You are on page 1of 37

Eur. Actuar. J.

(2013) 3:321–357
DOI 10.1007/s13385-013-0079-3

ORIGINAL RESEARCH PAPER

Bivariate lower and upper orthant value-at-risk

Hélène Cossette • Mélina Mailhot •


Étienne Marceau • Mhamed Mesfioui

Received: 10 June 2013 / Revised: 7 September 2013 / Accepted: 18 September 2013 /


Published online: 31 October 2013
 DAV / DGVFM 2013

Abstract Value-at-risk (VaR) is an important risk measure widely used in actu-


arial science and quantitative risk management. Embrechts and Puccetti (J Multivar
Anal 97(2):526–547, 2006a) have introduced the multivariate lower and upper
orthant VaR. The practical applications of these risk measures is very promising,
especially in actuarial science and quantitative risk management. Our objective is to
study in details the multivariate lower and upper orthant VaR in the bivariate
setting, their properties and their applications. In particular, new characterizations of
the bivariate lower and upper orthant VaR and desirable properties are given, such
as translation invariance, positive homogeneity and comonotonic additivity. Lower
and upper confidence regions for random vectors are developed and used to provide
new results on the convexity conditions and to suggest capital allocation techniques.
We provide bounds on functions of random pairs and derive interesting relations
with existing results. We motivate the use of the bivariate lower and upper ortant
VaR for risk allocation, to represent bivariate ruin probabilities and for risk com-
parison. Practical illustrations and examples of the results are presented throughout
the article.

Keywords Multivariate value-at-risk  Risk measures  Copulas  Bounds 


Convexity

H. Cossette  É. Marceau


École d’Actuariat, Université Laval, Quebec, Canada

M. Mailhot (&)
Department of Mathematics and Statistics, Concordia University, Montreal, QC, Canada
e-mail: melina.mailhot@concordia.ca

M. Mesfioui
Département de mathématiques et informatique, Université du Québec à Trois-Rivières,
Trois-Rivières, Canada

123
322 H. Cossette et al.

1 Introduction and motivation

Under the univariate hypothesis, several tools have been studied to measure risk and
allocate capital, especially for insurance companies and other financial institutions.
Value-at-risk (VaR) is an important risk measure in actuarial science and in finance
first because of regulatory reasons, but also because it is easy to understand. Let
X be a random variable (rv) with cumulative distribution function (cdf) FX. The
VaR, at level a, 0 \ a \ 1, of X is defined by
VaRa ðXÞ ¼ inffx 2 R; FX ðxÞ  ag:
This risk measure is well understood and has been a point of interest for several
years. See e.g. McNeil et al. [21] for a review.
Some situations require that each component of a portfolio be fixed such that the
joint cdf does not to exceed a given level a. Notably, one may want to describe
relationships using inverse quantile functions, or allocate capital for each
component of a portfolio. As noted in Jouini et al. [16], investors might not be
able to aggregate their risks, due to liquidity problems and/or transaction costs
between the different security markets. In those situations, the use of a multivariate
VaR is more appropriate.
Embrechts and Puccetti [10] have introduced the multivariate lower and upper
orthant VaR associated to a loss X. Let X ¼ ðX1 ; . . .; Xk Þ be a random vector with joint
cdf FX and joint survival function (sf) FX : For a 2 ð0; 1Þ; the multivariate lower
orthant VaR at probability level a is the boundary of its a-level sets, defined by
 
VaRa ðXÞ ¼ o x 2 Rk : FX ðxÞ  a : ð1:1Þ
Analogously, the multivariate upper orthant VaR at probability level a is given by
 
VaRa ðXÞ ¼ o x 2 Rk : FX ðxÞ  1  a : ð1:2Þ
The practical applications of these risk measures seem to be very promising,
especially in finance, quantitative risk management and actuarial science. As stated
in Cherubini and Luciano [5], one could be interested in a trade-off between the
VaR of two stock indices or in comparing a-level sets. Gugan and Hassani [14]
justify the use of multivariate risk measures by linking capital requirements with
operational risk capital, using pair-copulas. Frees and Valdez [13] consider the
bivariate allocation to losses and the related expense and adjustment costs. The
example in this paper has also been used in Di Bernardino et al. [8] to illustrate the
multivariate CTE presented in this article. The latter article studies the empirical
bivariate distributions of risks, and provide an estimation method for a multivariate
Conditional Tail Expectation. The applications of multivariate risk measures can be
used in other contexts than multivariate capital allocation or risk comparison, as for
studying reinsurance premiums and building asset portfolios.
Our objective is to study the behavior of VaRa (X) and VaRa (X) in a bivariate
setting to begin with. Results could be extended to higher dimensions (k [ 2) by, for
example, successively fixing each rv. Such an extension to a higher dimension will
be addressed in future work. Note that in certain contexts, such as vine copulas,

123
Bivariate lower and upper orthant value-at-risk 323

higher dimensions result in a study of many two-dimensions. In this paper, we


consider continuous rv’s and leave the non continuous case (suitable e.g. for layered
risks or risks covered by reinsurance stop loss contracts) for further research work.
The paper is organized as follows. In Sect. 2.1 definitions and characterizations
of the bivariate lower and upper orthant VaR are given. Section 2.2 presents
interesting and desirable properties of the bivariate lower and upper orthant VaR,
such as behaviors under transformations of the bivariate set, translation invariance
and positive homogeneity. In Sect. 2.3 lower and upper confidence regions are
presented. In Sect. 2.4 the convexity conditions are studied, based on the joint and
marginal distributions. Also, special cases are presented. In Sect. 2.5 the impact of
the dependence and of marginal distributions on the bivariate lower and upper
orthant VaR is provided. Section 3 studies the bivariate lower and upper VaR for
sums of random pairs. This section also provides bounds for those sums. Finally,
Sect. 4 illustrates some results of the previous sections about the lower and upper
orthant VaR. More specifically, Sect. 4.1 presents two methods using VaRa (X) and
VaRa (X) to obtain an allocation couple, Sect. 4.3 treats the special case of two lines
of business covering the same number of risks, Sect. 4.2 links bivariate lower and
upper orthant VaR with bivariate ruin probabilities and Sect. 4.4 illustrates the
concept of confidence regions and allocation sets.

2 Characterization and resulting properties

2.1 Bivariate lower and upper orthant value-at-risk

In this section, we propose an alternative way to define the lower and upper orthant
VaR, i.e. VaRa (X) and VaRa (X), defined in Embrechts and Puccetti [10].
Let X = (X1, X2) be a random vector with joint cdf FX and sf FX : Denote Fx1 and
Fx2 the marginal cdf’s of X. For fixed x1, define the functions x2 7! Fx1 ðx2 Þ ¼
FX ðx1 ; x2 Þ and x2 7! Fx1 ðx2 Þ ¼ FX ðx1 ; x2 Þ: Let Fx1
1
(a) and Fx1
1
ðaÞ be their corre-
sponding generalized inverse functions given by
Fx1
1
ðaÞ ¼ inf ft 2 R : Fx1 ðtÞ  ag and Fx1
1
ðaÞ ¼ inf ft 2 R : Fx1 ðtÞ  ag

respectively. Note that the inequality FX ðx1 ; x2 Þ  a and FX ðx1 ; x2 Þ  1  a are
equivalent to x2  Fx1 1
ðaÞ and x2  Fx1
1
ð1  aÞ respectively. Moreover, if FX is
continuous, then we have
   
FX x1 ; Fx1
1
ðaÞ ¼ a; X x1 ; Fx1 ð1  aÞ ¼ 1  a; x1  VaRa ðX1 Þ:
F 1

Throughout the paper we adopt the notations


Fx1
1
ðaÞ ¼ VaRa;x1 ðXÞ and Fx1
1
ð1  aÞ ¼ VaRa;x1 ðXÞ:
Hereafter, we propose a convenient characterization of the boundary a-level sets
VaRa (X) and VaRa (X) given in (1.1) and (1.2) in terms of the a-level curves
x1 7! VaRa;x1 ðXÞ; x1 7! VaRa;x1 (X) and x2 7! VaRa;x2 ðXÞ; x2 7! VaRa;x2 (X), namely

123
324 H. Cossette et al.

  
VaRa ðXÞ ¼ x1 ; VaRa;x1 ðXÞ ; x1  VaRa ðX1 Þ ; ð2:1Þ
  
VaRa ðXÞ ¼ x1 ; VaRa;x1 ðXÞ ; x1  VaRa ðX1 Þ ð2:2Þ
and
  
VaRa ðXÞ ¼ VaRa;x2 ðXÞ; x2 ; x2  VaRa ðX2 Þ ; ð2:3Þ
  
VaRa ðXÞ ¼ VaRa;x2 ðXÞ; x2 ; x2  VaRa ðX2 Þ : ð2:4Þ
The latter approach defines the values obtained to represent the bivariate lower and
upper orthant VaR respectively, for X2 and X1. More precisely, (2.1) and (2.2) allow
to establish the lower and upper curves for X2, by fixing x1 and isolating x2. The
lower and upper curves for X1 are provided by (2.3) and (2.4). They are obtained by
fixing x2 and isolating x1. To simplify the presentation, we will systematically use
(2.1) and (2.2) to express our results.
Definitions (2.1) and (2.2) of the bivariate VaR allow us to derive the results
presented in this paper. We now investigate the behavior of the a-level curves
x1 7! VaRa;x1 ðXÞ and x1 7! VaRa;x1 ðXÞ: For that purpose, denote the supports of X1
and X2 by supp(X1) and supp(X2). Let lX1 and uX1 be the essential supremum and
essential infimum of X1 defined by lX1 ¼ inf fx : x 2 suppðX1 Þg and uX1 ¼
supfx : x 2 suppðX1 Þg; and define lX2 and uX2 for X2 similarly.
Proposition 2.1 Let X ¼ ðX1 ; X2 Þ be a pair of rv’s with joint cdf’s FX and
marginal cdf’s FX1 and FX2 . Then, the a-level curves
x1 7! VaRa;x1 ðXÞ and x1 7! VaRa;x1 ðXÞ

are decreasing functions. Moreover, if FX is strictly increasing, then


ð1Þ lim VaRa;x1 ðXÞ ¼ VaRa ðX2 Þ and lim VaRa;x1 ðXÞ ¼ uX2 ; ð2:5Þ
x1 !uX1 x1 !VaRa ðX1 Þ

ð2Þ lim VaRa;x1 ðXÞ ¼ VaRa ðX2 Þ and lim VaRa;x1 ðXÞ ¼ lX2 : ð2:6Þ
x1 !lX1 x1 !VaRa ðX1 Þ

Proof Since the joint cdf FX is assumed to be continuous, we have


   
FX x1 ; VaRa;x1 ðXÞ ¼ a; FX x1 ; VaRa;x1 ðXÞ ¼ 1  a: ð2:7Þ

Given that FX is an increasing function (respectively FX decreasing), then one has
necessarily from (2.7) that x1 7! VaRa;x1 ðXÞ and x1 7! VaRa;x1 ðXÞ are decreasing
functions of x1 (if not, we have a contradiction). In addition, since FX is continuous,
then
lim VaRa;x1 ðXÞ ¼ FX1
2
ðaÞ ¼ VaRa ðX2 Þ:
x1 !uX1

Also, since FX is continuous and strictly increasing, then


lim VaRa;x1 ðXÞ ¼ uX2 ;
x1 !VaRa ðX1 Þ

hence (2.5). The result (2.6) is obtained similarly. h

123
Bivariate lower and upper orthant value-at-risk 325

The interest of Proposition 2.1 is to show that


VaRa ðXÞ  ½VaRa ðX1 Þ; uX1 ½  ½VaRa ðX2 Þ; uX2 ½; ð2:8Þ
VaRa ðXÞ  lX1 ; VaRa ðX1 ÞlX2 ; VaRa ðX2 Þ: ð2:9Þ

