Pastor, L., and Veronesi, P. (2003), Stock Valuation and Learning about Prof-
itability, Journal of Finance, 58, 17491789. [110,111,115]
Ross, S. (1976), The Arbitrage Theory of Capital Asset Pricing, Journal of
Economic Theory, 13, 341360. [110]
Schwert, G. (1989), Why Does Stock Market Volatility Changes over Time?,
Journal of Finance, 44, 11151153. [111]
Sharpe, W. F. (1964), Capital Asset Prices: A Theory of Market Equi-
librium under Conditions of Risk, Journal of Finance, 19, 425442.
[110]
Shephard, N. (2004), Stochastic Volatility: Selected Readings, Oxford: Oxford
University Press. [109]
Silverman, B. W. (1984), Spline Smoothing: The Equivalent Variable Kernel
Method, Annals of Statistics, 12, 898916. [120]
Syverson, C. (2004), Market Structure and Productivity: AConcrete Example,
Journal of Political Economy, 112, 11811222. [116]
Wei, S., and Zhang, C. (2006), Why Did Individual Stocks Become More
Volatile?, Journal of Business, 79, 259292. [111]
D
o
w
n
l
o
a
d
e
d

b
y

[
b
-
o
n
:

B
i
b
l
i
o
t
e
c
a

d
o

c
o
n
h
e
c
i
m
e
n
t
o

o
n
l
i
n
e

U
P
]

a
t

1
5
:
5
6

2
7

M
a
y

2
0
1
2