Example 2.2 Consider X1*exponentialðk1 ¼ 0:2Þ and X2*exponentialðk2 ¼


0:25Þ; linked by a Clayton copula with h = 2. Figure 1 represents the 95 %-level
curves of VaRa ðXÞ (dashed line) and VaRa ðXÞ (solid line). Note that bivariate risks
linked by an Archimedean copula have concave marginal cdf’s. The horizontal and
vertical solid lines represent the univariate VaRa(X1) and VaRa(X2). This represents
a typical situation. Nevertheless, we will see further that one prefers the situation
where VaRa ðXÞ is convex and VaRa ðXÞ is concave. The desirable conditions are
more easily obtained with VaRa ðXÞ than with VaRa ðXÞ. h
Remark 2.3 It is also possible to verify that VaRa ðXÞ is always smaller than
VaRa ðXÞ: Using q to represent the boundary of a set, we have that
  
VaRa ðXÞ ¼ o x 2 R2 : FX ðxÞ  a : Then, the boundary sets must verify
  T 
o x 2 R2 : PðX1  x1 X2  x2 Þ  a : Analogously, we have that
  
VaRa ðXÞ ¼ o x 2 R : FX ðxÞ  1  a : Then, the boundary sets must verity
2 
  S 
o x 2 R2 : PðX1  x1 X2  x2 Þ  a :
Example 2.4 Consider X1*Beta(a1 = 5, b1 = 1) and X2*Beta(a2 = 2, b2 = 1),
linked by a Frank copula with h = -5. Figure 2 illustrates the bivariate lower and
upper orthant VaR, using convex marginal cdf’s.
One sees that a lower dependence parameter provides a lower curve for the
bivariate upper orthant VaR and an upper curve for the bivariate lower orthant VaR.
Also, we observe a concave bivariate upper orthant VaR, and slightly convex lower
orthant VaR. In this case, it is clear that the convexity, based on the marginals and
dependence structure, differs from that of Example 2.2. h

2.2 Properties of the bivariate lower and upper orthant VaR

In this section, we present analogous properties of the bivariate lower and upper
orthant VaR to those of the univariate VaR.
The following proposition states that the bivariate lower and upper orthant VaR
of a transformation of the bivariate set, through increasing functions, modifies the
curves by this same transformation. It also shows analogous results for decreasing
functions, where the bivariate lower (upper) orthant curve at level a, results in the
bivariate upper (lower) orthant at level 1 - a.
Proposition 2.5 Let X = (X1, X2) be a continuous random vector and
/ðXÞ ¼ ð/1 ðX1 Þ; /2 ðX2 ÞÞ;
where /1 and /2 are real functions defined on the supports of X1 and X2
respectively.

123
326 H. Cossette et al.

35

30

25

20

15

10

0
5 10 15 20 25 30 35

Fig. 1 Graphical representation of the bivariate lower and upper VaR. (a-curves of VaRa (X) (dashed)
and VaRa (X) (solid))

1 1
↑ θ=−5
0.9 VaRα(X2) θ=5

0.995
0.8

0.7
0.99
0.6
θ=−5
θ=5 VaRα(X1)→
0.5
0.985 ←VaRα(X1)

0.4

0.3 0.98

0.2
VaRα(X2)
0.1 0.975 ↓
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.99 0.991 0.992 0.993 0.994 0.995 0.996 0.997 0.998 0.999 1

Fig. 2 Graphical representation of the upper and lower orthant VaR with h = -5 and h = 5

1.For increasing functions /i and /j, i, j = 1, 2, i = j, we have


   
VaRa;/j ðxj Þ ð/ðXÞÞ ¼ /i VaRa;xj ðXÞ and VaRa;/j ðxj Þ ð/ðXÞÞ ¼ /i VaRa;xj ðXÞ :

2. For decreasing functions /i and /j, i, j = 1, 2, i = j, we have


 
VaRa;/j ðxj Þ ð/ðXÞÞ ¼ /i VaR1a;xj ðXÞ and
 
VaRa;/j ðxj Þ ð/ðXÞÞ ¼ /i VaR1a;xj ðXÞ :

123
Bivariate lower and upper orthant value-at-risk 327

Proof
1. Let us condition on X2 = x2 and consider increasing functions /i, i = 1, 2.
Then, one has VaRa;x2 7! Fx2 ðVaRa;x2 Þ such that
a ¼ FX ðVaRa;x2 ðXÞ; x2 Þ
¼ PðX1  VaRa;x2 ðXÞ; X2  x2 Þ
¼ Pð/1 ðX1 Þ  /1 ðVaRa;x2 ðXÞÞ; /2 ðX2 Þ  /2 ðx2 ÞÞ
¼ Pð/1 ðX1 Þ  VaRa;/2 ðx2 Þ ð/ðXÞÞ; /2 ðX2 Þ  /2 ðx2 ÞÞ:

Analogous arguments are used for a similar result with the bivariate upper orthant
VaR.
2. Let us condition on X2 = x2 and consider decreasing functions /i, i = 1, 2.
Then, one has VaRa;x2 7! F x2 ðVaRa;x2 Þ such that
1  a ¼ F X ðVaRa;x2 ðXÞ; x2 Þ;
and
1  a ¼ PðX1 [ VaR1a;x2 ðXÞ; X2 [ x2 Þ:
Because /i, i = 1, 2 are decreasing functions, one has
1  a ¼ Pð/1 ðX1 Þ  /1 ðVaR1a;x2 ðXÞÞ; /2 ðX2 Þ  /2 ðx2 ÞÞ
¼ Pð/1 ðX1 Þ  VaRa;/2 ðx2 Þ ð/ðXÞÞ; /2 ðX2 Þ  /2 ðx2 ÞÞ:

Analogous arguments are used for the result with the bivariate upper orthant
VaR. h
Corollary 2.6 ensures the translation invariance. For any additional risk to the set,
the curves will also translate for this same values c ¼ ðc1 ; c2 Þ 2 R:
Corollary 2.6 As a special case of Proposition 2.5, one has for all c ¼
ðc1 ; c2 Þ 2 R and i, j = 1, 2, i = j, then
VaRa;xj þcj ðX þ cÞ ¼ VaRa;xj ðXÞ þ ci ; VaRa;xj þcj ðX þ cÞ ¼ VaRa;xj ðXÞ þ ci :

Obtained with the application of Proposition 2.5 (item 1.), corollary 2.7 relates
the conditions for the bivariate lower and upper orthant VaR to be homogeneously
invariant. Through positive transformations of the bivariate set, the bivariate lower
and upper orthant VaR will vary with the same transformations, with level a.
Corollary 2.7 For all c ¼ ðc1 ; c2 Þ 2 Rþ  Rþ and i, j = 1, 2, i = j, then
VaRa;cj xj ðcXÞ ¼ ci VaRa;xj ðXÞ; VaRa;cj xj ðcXÞ ¼ ci VaRa;xj ðXÞ:

Corollary 2.8 results from an application of Proposition 2.5 (2.), for negative
transformations of the bivariate set.

123
328 H. Cossette et al.

Corollary 2.8 For all c ¼ ðc1 ; c2 Þ 2 R  R and i, j = 1, 2, i = j, then


VaRa;cj xj ðcXÞ ¼ ci VaR1a;xj ðXÞ; VaRa;cj xj ðcXÞ ¼ ci VaR1a;xj ðXÞ:

2.3 Lower and upper confidence regions

In this section, we want to study other properties of the bivariate lower and upper
orthant VaR and explain how to enlarge the use of the bivariate lower and upper
orthant VaR, by using confidence regions. The bivariate lower and upper confidence
regions of level a represents the sets of points covering a % of the possible values of
a bivariate set of rv’s. Those regions can represent an acceptance region as defined
in Jouini et al. [16] and Bentahar (2006), based on the bivariate VaR curves. The
bivariate lower orthant confidence region is bounded by a bivariate lower orthant
VaR, up to which a % of the sets are under that curve. The bivariate upper orthant
confidence region is bounded by the bivariate upper orthant VaR, up to which (1 -
a)% of the sets are over that curve.
 
Here and in the sequel, we denote CFX ;a ¼ ðx1 ; x2 Þ 2 R2 : FX ðx1 ; x2 Þ  a and
 
CFX ;a ¼ ðx1 ; x2 Þ 2 R2 : FX ðx1 ; x2 Þ  1  a : In this section, we discuss how one
can derive a lower and upper confidence region based on the lower and upper
orthant VaR. The objective is to find a level curve such that the probability that the
random vector X is below (respectively above) this curve equals the level a.
To this end, define the bivariate order, denoted ; by X VaRa ðXÞ if and only if
X 2 CcFX ;a ; where CcFX ;a denotes the complement of the event CFX ;a : Clearly,
X VaRa ðXÞ is equivalent to X being below the curve VaRa ðXÞ: One has
 
PðX VaRa ðXÞÞ ¼ P X 2 CcFX ;a ¼ A1 þ A2
VaRa;s ðXÞ
Z1 Z
¼aþ f ðs; tÞdtds ¼ kðaÞ;
FX1 ðaÞ 1
1

ð2:10Þ
where f(s, t) denotes the joint probability density function (pdf) of random vector
X. Figure 3 illustrates
 the regions A1 and A2 : Consequently, one has
P X VaRk1 ðaÞ ðXÞ ¼ a: In other words, the level curve VaRk1 ðaÞ ðXÞ can be
considered as a lower confidence region of the random vector X at level a.
Moreover, one observes that k1 : ½0; 1 7! ½0; 1 is an increasing function such that
k-1(a) B a. This means that the lower confidence region curve VaRk1 ðaÞ (X) is
smaller than the lower orthant a-level curve VaRa ðXÞ: Similarly, one can obtain the
upper confidence region of random vector X at level 1 - a, in terms of the upper
orthant VaR. In fact, let
be a bivariate order defined by X
VaRa ðXÞ if and only
if X is above the a-curve VaRa ðXÞ: It follows that

123
Bivariate lower and upper orthant value-at-risk 329

30
ψα(s)
A1
25
A2

20

15
t

10

5 → F−1
1
(α)

0
0 5 10 15 20 25 30 35
s

Fig. 3 Lower orthant confidence region for rv’s with positive supports

FX1 ðaÞ
Z1 Z1
 
P X
VaRa ðXÞ ¼ 1  a þ 
f ðs; tÞdtds ¼ kðaÞ:
1 VaRa;s ðXÞ

The upper confidence region of random vector X is then given by VaRk1 ðaÞ ðXÞ; that
is,
 
P X
VaRk1 ðaÞ ðXÞ ¼ 1  a:

Hence, VaRk1 ðaÞ ðXÞ (respectively VaRk1 ðaÞ ðXÞ) can be viewed as a threshold curve
such that the probability that the components of the loss X over the given time
horizon are simultaneously below (respectively above) this curve with probability a
(respectively 1 - a).
Example 2.9 Let us consider the model chosen in Cherubini and Luciano [5] for
the returns on two different indexes, S&P100 and FTSE100 (historical data
downloaded from YahooFinance) represented by a Clayton copula (h = 0.5) and
normal marginal distributions. The copula allows to separate the impact on the joint
distribution of the marginal cdf’s. We use this example in order to illustrate the
confidence region at level 1 %, that could be of interest to study the effect of moving
capital from one desk to the other, using the trade-off of VaR. The confidence
bivariate lower orthant confidence region illustrates the 1 % acceptable scenarios,
for which trading from one desk to the other does not produce an undesirable
outcome. The same exercise could be done using portfolios of assets.

123
330 H. Cossette et al.

−3
x 10

log−returns of FTSE100

−5

−10

−15
−12 −10 −8 −6 −4 −2 0
−3
log−returns of S&P100 x 10

Fig. 4 Lower orthant confidence region at level 1 %

Figure 4 represents the trade-off curves at level 1 % and area under 0.535 % level
curve. It shows that 1 % of the sets under the bivariate lower orthant VaR at level
a = 1 % can be represented by the bivariate lower orthant VaR at level 0.535 %. h

2.4 Convexity of the bivariate lower and upper orthant value-at-risk

As shown by Examples 2.2 and 2.4, the convexity of the joint cdf of X ¼ ðX1 ; X2 Þ
has an impact on the shape of the bivariate lower and upper orthant VaR. It will also
affect the allocation sets, when one wants to select a bivariate vector from the lower
or upper orthant VaR. This preoccupation is covered in Sect. 4.1 We need to study
the convexity of the curves first.
In what follows, we examine the shape of the a-level curves x1 7! VaRa;x1 ðXÞ and
x1 7! VaRa;x1 ðXÞ: The following result establishes sufficient conditions to ensure the
convexity and the concavity of these a-level curves.
Proposition 2.10 Let X ¼ ðX1 ; X2 Þ be a random vector with joint cdf FX and joint
sf FX : One has
(1) If FX is concave (respectively convex) then x1 7! VaRa;x1 ðXÞ is convex
(respectively concave).
(2) If FX is convex (respectively concave) then x1 7! VaRa;x1 ðXÞ is convex
(respectively concave).

Proof To show (1), suppose that FX is a concave function, with the confidence
region

123
Bivariate lower and upper orthant value-at-risk 331

 
CFX ;a ¼ ðx1 ; x2 Þ 2 R2 : FX ðx1 ; x2 Þ  a
and let x ¼ ðx1 ; x2 Þ 2 CFX ;a ; y ¼ ðy1 ; y2 Þ 2 CFX ;a and k 2 ½0; 1: Then, one has
FX ðkx þ ð1  kÞyÞ  kFX ðxÞ þ ð1  kÞFX ðyÞ  ka þ ð1  kÞa ¼ a:
Thus, the confidence region CFX ;a is a convex set and its boundary is a convex
function, so x1 7! VaRa;x1 ðXÞ is convex. Now, if FX is convex, then the complement
of the confidence region CFX ;a is a convex set so the boundary of CFX ;a is concave,
thus (1) holds. Similar arguments may be used to show (2). h
We propose a practical criterion that ensures the convexity of these risk measures
and we set x1* = FX11 (a) and x2* = FX1
2
(a).
Proposition 2.11 Let X ¼ ðX1 ; X2 Þ be a random vector with joint cdf FX and joint
sf FX : Denote FX1 and FX2 the marginal cdf’s. Assume that FX is twice
differentiable. Then,
o2 FX
(1) If ox2i
ðx1 ; x2 Þ  0 for all x1 C x*1 and x2 C x*2, then the a-level curve
x1 7! VaRa;x1 ðXÞ is convex.
o2 FX
(2) If ox2i
ðx1 ; x2 Þ  0 for all x1 B x*1 and x2 B x*2, then the a-level curve
x1 7! VaRa;x1 ðXÞ is concave.

Proof Using implicit differentiable calculus rules, one deduces that VaRa;x1 ðXÞ is also
twice differentiable (because F is twice differentiable and FX ðx1 ; VaRa;x1 ðXÞÞ ¼ a).
 
Now using the fact that d2 FX x1 ; VaRa;x1 ðXÞ =dx21 ¼ 0; then
d   d2 VaRa;x1 ðXÞ
 FX x1 ; VaRa;x1 ðXÞ
dx2 dx21
d2   d2   dVaRa;x1 ðXÞ
¼ 2 FX x1 ; VaRa;x1 ðXÞ þ 2 FX x1 ; VaRa;x1 ðXÞ ð2:11Þ
dx1 dx1 dx2 dx1

d2   dVaRa;x1 ðXÞ 2
þ 2 FX x1 ; VaRa;x1 ðXÞ
dx2 dx1
hence (1), because
d 2 F X ðx 1 ; x2 Þ d dVaRa;x1 ðXÞ
 0; FX ðx1 ; x2 Þ  0;  0:
dx1 dx2 dx2 dx1
The statement (2) is obtained similarly. h
Many bivariate distributions satisfy the criteria of Proposition 2.10, as for the
bivariate Eynaud–Farlie–Gumbel–Morgenstein (EFGM) bivariate exponential dis-
tribution, as presented in Balakrishnan and Lai [1].
Example 2.12 Consider the bivariate EFGM exponential distribution, with
parameters (b1 = 10, b2 = 15, h = 3). Figure 5 illustrates the curves of the
bivariate lower and upper orthant VaR at level 95 %, and for h = –0.9 and

123
332 H. Cossette et al.

100
θ=0.9
90 θ=−0.9

80

70

60

50


40 VaRα(X2)

30
←VaRα(X1)
20

10

0
0 10 20 30 40 50 60

Fig. 5 Graphical representation of the lower and upper bounds with the bivariate FGM distribution, for
h = -0.9 and h = 0.9

h = 0.9. One sees that the bivariate lower orthant VaR is convex for h [ 0 and for
h \ 0. However, the bivariate upper orthant VaR is concave only when h [ 0,
which is a desirable scenario to obtain optimized values, as we will see further in
this chapter. The bivariate upper orthant VaR is convex when h \ 0. Moreover,
Fig. 5 clearly shows that the bivariate upper orthant VaR is more affected by
changes in the dependence parameter than the bivariate lower orthant VaR. h
As shown in Example 2.12, when there is a positive dependence between the
rv’s, the bivariate EFGM exponential distribution satisfies the two conditions of
Proposition 2.11, in order to have convenient lower and upper orthant VaR curves.
Note that the bivariate EFGM copula is generated from the bivariate EFGM
exponential distribution.
It is now well recognized that copulas provide a flexible approach to model the
joint behavior of rv’s. In fact, they allow the representation of a multivariate
distribution as a function of its univariate marginal cdf’s through a linking function
called a copula. Let X ¼ ðX1 ; X2 Þ be a random vector with joint cdf FX and marginal
cdf’s FX1 and FX2 . A well known theorem in Sklar [25] ensures that there exists a
unique copula C : ½0; 12 ! ½0; 1 such that for all x1 ; x2 2 R
FX ðx1 ; x2 Þ ¼ CðFX1 ðx1 Þ; FX2 ðx2 ÞÞ: ð2:12Þ
As a consequence of Proposition 2.11, the shape of the a-level curves
x1 7! VaRa;x1 ðXÞ and x1 7! VaRa;x1 ðXÞ may be studied in terms of copulas and
marginal cdf’s as stated next.

123
Bivariate lower and upper orthant value-at-risk 333

Corollary 2.13 Let X ¼ ðX1 ; X2 Þ be a random vector with joint cdf FX and
marginal cdf’s FX1 and FX2 connected by a copula C. Suppose that the copula C and
FXi , i = 1, 2 are twice differentiable.
o2 C
(1) If ou2i
ðu1 ; u2 Þ  0 for all u1 ; u2 2 ½a; 1; i ¼ 1; 2 and FX1 (x1) and FX2 (x2) are
concave for all x1 C x*1 and x2 C x*2, then the a-level curve x1 7! VaRa;x1 ðXÞ is
convex.
2
(2) If oouC2 ðu1 ; u2 Þ  0 for all u1 ; u2 2 ½0; a; i ¼ 1; 2 and FX1 (x1) and FX2 (x2) are
i

convex for all x1 B x*1 and x2 B x*2, then the a-level curve x1 7! VaRa;x1 ðXÞ is
concave.

Proof The result is immediate from Sklar’s theorem stated in (2.12). In fact, one
has
o2 F X o2 C
ðx ;
1 2x Þ ¼ ðFX1 ðx1 Þ; FX2 ðx2 ÞÞðFX0 i ðxi ÞÞ2
ox2i oFX2 i ðxi Þ
oC
þ 2 ðFX1 ðx1 Þ; FX2 ðx2 ÞÞFX00i ðxi Þ; i ¼ 1; 2; ð2:13Þ
oFXi ðxi Þ

so (1) holds. Similar arguments ensure (2). h


Note that the assumption that F1(x1) and F2(x2) are concave for all x1 C x*1 and
x2 C x*2 is fulfilled for many important univariate distributions in actuarial science
and quantitative risk management, such as the exponential, Pareto and gamma
distributions. Moreover, several multivariate distributions, as the EFGM bivariate
exponential distribution, satisfy Corollary 2.13.
An important class of copulas are the Archimedean copulas (e.g. Nelsen [23]). A
copula C is archimedean with generator /, if for all u; v 2 ½0; 1
Cðu; vÞ ¼ /1 ð/ðuÞ þ /ðvÞÞ;
where / : ½0; 1 ! Rþ is a continuous, possibly infinite, strictly decreasing convex
function such that /(1) = 0. For this class of copulas, the convexity of the a-level
curve x1 7! VaRa;x1 ðXÞ depends only of the behavior of the marginal cdf’s as stated
below.
Proposition 2.14 Let X ¼ ðX1 ; X2 Þ be a random vector with joint cdf FX and
marginal cdf’s FX1 and FX2 connected by an archimedean copula C with generator
/. If FX1 (x1) and FX2 (x2) are concave for all x1 C x*1 and x2 C x*2, then
x1 7! VaRa;x1 ðXÞ is convex.
Proof Let La(u) be the a-level curve of the copula C, that is, C(u, La(u)) = a. One
can show that La(u) = /-1(/(a) - /(u)). Theorem 4.3.2 in Nelsen [23] states that
for any archimedean copula, the a-level curve u 7! La ðuÞ is convex. Let (F 
G)(x) = F(G(x)), be the composite function. Then, VaRa;x1 ðXÞ ¼ FX1 2
La
FX1 ðx1 Þ: If FX1 and FX2 are concave, then x1 7! VaRa;x1 ðXÞ is convex. h

123
334 H. Cossette et al.

We refer the reader to Example 2.2 for an illustration of Proposition 2.14.


We now investigate the concavity of the a-curve x1 7! VaRa;x1 ðXÞ in terms of the
generator /. Here also, we have VaRa;x1 ðXÞ ¼ FX1
2
La FX1 ðx1 Þ where La ðuÞ is the
a-level curve associated to the bivariate sf associated to the copula C, that is
 vÞ ¼ 1  u  v þ Cðu; vÞ; so Cðu;
Cðu;  La ðuÞÞ ¼ a; as mentioned in Remark 2.3.
Hereafter, we derive conditions that ensure the concavity of the upper orthant VaR.
Proposition 2.15 Let X ¼ ðX1 ; X2 Þ be a random vector with joint cdf FX and
marginal cdf’s FX1 and FX2 connected by an archimedean copula C with generator
/. Suppose that /, FX1 and FX2 are twice differentiable. If FX1 and FX2 are convex
for all x1 B x*1 and x2 B x*2, and t 7! /00 ðtÞ=ð/0 ðtÞÞ2 is increasing for t 2 ½0; a; then
x1 7! VaRa;x1 ðXÞ is concave.
Proof We use (2) in Corollary 2.13 to prove this result. We have that
o2 C
ou2i
¼ 1; 2; if and only if /00 ðwÞ=ð/0 ðwÞÞ2  /00 ðui Þ=ð/0 ðui ÞÞ2 ; i ¼
ðu1 ; u2 Þ  0; i
1; 2; where w = /-1((/(u1) ? /(u2)) and w B min(u1, u2). It follows that if
2
t 7! /00 ðtÞ=ð/0 ðtÞÞ2 is increasing, then oouC2 ðu1 ; u2 Þ  0; i ¼ 1; 2; hence the desired
i
result. h
Refer to Example 2.4 for an illustration of Proposition 2.15.
Note that it is easy to verify the monotonicity of the function t 7! /00 ðtÞ=ð/0 ðtÞÞ2
for copulas that are members of families with real parameters. For example, for
Clayton’s family with generator /(t) = (th - 1)/h, one has /00 ðtÞ=ð/0 ðtÞÞ2 ¼ ðh þ
1Þth ; so that t 7! /00 ðtÞ=ð/0 ðtÞÞ2 is increasing if and only if h C 0. Also, for Frank’s
family with generator /ðtÞ ¼ lnðeh  1Þ  lnðeth  1Þ; one has /00 ðtÞ=ð/0 ðtÞÞ2 ¼
eht ; which is increasing if and only if h C 0. For these families of copulas, h C 0
generates a positive dependence.

2.5 Impact of dependence and marginals

In this section, we study the effect of the dependence level and the marginal cdf’s on
the bivariate upper and lower orthant VaR. Explicit bounds on these risk measures
are also obtained. In the following, we introduce stochastic ordering in order to
compare bivariate lower orthant VaR’s (respectively upper orthant VaR’s). The
latter is based on the confidence regions, presented in Sect. 2.3.
Definition 2.16 Let X1 = (X1,1, X2,1) and X2 = (X1,2, X2,2) be two pairs of risks
with joint cdf’s FX1 and FX2 ; respectively. Then, VaRa ðX1 Þ is smaller than
VaRa ðX2 Þ; denoted VaRa ðX1 Þ VaRa ðX2 Þ; if CFX2 ;a  CFX1 ;a (equivalently
VaRa;x1 ðX1 Þ  VaRa;x1 ðX2 Þ for all x1). Similarly, VaRa ðX1 Þ is smaller than
VaRa ðX2 Þ; denoted VaRa ðX1 Þ VaRa ðX2 Þ; if CFX2 ;a  CFX1 ;a (or equivalently
VaRa;x1 ðX1 Þ  VaRa;x1 ðX2 Þ for all x1).

123
Bivariate lower and upper orthant value-at-risk 335

Given two random vectors X1 = (X1,1, X2,1) and X2 = (X1,2, X2,2) with joint
cdf’s FX1 and FX2 respectively, X1 is said to be more concordant than X2, denoted
X1 co X2 ; if FX1 ðx1 ; x2 Þ  FX2 ðx1 ; x2 Þ holds for all x1 ; x2 2 R:
It is easy to see that if X1 co X2 ; then CFX1 ;a  CFX2 ;a and CFX2 ;a  CFX1 ;a :
Moreover, if CX1 and CX2 denote the copulas of X1 and X2 respectively, then
X1 co X2 if and only if CX1 ðu; vÞ  CX2 ðu; vÞ for all u; v 2 ½0; 1:
We define the Fréchet class, denoted by CðFX1 ; FX2 Þ; as the set of joint cdf’s FX1
with fixed marginals FX 1 and FX 2 . We also denote by
M(x1, x2) = min(FX1 (x1), FX2 (x2)) and W(x1, x2) = max(FX1 (x1) ? FX2 (x2) - 1, 0)
the Frchet upper and lower bounds respectively. It is well known that
Wðx1 ; x2 Þ  FX ðx1 ; x2 Þ  Mðx1 ; x2 Þ;
for all F 2 CðFX1 ; FX2 Þ and x1 ; x2 2 R:
The following result follows from the definition of the concordance ordering. It
shows the impact when the dependence structures within the vectors X1 and X2 are
different, but they have the same componentwise marginal cdf’s.
Lemma 2.17 (Impact of dependence) Let X1 = (X1,1, X2,1) and X2 = (X1,2, X2,2)
be two pairs of risks with joint cdf’s FX1 2 CðFX1 ; FX2 Þ and FX2 2 CðFX1 ; FX2 Þ;
respectively. Then, we have
X1 co X2 ) VaRa ðX2 Þ VaRa ðX1 Þ for all a 2 ½0; 1; ð2:14Þ
X1 co X2 ) VaRa ðX1 Þ VaRa ðX2 Þ for all a 2 ½0; 1: ð2:15Þ

Let us also discuss the effect of the marginal cdf’s on VaRa ðXÞ and VaRa ðXÞ
when the dependence between the components of X is fixed.
Lemma 2.18 (Impact of marginals) Let X1 = (X11, X21) and X2 = (X12, X22) be
continuous random vectors with the same copula C, within X1 and X2. Also,
consider the respective joint cdf’s FX1 2 CðFX1 ; FX2 Þ and FX2 2 CðGX1 ; GX2 Þ: Then,
for fixed a 2 ð0; 1Þ; we have
VaRa ðXi;1 Þ  VaRa ðXi;2 Þ; i ¼ 1; 2 , VaRa ðX1 Þ VaRa ðX2 Þ; ð2:16Þ
VaRa ðXi;1 Þ  VaRa ðXi;2 Þ; i ¼ 1; 2 , VaRa ðX1 Þ VaRa ðX2 Þ: ð2:17Þ

Proof To verify ()) in (2.16) and (2.17), we use the fact that if VaRa(X-
 
i,1) B VaRa(Xi,2), i = 1, 2, then C FX1 ;a  C FX2 ;a and C FX1 ;a  C FX2 ;a : To show (()

in (2.16), let VaR1


x2 ;a ðXi Þ be the inverse function of x1 ! VaRx1 ;a ðXi Þ: One sees that

lim VaRx1 ;a ðX1 Þ ¼ VaRa ðX1;1 Þ; lim VaRx1 ;a ðX2 Þ ¼ VaRa ðX1;2 Þ; ð2:18Þ
x1 !1 x1 !1

lim VaR1
x2 ;a ðX1 Þ ¼ VaRa ðX2;1 Þ; lim VaR1
x2 ;a ðX2 Þ ¼ VaRa ðX2;2 Þ: ð2:19Þ
x2 !1 x2 !1

Now, VaRa ðX1 Þ VaRa ðX2 Þ implies VaRx1 ;a ðX1 Þ  VaRx1 ;a ðX2 Þ and
VaRx2 ;a ðX1 Þ  VaR1
1
x2 ;a ðX2 Þ: Thus, from (2.18) and (2.19), we have

123
336 H. Cossette et al.

VaRa(Xi1) B VaRa(Xi2), i = 1, 2, hence the result. The implication (() in (2.17)


can be obtained similarly. h
Lemma 2.18 considers a fixed dependence structure C between the components
of the random sets, with fixed joint cdf FXi, i = 1, 2 within each random set. Based
on the latter assumption, we established the order of the bivariate lower and upper
orthant VaR of X1 and X2, based on the univariate VaR’s of the individual
components of each random set, vice versa.
Let X = (X1, X2) be a random vector with joint cdf FX and marginal cdf’s FX1
and FX2 . As a consequence of Lemma 2.17, one obtains
VaRa ðXM Þ VaRa ðXÞ VaRa ðXW Þ; VaRa ðXW Þ VaRa ðXÞ VaRa ðXM Þ:
ð2:20Þ
An interesting property of positive dependence is the concept of positive
quadrant dependence (PQD) introduced by Lehmann [18]. The random vector
X = (X1, X2) is said to be positively quadrant dependent if and only if
FX ðx1 ; x2 Þ  Pðx1 ; x2 Þ ¼ FX1 ðx1 ÞFX2 ðx2 Þ for all x1 ; x2 2 R: In such a situation and
using Lemma 2.18, inequalities in (2.20) become
VaRa ðXM Þ VaRa ðXÞ VaRa ðXP Þ; VaRa ðXP Þ VaRa ðXÞ VaRa ðXM Þ;
where XP is the random vector with the same marginal cdf’s than X, but with
independent components.

3 Lower and upper orthant value-at-risk for sums of random pairs

In this section, we motivate the use of bivariate VaR’s to obtain accurate values for
risk allocation and comparison, in the case of sum of random pairs. Models should
try to capture important characteristics such as the marginal cdf’s of homogeneous
classes and the appropriate dependence structure between classes. Using bivariate
VaR’s allows to consider each homogeneous structure of a dependent set during the
modeling process. We initiate this section by setting the framework.
Let X1 ¼ ðX1;1 ; X1;2 Þ; . . .; Xn ¼ ðXn;1 ; Xn;2 Þ be a sequence of n random pairs with
distributions FX1 ; . . .; FXn and marginal cdf’s FX1;1 ; . . .; FXn;1 and FX1;2 ; . . .; FXn;2 :
Denote S1 ¼ X1;1 þ    þ Xn;1 and S2 ¼ X1;2 þ    þ Xn;2 : In this section, we
examine the lower and upper orthant VaR of the random vector. We have S ¼
X1 þ    þ Xn ; where





S1 X11 ... Xn1
S¼ ¼ þ þ ;
S2 X12 ... Xn2
and where X1 ; . . .; Xn are linked by a copula C. The computation of the joint cdf of
S is not obvious even for specific dependence structures assumed for the random
vectors X1 ; . . .; Xn : Therefore, it is not easy to evaluate the univariate VaR for sums
of random variables. The problem is even more challenging when considering
bivariate sums of random pairs.

123
Bivariate lower and upper orthant value-at-risk 337

In Sect. 3.1 we study the sum of random vectors with comonotonic components. In
Sect. 3.2, we develop bounds on the bivariate lower and upper orthant VaR for pairs
representing random sums, in terms of the univariate VaR. We also establish the
relation with existing stochastic bounds. In Sect. 3.3, we provide bounds for the sum
of a bivariate set, where each component represents aggregated homogeneous risks.

3.1 Sum of random vectors with comonotonic components

The objective of this subsection is to provide the bivariate version of


Pn Pn
VaRa i¼1 Xi ¼ i¼1 VaRa ðXi Þ; when X1 ; . . .; Xn are comonotonic, that is if Xi ¼
1
FXi ðUÞ; i ¼ 1; . . .; n; where U is uniformly distributed on the interval [0, 1].
Hereafter, we have an analogous result in a bivariate setting.
Proposition 3.1 Let X1 ¼ ðX1;1 ; X1;2 Þ; . . .; X n ¼ ðXn;1 ; Xn;2 Þ be a sequence of
random pairs with distributions FX1 ; . . .; FXn and marginal cdf’s FX1;1 ; . . .; FXn;1 and
FX1;2 ; . . .; FXn;2 : Suppose that X1;1 ; . . .; Xn;1 and X1;2 ; . . .; Xn;2 are comonotonic
respectively, meaning that there exists uniform rv’s U1 and U2 such that
Xi;1 ¼ FX1i;1 ðU1 Þand Xi;2 ¼ FX1 i;2
ðU2 Þ; i ¼ 1; . . .; n: Suppose that U1 and U2 are
connected with a copula C. Then,
X n
VaRa;sj ðSÞ ¼ VaRa;xk;j ðXk Þ; sj  VaRa ðSj Þ; ð3:1Þ
k¼1

X
n
VaRa;sj ðSÞ ¼ VaRa;xk;j ðXk Þ; sj  VaRa ðSj Þ; ð3:2Þ
k¼1
P n Pn 1
Pn
where k=1 xk,j = sj = k=1Fxk;j  FSj (sj) and FS1
FX1 j
k;j
(u) =
(u), j = 1, 2.
k=1
P n
Proof Using the fact that S ¼ ðS1 ; S2 Þ ¼ ðFS1 ðU1 Þ; FS1 ðU2 ÞÞ and k=1
Pn 1 1 2

xk,j = sj = k=1FXk;j  FSj (sj), then from Proposition 2.5, we have for i, j = 1, 2
and i = j,
VaRa;sj ðSÞ ¼ VaRa;sj ðFS11
ðU1 Þ; FS12
ðU2 ÞÞ
 
1
¼ FSi VaRa;FSj ðsj Þ ðU1 ; U2 Þ
Xn  
¼ FX1k;i VaRa;FSj ðsj Þ ðU1 ; U2 Þ
k¼1
X
n

¼ FX1k;i 1 1
VaRa;FX1 ðFSj ðsj ÞÞ ðFXk;1 ðU1 Þ; FXk;2 ðU2 ÞÞ
k;j
k¼1
Xn
¼ VaRa;xk;j ðFX1k;1 ðU1 Þ; FX1k;2 ðU2 ÞÞ
k¼1
Xn
¼ VaRa;xk;j ðXk Þ:
k¼1

123
338 H. Cossette et al.

Hence, one obtain (3.1). Similar arguments lead to (3.2). h


This interesting proposition shows that the relation for the univariate VaR of a
sum of comonotonic rv’s also holds for the bivariate lower and upper orthant VaR of
a sum of comonotonic random couples.

3.2 Bounds on the bivariate lower and upper orthant value-at-risk for sums
of random pairs

Now, let us examine bounds on VaRa;si ðSÞ and VaRa;si ðSÞ; i ¼ 1; 2: We link
univariate and multivariate results using the bivariate lower and upper orthant VaR
and stochastic bounds for FSi , i = 1, 2. It allows to take into account homogeneous
groups of risks, part of a global dependent set. One can easily derive the following
bounds on the a-level curves si 7! VaRa;si ðSÞ and si 7! VaRa;si ðSÞ in terms of
VaRa(Si) and FSi :
VaRa ðSj Þ  VaRa;si ðSÞ  VaRaFSi ðsi Þþ1 ðSj Þ; si  VaRa ðSi Þ; ð3:3Þ

VaRaFSi ðsi Þ ðSj Þ  VaRa;si ðSÞ  VaRa ðSj Þ; si  VaRa ðSi Þ; ð3:4Þ

for i, j = 1, 2, i = j.
It is not easy to evaluate explicitly VaRa(Si) and FSi , i = 1, 2. However, several
authors have examined explicit formulas and the estimation of these quantities by
deriving stochastic bounds on the distribution of Si, i = 1, 2. Makarov [20] and
independently Rschendorf (1982) obtained stochastic bounds on FSi , i = 1, 2.
Williamson and Downs [26] also studied stochastic bounds and extended previous
results, using the duality principle. Denuit et al. [7] applied the stochastic bounds in
actuarial science, applying their results to insurance problems. Embrechts and
Puccetti [11] uses stochastic bounds on FSi , i = 1, 2 to improve the results obtained
in Embrechts et al. [9]. To recall these results, let Ci be the copula associated to the
rv’s X1;i ; . . .; Xn;i ; i ¼ 1; 2: Denote by Cdi the dual of Ci, i = 1, 2 defined by
!
[
n
Cid ðu1;i ; . . .; un;i Þ ¼ P Uj;i  uj;i ; i ¼ 1; 2;
j¼1

where ðU1;i ; . . .; Un;i Þ denotes a random vector with distribution Ci, i = 1, 2. It will
be supposed, however, that partial information is available about Ci, namely that
there are copulas Ci,L and Ci,U such that Ci C Ci,L and Cdi B Cdi,U, i = 1, 2.
Any multivariate distribution function can be represented in a way that
emphasizes the separate roles of the marginal cdf’s and the dependence structure.
For all s 2 R; Fmin;Si ðsÞ  FSi ðsÞ  Fmax;Si ðsÞ such that
 
Fmin;Si ðsÞ ¼ sup Ci;L F1;i ðu1 Þ; . . .; Fn;i ðun Þ ; i ¼ 1; 2; ð3:5Þ
u1 þþun ¼s
d
 
Fmax;Si ðsÞ ¼ inf Ci;U F1;i ðu1 Þ; . . .; Fn;i ðun Þ ; i ¼ 1; 2: ð3:6Þ
u1 þþun ¼s

Williamson and Downs [26] presented bounds for the VaR of the sum of two risks

123
Bivariate lower and upper orthant value-at-risk 339

using the duality principle. The n-dimensional formulation of this result is stated
formally by
X
n
1
VaRmin;a ðSi Þ ¼ sup Fj;i ðuj Þ; i ¼ 1; 2; ð3:7Þ
d ðu ;...;u Þ¼a
Ci;U 1 n j¼1

X
n
1
VaRmax;a ðSi Þ ¼ inf Fj;i ðuj Þ; i ¼ 1; 2: ð3:8Þ
Ci;L ðu1 ;...;un Þ¼a
j¼1

This is in fact a special case of Theorem 3.1 of Embrechts et al. [9], where the VaR
of a function wðx1 ; . . .; xn Þ of n-dependent risks was treated, applying the duality
principle of Frank and Schweizer [12].
In practical situations, the dependence structure of ðX1i ; . . .; Xni Þ; i ¼ 1; 2 is often
unknown. However, for any copula, the inequalities Ci C W and Cid  W ~ d hold,
where
Wðu1 ; . . .; un Þ ¼ minðu1 þ    þ un  1; 0Þ and
W~ d ðu1 ; . . .; un Þ ¼ minð1; u1 þ    þ un Þ:

~ d and W respec-
Practical bounds can be obtained by replacing Cdi,U and Ci,L by W
tively in (3.7) and (3.8).
Proposition 3.2 We obtain bounds for the lower and upper orthant VaR, using
bounds on the univariate VaR, that is
1
VaRmin;a ðS2 Þ  VaRa;s1 ðSÞ  VaRmax;aFmin;S1 ðs1 Þþ1 ðS2 Þ; s1  Fmin;S 1
ðaÞ; ð3:9Þ

and
1
VaRmin;aFmax;S1 ðs1 Þ ðS2 Þ  VaRa;s1 ðSÞ  VaRmax;a ðS2 Þ; s1  Fmax;S1
ðaÞ: ð3:10Þ

Proof Since a 7! VaRa ðS2 Þ is increasing, then


VaRaFS1 ðs1 Þþ1 ðS2 Þ  VaRaFmin;S1 ðs1 Þþ1 ðS2 Þ  VaRmax;aFmin;S1 ðs1 Þþ1 ðS2 Þ

Moreover, one has


VaRmin;a ðS2 Þ  VaRa ðS2 Þ:
These informations together with (3.3) imply (3.9). The inequalities (3.10) may be
obtained analogously. h

3.3 Bounds on the sum of aggregated risks

In this subsection, we consider a univariate framework. We establish bounds on the


sum of two dependent classes of homogeneous aggregated risks. This setting differs
from Sects. 3.1 and 3.2, since we obtain one-dimensional results that are related to
the sum of all risks. Moreover, we can now deal with different lengths of aggregated
risks. Our approach provides a simple way of bounding the sum of hypothetically
homogeneous classes of risks, based on the bivariate lower and upper orthant VaR.

123
340 H. Cossette et al.

Also, we provide results when the dependence structures between and within the
classes of risks are known or not.
Consider a portfolio divided into two classes comprising n1 and n2 contracts, and
let Xi,j represent the risk associated to the ith contract in the jth class, j = 1, 2. In
many situations, it is convenient to model separately the distribution of each random
vector X1 ¼ ðX1;1 ; . . .; Xn1 ;1 Þ and X2 ¼ ðX1;2 ; . . .; Xn2 ;2 Þ instead of the distribution of
the random vector X ¼ ðX1;1 ; . . .; Xn1 ;1 ; X1;2 ; . . .; Xn2 ;2 Þ: This is because the classes
are often homogeneous, and it can be easier to identify the structure of each vectors
X1 and X2 instead of the structure of the random vector X. In such a case, the lower
and upper orthant VaR can be used to derive bounds on VaR(S1 ? S2), for example,
instead of modeling the aggregate distribution of all the risks.
Corollary 3.3 When the structure of dependence of (S1, S2) is known, one obtains
( )
max s1 þ VaRa;s1 ðSÞg  VaRa ðS1 þ S2 Þ  min fs1 þ VaRa;s1 ðSÞ :
s1 \FS1 ðaÞ s1 [ FS1 ðaÞ
1 1

ð3:11Þ
When the structure of the dependence of the random vector (S1, S2) is unknown, one
can use Proposition 3.2 to derive bounds on VaRa(S1 ? S2) that are expressed in
terms of the bounds on the cdf’s of S1 and S2 as shown next:
Amin max
a  VaRa ðS1 þ S2 Þ  Aa ;

where
Amin
a ¼ max fs1 þ FS1
2
ða  FS1 ðs1 ÞÞg and
s1 \FS1 ðaÞ
1

Amax
a ¼ min fs1 þ FS1
2
ða  FS1 ðs1 Þ þ 1Þg:
s1 [ FS1 ðaÞ
1

Note that if FSi and FS1 ,


i = 1, 2 are not available, then one can substitute them by
i
their bounds Fmin;S1 ; Fmax;S1 , and Fmin;S1 ; Fmax;S1 , respectively. Hence, one obtains
Dmin max
a  VaRa ðS1 þ S2 Þ  Da ;

where
n o
1
Dmin
a ¼ max s1 þ Fmin;S2
ða  Fmax;S1 ðs1 ÞÞ
1
s1 \Fmin;S ðaÞ
1

and
n o
1
Dmax
a ¼ min s1 þ Fmax;S 2
ða  F min;S 1
ðs1 Þ þ 1Þ :
1
s1 [ Fmin;S ðaÞ
1

We want to highlight the fact that using our approach to provide bounds on the
lower and upper orthant VaR gives the opportunity to consider the dependence
within different sectors X1 and X2, and also to consider the model that represents the
dependence between sectors. Traditional methods consider a different dependence

123
Bivariate lower and upper orthant value-at-risk 341

structure, that is possibly hard to fit, because they consider heterogeneous variables,
which is not the case in Eqs. (3.7) and (3.8). This method leaves aside the depen-
dence within and between X1 and X2.
For illustration, let us consider the case where X1;i ; . . .; Xn;i are identically
distributed with common cdf FXi ; i ¼ 1; 2: Suppose that there exists x i 2 R such that
the density function fXi ðxÞ ¼ dFXi ðxÞ=dx is non-increasing for all x  x i ; i ¼ 1; 2:
This assumption is fulfilled for many important models in actuarial science and
quantitative risk management like exponential, Pareto and gamma models. Suppose
also that X1;i ; . . .; Xn;i ; i ¼ 1; 2 are positively lower orthant dependent (PLOD).
Then, from Remark 3.2 in Mesfioui and Quessy [22] we have
 
VaRmax;a ðS2 Þ ¼ nFX12 a1=n ; Fmin;S1 ðs1 Þ ¼ ½FX1 ðs1 =nÞn : ð3:12Þ

Combining (3.9) and (3.12), we get the next explicit upper bound of VaRa;s1 ðSÞ
h i
1=n
VaRa;s1 ðSÞ  nFX12 ða  ½FX1 ðs1 =nÞn þ1Þ ; s1  nFX11 ða1=n Þ: ð3:13Þ

Example 3.4 In order to appreciate the influence of the dimension n on the upper
bound given in (3.13), consider FX1  Expðk1 Þ; that is, FX1 ðxÞ ¼ 1  ek1 x ; x [ 0
and FX2 * Pareto(a), namely, FX2 ðxÞ ¼ 1  xa ; x [ 1 and a [ 0. Figure 6 provides
the a-curves of this upper bound, with a ¼ 0:95; k ¼ 0:2 and a = 1.5, for n = 2,
n = 3 and n = 4, respectively, and n1 = n2 = n. We remark that increasing the
dimension n increases the proposed upper bound with respect to the order : h

400
n=2
n=3
350
n=4

300
Lower α−curve values

250

200

150

100

50

0
0 10 20 30 40 50 60 70 80 90 100
s1

Fig. 6 Graphical representation of the upper bounds of the lower orthant VaR Graphical representation
of the upper bounds of VaRa;s1 (S) for n = 2, 3, 4

123
342 H. Cossette et al.

4 Applications

4.1 Bivariate value-at-risk and allocation

In this section, we suggest methods based on the bivariate lower and upper orthant
VaR to obtain optimal capital allocation sets. Two optimization criteria are
developed to select a bivariate set of values from these curves. The objective is to
allocate a value to each homogeneous risk, that could be used for comparison or for
capital requirements of a company with several business lines. These criteria are
developed mainly to fulfil practical needs. The bivariate lower and upper orthant
VaR curves are useful for risk comparison, but in several situations, companies or
regulators need to allocate a single amount to each business line or risk of a
portfolio. Therefore, a set has to be selected from the bivariate VaR. As shown in
Sect. 2.4, we must have a convex lower orthant VaR and a concave upper orthant
VaR, and when this condition is respected, one can obtain a bivariate set of values
from the curves, based on two different approaches. Using the bivariate lower
orthant VaR, the allocation couple has to be such that the probability that X1 is
smaller than x1 and X2 is smaller than x2 equals a: To find an allocation couple from
the curve ðx1 ; VaRa;x1 ðXÞÞ; we propose two different criteria of interest in finance,
actuarial science and quantitative risk management. Analogous results can be
obtained from the curve ðx1 ; VaRa;x1 ðXÞÞ:

4.1.1 Orthogonal projection

We start by the Orthogonal projection allocation, which consists in finding the


closest point from ðx 1 ; VaRa;x 1 ðXÞÞ to the couple (VaRa(X1), VaRa(X2)), by solving
the following minimization problem,
n  2 o
min ðVaRa ðX1 Þ  x1 Þ2 þ VaRa ðX2 Þ  VaRa;x1 ðXÞ :
x1 [ F11 ðaÞ

To find the solution, we have to solve


dVaRa;x1 ðXÞ
2ðx1  VaRa ðX1 ÞÞ þ 2ðVaRa;x1 ðXÞ  VaRa ðX2 ÞÞ ¼ 0; ð4:1Þ
dx1
and verify that the second derivative

  d2 VaRa;x1 ðXÞ dVaRa;x1 ðXÞ 2
2 þ 2 VaRa;x1 ðXÞ  VaRa ðX2 Þ þ2 ð4:2Þ
dx21 dx1
is positive, in order to get a minimum. The convexity of VaRa;x1 ðXÞ plays a central
role to obtain the optimal solution. In fact, we see that the last expression in (4.2) is
positive if VaRa;x1 ðXÞ is convex. In this situation, Eq. (4.1) will provide the
orthogonal projection from the couple ðVaRa ðX1 Þ; VaRa ðX2 ÞÞ to ðx 1 ; VaRa;x 1 ðXÞÞ:
An intuitive interpretation of this minimization is to calculate the simultaneous set
for which each rv reaches the closest values to the one that would be obtained on a

123
Bivariate lower and upper orthant value-at-risk 343

stand-alone basis, with the strongest dependence level and the possibility of risk
mitigation. This interpretation allows the user to quantify the impact of protecting
each homogeneous risk, without the possibility of aggregation.

4.1.2 Proportional allocation

The second approach is the Proportional allocation. The idea of this method is to
preserve the same ratio of the univariate VaR, that is to consider ðx 1 ; VaRa;x 1 ðXÞÞ
solution of

2
VaRa ðX1 Þ
min x1  VaRa;x1 ðXÞ :
x1 [ F11 ðaÞ VaRa ðX2 Þ

Again, if VaRa;x1 ðXÞ is convex, the minimum is obtained by solving the equation
d
VaRa;x1 ðXÞ ¼ VaRa ðX2 Þ=VaRa ðX1 Þ:
dx1
Explicit forms for inverse bivariate cdf’s do not exist for most cases. Copulas are
more tractable for that purpose, as shown in an example in the last section. The
intuitive interpretation of this minimization is to calculate the set that allocates the
same proportion to each risk as if they were comonotonic and could be mitigated.
The latter framework represents the strongest dependence level between risks that
are aggregated, which is often considered for capital requirement purposes. If no
closed-form expression exists, bounds can be found and optimization methods can
be used, as shown in the next example.
Example 4.1 Consider the random couple (X1, X2), following a bivariate mixture
of Erlang distributions with the same scale parameter h. The joint pdf is
1 X
X 1 Y
2
fX1 ;X2 ðx1 ; x2 j sm ; hÞ ¼ sm1 ;m2 hðxj ; mj ; hÞ;
m1 ¼1 m2 ¼1 j¼1

with



s1;1 s1;2 s1;3 0:2 0:1 0
sm ¼ ¼ ;
s2;1 s2;2 s2;3 0:4 0 0:3
and sm1 m2 = 0 for m1 = 3, 4, … and m2 = 4, 5, …. Also, hð; a; bÞ represents the
pdf of a gamma distribution with shape and scale parameters a and b respectively.
We obtain the following results for the allocation couples resulting from the two
allocation criteria previously presented.
Since VaRa(X1) [ VaRa(X2), the allocation to X1 is always higher in the
Proportional allocation couples. The Orthogonal projection allocation provides the
closest couple from (VaRa(X1), VaRa(X2)), resulting in a smaller total than with the
Proportional allocation, but not preserving the proportion of each risk on the
aggregate risks. In Table 1, we see that the sum of the components of the orthogonal

123
344 H. Cossette et al.

Table 1 Couples resulting from the Orthogonal projection and Proportional allocation criteria
a Orthogonal projection Total Proportional Total VaRa (X1 ? X2)

0.95 (5.2762, 5.7244) 11.0006 (5.2252, 5.7788) 11.0040 9.2699


0.99 (7.1314, 7.8442) 14.9756 (7.0794, 7.8994) 14.9788 13.3138
0.995 (7.9071, 8.7167) 16.6238 (7.8581, 8.7687) 16.6268 14.9863

projection and proportional allocation couples are higher than the VaR of the sum of
the components. This is because each component is always protected to the level
a, without considering the value of the remaining rv. h
Note that x 1 þ VaRa;x 1 ðXÞ  VaRa ðX1 Þ þ VaRa ðX2 Þ and x 1 þ VaRa;x 1 ðXÞ 
VaRa ðX1 Þ þ VaRa ðX2 Þ: This is because the protection level a is dedicated to both
risk, without embedding any possibility of risk mitigation. The minimal value of
x 1 þ VaRa;x 1 ðXÞ and x 1 þ VaRa;x 1 ðXÞ are obtained in the situation where X1 and X2
are comonotonic. Then, as the dependence gets less positive between the rv’s, the
bivariate lower orthant curve would be higher than in the comonotic scenario, and it
means that if X1 takes a smaller value, X2 should take higher values with a higher
probability, and a higher amount would be necessary to cover both risks at the same
level a.

4.2 Bivariate value-at-risk and Ruin probabilites for a portfolio of bivariate risks

We consider a portfolio with two lines of business. The aggregate claim amounts for
the next period (e.g. a month, three months or a year) for the line i is defined by the
rv Si and the corresponding premium income is denoted by pi, i = 1, 2. We assume
that pi ¼ ð1 þ gi ÞE½Si ; i ¼ 1; 2: The initial reserve allocated to the line i is denoted
by ui, i = 1, 2. Inspired from Chan et al. [4] and Cai and Li [2, 3], we define the
two following ruin probabilities of the next periods
!
[2
wor ðu1 ; u2 Þ ¼ Pr f S i  pi [ u i g
i¼1

and
!
\
2
wand ðu1 ; u2 Þ ¼ Pr f S i  pi [ u i g :
i¼1

We can relate wor ðu1 ; u2 Þ and wand ðu1 ; u2 Þ to VaRa ðSÞ and VaRa ðSÞ; where
S ¼ðS1 ; S2 Þ: We fix a given value a 2 ð0; 1Þ and assume that the premium rates p1
and p2 are fixed such that pi \VaRa ðSi Þ; i ¼ 1; 2: Using the representations
 in (2.1)
ðor Þ ðor Þ
and (2.3) we can express for the given value a, the set of couples u1 ; u2 such
that wor ðu1 ; u2 Þ ¼ 1  a coincides with the curve of VaRa ðS1  p1 ; S2  p2 Þ: It
means that either S1 or S2 stay smaller than their respective (1 - a)% largest value,
without considering the other rv’s value, but consideringthat its valuemight affect
ðandÞ ðand Þ
the rv of interest. Similarly, the set of couples u1 ; u2 such that

123
Bivariate lower and upper orthant value-at-risk 345

wand ðu1 ; u2 Þ ¼ a coincides with the curve of VaRa ðS1  p1 ; S2  p2 Þ: This means
that both S1 and S2 do not exceed the level a. For example, at a level of 99 %, if S1
reaches its 99 % greatest value but S2 is smaller than its 99 % greatest value, wor is
respected, but not wand. Since there are more scenarios
 where
 either S1 or S2 can
ðor Þ ðor Þ
reach their level a, using wand, the set of couples u1 ; u2 will always provide
 
ðandÞ ðand Þ
higher values than u1 ; u2 ; as mentioned in Remark 2.3.
In Sect. 2.5, we have demonstrated the impact of the dependence and the
marginal cdf’s on the bivariate  VaR’s.
 This is therefore
 also true for the values
ðor Þ ðor Þ ðand Þ ðandÞ
within the sets of u1 ; u2 and u1 ; u2 :
We illustrate this situation in the following example, using a bivariate compound
Poisson model.
We assume ðS1 ; S2 Þ to be a vector of rv’s following a bivariate compound
distribution where
(P
Mi
ji ¼0 Bi;ji ; Mi [ 0
Si ¼ ; ð4:3Þ
0; Mi ¼ 0

where the joint probability mass function (pmf) of ðM1 ; M2 Þ is given by


fM1 ;M2 ðm1 ; m2 Þ ¼ PrðM1 ¼ m1 ; M2 ¼ m2 Þ ¼ qm1 ;m2 ;
 
for m1 ; m2 2 N: Also, for each i; Bi;1 ; Bi;2 ; . . .; form a sequence of independent and
identically distributed
  rv’s, and Bi,1, Bi,2, …* Bi, i = 1, 2. We have that
B1;1 ; B1;2 ; . . .; ; B2;1 ; B2;2 ; . . .; are independent between themselves and inde-
pendent of ðM1 ; M2 Þ: The rv’s B1, B2 are assumed to be continuous and strictly
positive. We assume that ðM1 ; M2 Þ follows a multivariate Poisson distribution which
is defined with a common shock as explained in e.g. Johnson et al. [15], Lindskog
and McNeil [19]. We briefly recall the definition of this distribution. Let J0, J1, J2
be independent rv’s with J0  Poisðc0 Þ and Ji  Poisðci ¼ ki  c0 Þ with
0  c0  minðk1 ; k2 Þ: The components of the random vector ðM1 ; M2 Þ are defined by
M1 ¼ J1 þ J0 ;
M2 ¼ J2 þ J0 :
It implies that Mi  Poisðki Þ; i ¼ 1; 2: Also, the joint pmf and the joint probability
generating function of ðM1 ; M2 Þ are respectively given by
ðm1 ;m2 Þ
minX
c0j Y
2
ðki  c0 Þðmi jÞ
qm1 ;m2 ¼ ec0 eðki c0 Þ ;
j¼0
j! i¼1 ðmi  jÞ!
ð4:4Þ
1 M2 Y
2
c0 ðs1 s2 1Þ
PM1 ;M2 ðs1 ; s2 Þ ¼ E sM s
1 2 ¼ e eðki c0 Þðsi 1Þ ;
i¼1

where c0 corresponds to the dependence parameter. When c0 = 0, it means that the


components of ðM1 ; M2 Þ are independent.

123
346 H. Cossette et al.

Since ðM1 ; M2 Þ follows a bivariate Poisson distribution, it implies that ðS1 ; S2 Þ


has a bivariate compound Poisson distribution. We also assume that
Bi  gammaðsi ; bi Þ with E½Bi  ¼ bsi ; i ¼ 1; 2: Then the joint cdf of ðS1 ; S2 Þ is given
i
by
1 X
X 1
FS1 ;S2 ðx1 ; x2 Þ ¼ qm1 ;m2 H ðx1 ; s1 ; b1 ÞH ðx2 ; s2 ; b2 Þ:
m1 ¼0 m2 ¼0

Using the previous model, we illustrate the bivariate VaR’s in a context of ruin
probabilities for a portfolio of bivariate risks.
Example 4.2 Let us fix k1 = 2, k2 = 3, s1 = s2 = 1, b1 = b2 = 0.6, g1 = g2 =
25 %. It implies that p1 ¼ p2 ¼ 1:25  E½Si : We obtain the following values
Tables 2 and 3 both illustrate Lemma 2.17. The allocation values obtained from
the bivariate lower and upper orthant VaR are always smaller for S1 than for S2,
because the marginal cdf of S2 that is always smaller than the marginal cdf of S1.
Note that this is the same result as for the univariate VaR. Also, one sees that
the sum of the components of the allocation couples is always smaller from
the orthogonal projection than for the proportional allocation. Moreover, as the
dependence parameter increases, the bivariate lower orthant VaR decreases and the
bivariate upper orthant increases (Tables 2, 3).

 
ðorÞ ðorÞ
Table 2 Couples u1 ; u2 resulting from the Orthogonal projection and Proportional allocation
criteria for wor ðu1 ; u2 Þ ¼ 1 %:
c Orthogonal Total Proportional Total
projection

0 (18.3767, 23.9027) 42.2794 (18.7214, 23.5904) 42.3118


1 (18.3547, 23.8858) 42.2405 (18.7045, 23.5693) 42.2738
2 (18.3221, 23.8442) 42.1663 (18.6723, 23.5293) 42.2016

 
ðandÞ ðandÞ
Table 3 Couples u1 ; u2 resulting from the Orthogonal projection and Proportional allocation
criteria for wand ðu1 ; u2 Þ ¼ 1 %:
c Orthogonal Total Proportional Total
projection

0 (9.2946, 15.7539) 25.0785 (10.2339, 14.4112) 24.6451


1 (10.4030, 16.3737) 26.7767 (11.1460, 15.5437) 26.6897
2 (11.3596, 17.1804) 28.5400 (11.9447, 16.5355) 28.4802

123
Bivariate lower and upper orthant value-at-risk 347

Table 2 having higher values than Table 3 show that to make sure that both risks
do not exceed their 99 % respective worst possible values, without restriction on the
values of the remaining risk, more has to be kept aside than to make sure that the
worst 1 % is not reached by any of the two risks. h

4.3 Two lines of business with dependent risks

Now, we consider an insurance company with two lines of business (i = 1, 2). The
two vectors of rv’s (risks) associated to the lines 1 and 2 are
0 1 0 1
X1;1 X2;1
BX C BX C
B 1;2 C B 2;2 C
X1 ¼ B C and X2 ¼ B C;
@ ... A @ ... A
X1;n X2;n
Line1 Line2

where Xi,j corresponds to the cost associated to the policy j in line


i, j = 1, 2, …, n and i = 1, 2. Such a framework is suitable for insureds in non-life
insurance companies exposed to the same pair of risks (e.g. net costs and loss
adjustment expenses) or same pair of coverages (e.g. car and property insurance).
Another context is in life insurance where insureds have invested in the same two
financial products.
We assume that the components within Xi are dependent and that X1 and X2 are
also dependent. P
Define the random pair S ¼ ðS1 ; S2 Þ where Si = nj=1Xi,j corresponds to the
aggregate claim amount for the whole line i, i = 1, 2. Also, Y = S1 ? S2 is the
aggregate claim amount for the whole portfolio.
In order to examine this problem, we use a risk model based on multivariate
mixed Erlang. For more information on the multivariate mixed Erlang, see Lee and
Lin [17].
Let H be a positive mixing rv with cdf FH and mgf MH : Given H ¼ h; we
assume
 
FX1;1 ;...;X1;n ;X2;1 ;...;X2;n jH¼h x1;1 ; . . .; x1;n ; x2;1 ; . . .; x2;n
X m X m
ðhÞ
Y
n     
¼ p1;2 ðl1 ; l2 Þ H x1;j ; l1 ; b1 H x2;j ; l2 ; b2 ;
l1 ¼1 l2 ¼1 j¼1
 
MX1;1 ;...;X1;n ;X2;1 ;...;X2;n jH¼h t1;1 ; . . .; t1;n ; t2;1 ; . . .; t2;n
X m X m Yn
l1
l2 !
ðhÞ b1 b2
¼ p1;2 ðl1 ; l2 Þ :
l ¼1 l ¼1 j¼1
b1  t1;j b2  t2;j
1 2

It implies that

123
348 H. Cossette et al.

  Xm X
m
ðhÞ    
FX1;j ;X2;j jH¼h x1;j ; x2;j ¼ p1;2 ðl1 ; l2 ÞH x1;j ; l1 ; b1 H x2;j ; l2 ; b2 ;
l1 ¼1 l2 ¼1
m X
X m
l1
l2
  ðhÞ b1 b2
MX1;j ;X2;j jH¼h t1;j ; x2;j ¼ p1;2 ðl1 ; l2 Þ :
l1 ¼1 l2 ¼1
b1  t1;j b2  t2;j

Let

ðhÞ
X
m
ðhÞ
p1 ð l 1 Þ ¼ p1;2 ðl1 ; l2 Þ;
l2 ¼1

and

ðhÞ
X
m
ðhÞ
p2 ð l 2 Þ ¼ p1;2 ðl1 ; l2 Þ:
l1 ¼1

Also, we have
!
  X
m
ðhÞ
Y
n  
FXi;1 ;...;Xi;n jH¼h xi;1 ; . . .; xi;n ¼ pi ðli Þ H xi;j ; li ; bi ;
li ¼1 j¼1

X
m n

Y li !
  ðhÞ bi
MXi;1 ;...;Xi;n jH¼h ti;1 ; . . .; ti;n ¼ pi ð l i Þ :
li ¼1 j¼1
bi  ti;j

We obtain
Z X
m X
m
  ðhÞ    
FX1;j ;X2;j x1;j ; x2;j ¼ p1;2 ðl1 ; l2 ÞH x1;j ; l1 ; b1 H x2;j ; l2 ; b2 dFH ðhÞ;
l1 ¼1 l2 ¼1
h2AH
Z X
m X
m
l1
l2
  ðhÞ b1 b2
MX1;j ;X2;j t1;j ; x2;j ¼ p1;2 ðl1 ; l2 Þ dFH ðhÞ;
l1 ¼1 l2 ¼1
b1  t1;j b2  t2;j
h2AH
Z !
  X
m
ðhÞ
Y
n  
FXi;1 ;...;Xi;n xi;1 ; . . .; xi;n ¼ pi ð l i Þ H xi;j ; li ; bi dFH ðhÞ;
li ¼1 j¼1
h2AH
Z X
m n

Y li !
  ðhÞ bi
MXi;1 ;...;Xi;n ti;1 ; . . .; ti;n ¼ pi ð l i Þ dFH ðhÞ
li ¼1 j¼1
bi  ti;j
h2AH

and

123
Bivariate lower and upper orthant value-at-risk 349

 
FX1;1 ;...;X1;n ;X2;1 ;...;X2;n x1;1 ; . . .; x1;n ; x2;1 ; . . .; x2;n
Z X !
m X m Y n    
ðhÞ
¼ p1;2 ðl1 ; l2 Þ H x1;j ; l1 ; b1 H x2;j ; l2 ; b2 dFH ðhÞ;
l1 ¼1 l2 ¼1 j¼1
h2AH
 
MX1;1 ;...;X1;n ;X2;1 ;...;X2;n t1;1 ; . . .; t1;n ; t2;1 ; . . .; t2;n
Z X m X m Y n
l1
l2 !
ðhÞ b1 b2
¼ p1;2 ðl1 ; l2 Þ dFH ðhÞ:
l ¼1 l ¼1 j¼1
b1  t1;j b2  t2;j
1 2
h2AH

The above leads to


MS1 ;S2 jH¼h ðt1 ; t2 Þ
¼ MX1;1 ;...;X1;n ;X2;1 ;...;X2;n jH¼h ðt1 ; . . .; t1 ; t2 ; . . .; t2 Þ
X m X m Yn
l1
l2 !
ðhÞ b1 b2
¼ p1;2 ðl1 ; l2 Þ
l1 ¼1 l2 ¼1 j¼1
b1  t1 b2  t2
X m X m
nl1
nl2
ðhÞ b1 b2
¼ p1;2 ðl1 ; l2 Þ ;
l ¼1 l ¼1
b1  t1 b2  t2
1 2

which implies
m X
X m
ðhÞ
FS1 ;S2 jH¼h ðs1 ; s2 Þ ¼ p1;2 ðl1 ; l2 ÞH ðs1 ; n  l1 ; b1 ÞH ðs2 ; n  l2 ; b2 Þ:
l1 ¼1 l2 ¼1

Also, we have
Z X
m X
m
ðhÞ
FS1 ;S2 ðs1 ; s2 Þ ¼ p1;2 ðl1 ; l2 ÞH ðs1 ; n  l1 ; b1 ÞH ðs2 ; n  l2 ; b2 ÞdFH ðhÞ
l1 ¼1 l2 ¼1
h2AH
0 1
m X
X m Z
B ðhÞ C
¼ H ðs1 ; n  l1 ; b1 ÞH ðs2 ; n  l2 ; b2 Þ@ p1;2 ðl1 ; l2 ÞdFH ðhÞA
l1 ¼1 l2 ¼1
h2AH
X
m X
m
¼ p1;2 ðl1 ; l2 ÞH ðs1 ; n  l1 ; b1 ÞH ðs2 ; n  l2 ; b2 Þ;
l1 ¼1 l2 ¼1

where
Z
ðhÞ
p1;2 ðl1 ; l2 Þ ¼ p1;2 ðl1 ; l2 ÞdFH ðhÞ:
h2AH

We consider the following example, to illustrate the allocation values for two
dependent lines of business, based on the bivariate VaR. The dependence model is
represented by a bivariate mixture of Erlang distributions, as presented above.

123
350 H. Cossette et al.

Example 4.3 We suppose that b1 ¼ 1; b2 ¼ 0:5; H 2 fh1 ; h2 g; with PrðH ¼ h1 Þ ¼


s; PrðH ¼ h2 Þ ¼ 1  s: Let
 
ðh1 Þ
0:75 0:05 ðh2 Þ
0:65 0:15
p ¼ and p ¼ :
0:05 0:15 0:15 0:05

This bivariate model can be interpreted as being constituted of two possible


bivariate mixture of Erlang distributions, based on the sets of weights pðh1 Þ and pðh2 Þ ;
with probabilities s and 1 - s respectively.
ðh Þ ðh Þ _ Also,
It implies that pi k ð1Þ ¼ 0:8 and pi k ð2Þ ¼ 0:2; for i = 1, 2 and k ¼ 1; 2:
we have
pð1; 1Þ ¼ 0:75  s þ 0:65  ð1  sÞ;
pð1; 2Þ ¼ pð2; 1Þ ¼ 0:05  s þ 0:15  ð1  sÞ;
pð2; 2Þ ¼ 0:15  s þ 0:05  ð1  sÞ:
Note that E½X1  ¼ 1:2; E½X2  ¼ 2:4; E½S1  ¼ 12 and E½S2  ¼ 24:
Figure 7 illustrates the lower and upper VaR curves for the set (X1, X2), with
a = 99 %. We see that the upper orthant VaR is more affected by s, that changes in
the set of weights of the model. The lower orthant VaR is calculated in terms of the
multivariate mixted Erlang, and the two different weight sets do not expose a strong
difference, which is reflected by the three almost overlapping lower orthant VaR.
The sum of the squared differences between the lower orthant VaR are the following
:0.000363 for the curves with s = 0 and s = 0.5, 0.000365 for the curves with
s = 0.5 and s = 1 and 0.001455 for the curves with s = 0 and s = 0.5.
The following tables illustrates many features of the bivariate lower and upper
VaR. Table 4 shows the values obtained from the orthogonal projection and
proportional allocation methods for the bivariate lower orthant VaR, whereas
Table 5 show the results for the bivariate upper orthant VaR. We see that the
amounts from Table 4 are always higher than those from Table 5, as explained in
Remark 2.3. As illustrated in Fig. 7, we can see that the values obtained for the
18 12
τ=0 τ=0
17 τ=0.5 τ=0.5
τ=1 10 ↑ τ=1
VaRα(X2)
16

8 VaRα(X1)→
15
X 2,i

X 2,i

14 6

13
4

12 →VaRα(X1)

2
VaRα(X2)
11 ↓

0
5.5 6 6.5 7 7.5 8 8.5 9 0 1 2 3 4 5 6

X1,i X1,i

Fig. 7 Graphical representation of the bivariate lower and upper orthant VaR with a = 99 %, with
different dependence levels, for single bivariate risks

123
Bivariate lower and upper orthant value-at-risk 351

Table 4 Optimal couples based on the bivariate lower orthant VaR for individual risks (Xi,1, Xi,2)
a c Orthogonal Total Proportional Total

0.99 0 (6.4489, 11.6468) 18.0957 (6.0924, 12.1846) 18.2770


0.5 (6.4469, 11.6461) 18.0930 (6.0914, 12.1828) 18.2742
1 (6.4465, 11.6435) 18.0900 (6.0904, 12.1807) 18.2711
0.95 0 (4.6533, 8.0513) 12.7046 (4.2956, 8.5911) 12.8867
0.5 (4.6491, 8.0446) 12.6937 (4.0292, 8.5833) 12.6125
1 (4.6393, 8.0442) 12.6835 (4.2877, 8.5753) 12.8630

Table 5 Optimal couples based on the bivariate upper orthant VaR for individual risks (Xi,1, Xi,2)
a c Orthogonal Total Proportional Total

0.99 0 (1.2343, 8.4532) 9.6875 (2.7573, 5.5147) 8.2720


0.5 (1.2698, 8.4377) 9.7075 (2.8427, 5.6853) 8.5280
1 (1.5301, 8.4225) 9.9526 (2.9216, 5.8433) 8.7649
0.95 0 (0.7984, 5.6057) 6.4041 (1.8166, 3.6332) 5.4498
0.5 (0.8968, 5.5704) 6.4672 (1.8679, 3.7358) 5.6037
1 (0.9843, 5.5674) 6.5517 (1.9186, 3.8371) 5.7557

upper orthant VaR are more varying in terms of s, which means that the probability
of exceeding a specified level for both rv’s is more affected by the dependence
structure.
Figure 8 illustrates the bivariate lower and upper orthant curves for the set
(S1, S2). We have the same small difference between the lower orthant VaR curves.
We can draw the same conclusions as for the individual bivariate risks.
Tables 6 and 7 show that the same relation is obtained as for Tables 4 and 5. We
also see that the amounts are higher for each aggregate line of business. Again, the
total allocation for both lines is always smaller using the orthogonal projection
method than the proportional allocation method. h

60
74 τ=0 τ=0
τ=0.5 55 τ=0.5

72 τ=1 VaRα(X2) τ=1
50
70
45 VaRα(X1)→
68
40
66
S2
S2

35
64
30
62
25
60
→VaRα(S1) 20
58
VaRα(S2) 15
56 ↓
10
28 29 30 31 32 33 34 35 36 37 38 10 12 14 16 18 20 22 24 26 28 30

S1 S1

Fig. 8 Graphical representation of the bivariate lower and upper orthant VaR with a = 99 %, with
different dependence levels for the business lines

123
352 H. Cossette et al.

Table 6 Optimal couples based on the bivariate upper orthant VaR for lines of business (S1, S2), with
n = 10
a c Orthogonal Total Proportional Total

0.99 0 (30.5148, 58.0388) 88.5536 (29.6666, 59.3334) 89.0000


0.5 (30.5036, 58.0291) 88.5327 (29.6588, 59.3177) 88.9765
1 (30.4915, 58.0202) 88.5117 (29.6509, 59.3019) 88.9528
0.95 0 (26.3009, 48.6328) 74.9337 (25.1809, 50.3618) 75.5427
0.5 (26.2186, 48.5780) 74.7966 (25.1286, 50.2572) 75.3858
1 (26.1370, 48.5135) 74.6505 (25.0734, 50.1467) 75.2201

Table 7 Optimal couples based on the bivariate upper orthant VaR for lines of business (S1, S2), with
n = 10
a c Orthogonal Total Proportional Total

0.99 0 (16.9283, 45.9186) 62.8469 (20.3982, 40.7964) 61.1946


0.5 (19.0284, 48.1650) 67.1934 (21.8845, 43.7689) 65.6534
1 (20.0769, 49.5461) 69.6230 (22.6883, 45.3766) 68.0649
0.95 0 (11.4444, 39.0837) 50.5281 (15.6172, 32.2343) 47.8515
0.5 (15.2038, 38.6930) 53.8968 (17.5943, 35.1885) 52.7828
1 (16.9518, 40.6229) 57.5747 (18.8556, 37.7113) 56.5669

4.4 Confidence regions, optimal couples and bounds

This subsection intends to illustrate the use of bivariate set-valued quantiles and
motivate our results on confidence regions, optimal couples and bounds. Describing
relationships among different dimensions of an outcome is a basic actuarial
technique for explaining the behavior of financial security systems to concerned
businesses and policy makers. A VaR trade-off between two stock indices have been
studied by Cherubini and Luciano [5] and Cherubini et al. [6]. Studying the
bivariate VaR is useful in the decision process, based on a predefine a-level. It
shows the possible vectors that can be obtained if each risk reaches the a-level,
considering the dependence between them. At level a, the VaR movement between
two stock indices can be studied to understand and quantify the impact of
transfering a stock from one class to another.
Example 4.4 Consider a random vector X with joint cdf F/ 2 CðFX1 ; FX2 Þ such
that
 
F/ ðx1 ; x2 Þ ¼ C/ FX1 ðx1 Þ; FX2 ðx2 Þ ; ð4:5Þ
where C/ is the archimedean copula with generator / defined as
   
C/ u1 ; u2 ¼ /1 /ðu1 Þ þ /ðu2 Þ ; ð4:6Þ

for 0 B u1, u2 B 1. The generator / : ½0; 1 ! Rþ in (4.6) is a continuous, possibly

123
Bivariate lower and upper orthant value-at-risk 353

infinite, strictly decreasing convex function such that /(1) = 0. Denote by


VaRa;/;x1 ðXÞ; the a-curve of the bivariate lower orthant VaR of X. The latter may be
expressed explicitly by using (4.5) and (4.6). Indeed,
 
VaRa;/;x1 ðXÞ ¼ FX12 /1 /ðaÞ  /ðFX1 ðx1 Þ ; x1  FX1 1
ðaÞ: ð4:7Þ

This situation could represent a simplified example of a financial institution having


two different possible investments. Consider the first possibility to be investing in a
fund (X1) composed of two assets and the second one to be one of the two assets
from the fund (X2). Requirements in capital allocation should be calculated in order
to cover each possible investment at level a. In that sense, the use of the bivariate
VaR is appropriate. As illustration, consider exponential marginal cdf’s with
parameters k1 = 0.2 and k2 = 0.25, that is Fi ðxi Þ ¼ 1  eki xi ; i ¼ 1; 2; linked by
a Clayton copula with dependence parameter h = 2 defined by
 1=h
Ch ðu1 ; u2 Þ ¼ uh h
1 þ u2  1 ; h [ 0:
This copula belongs to the archimedean class, generated by /h(x) = h-1(x-h - 1),
h [ 0. Standard computations show that
  1=h
1 h
 
kx1 h 1
VaRa;h;x1 ðXÞ ¼  ln 1  a  1  e þ1 ; x1   lnð1  aÞ:
k2 k1
ð4:8Þ
The aim of this example is to provide allocation couples and a lower orthant
confidence region.
As a consequence of the convexity of the bivariate risk measures, Fig. 1
represents the a-curves for a = 0.95. In this situation, values allocated from both
criteria do not expose a large difference. In Table 8, the results from our capital
allocation criteria are presented.
Moreover, we use this example to examine the lower risk confidence curve at
level a. Table 9 provides the level k-1(a) given by (2.10) in terms of a. This allows

Table 8 Couples resulting from the Orthogonal projection and Proportional allocation criteria
a Orthogonal projection Total Proportional Total

0.95 (17.11, 15.82) 32.81 (18.26, 14.61) 32.86


0.99 (26.02, 21.54) 47.56 (26.45, 21.16) 47.62
0.995 (29.51, 24.33) 53.84 (29.94, 23.95) 53.89

Table 9 Lower orthant


a k-1(a)
confidence region levels
0.95 .8114
0.99 .9172
0.995 .9417

123
354 H. Cossette et al.

us to compute the confidence region at level a. The values in Table 9 represent the
level curves covering a % of the area under VaRa ðXÞ:
Figure 9 shows the confidence curves for a = 0.95. This curve coincides with
VaR0:8114 ðXÞ:
Note that VaR:8114 ðXÞ can be viewed as a threshold curve such that the
probability that the components of the loss X over the given time horizon are
simultaneously under this curve with probability 0.95. h
Example 4.5 Consider a portfolio of 10 risks of a third party liability in motor
insurance, where the risks Xi (i = 1, …, 10) have to be split into a physical
P damage
claim PXi,1 and a material damage claim Xi,2. Let us define S1 = 10 i=1Xi,1 and
S2 = 10 X
i=1 i,2 , the rv’s representing the total amount for each type of claim. The
random vector (S1, S2) is studied in order to establish the capital allocated to each
party, knowing that Xi,j, i = 1, …, 10 and j = 1,2 are covered by different parties.
Suppose that each physical damage rv Xi,1 follows an exponential distribution
(k = 1.5) and each material damage rv Xi,2 follows a Pareto distribution (aP = 2).
Also, Xi,1, i = 1, …, 10 and j = 1,2 are independent within each class and the
dependence between S1 and S2 is unknown. The capital allocation for each party
based on the optimal couple criterion provided in the previous section is a set
(S1, S2) from the lower orthant VaR.
The distance between each a-level curve is large, showing the effect of heavy-
tails. Also, Fig. 10 illustrates that the bivariate risk measure increases with a, with
respect to the order ; similarly to Fig. 6. It is interesting to note the impact of the

30
VaR.95(X1,X2)
28
VaR.8114(X1,X2)
26

24

22

20

18

16

14

12

10

10 15 20 25 30 35

Fig. 9 Graphical representation of the lower orthant confidence region VaR:8114 (X) coinciding for
VaR:95 (X)

123
Bivariate lower and upper orthant value-at-risk 355

800
α=.95
α=.99
700
Lower α−curve values α=.995

600

500

400

300

200

100
40 50 60 70 80 90 100 110 120 130
s1

Fig. 10 Graphical representation of the upper bounds of the lower orthant VaR VaRa;s1 (S)

aggregation of fat-tailed distributions (Pareto) on VaRa;s1 ðSÞ and on the upper


bounds compared to the optimal couples.
Since the dependence between each party is unknown, the study of the worst-case
scenario provides a considerable difference between the allocations that have to be
made from each party (Table 10). Now, suppose that Xi,2, i = 1,…,10 follow a
gamma distribution with shape and scale parameters a = 2 and b = 1 respectively,
and that the dependence between the sums of risks S1 and S2 is represented by an
archimedean copula. In such a situation, (4.7) provides an explicit form for
VaRa;/;s1 ðSÞ; where S = (S1, S2). This can be used to establish from (3.11) an upper
bound of VaRa(S1 ? S2) in terms of the cdf’s FS1 and FS2 , and the generator / of the
copula of (S1, S2), namely
n 1   o
VaRa ðS1 þ S2 Þ  min s1 þ FS1
2
/ /ðaÞ  /ðF ðs
S1 1 ÞÞ : ð4:9Þ
s1  FS1 ðaÞ
1

As an illustration, let us consider a Clayton copula with dependence parameter h. In


Table 11, we provide the upper bound on VaR.95(S1 ? S2), for different dependence
levels, such that Kendall’s tau is 0.25, 0.5 and 0.75 respectively. Table 11 allows us

Table 10 Optimal couples


a Orthogonal Proportional
from the upper bound
0.95 (35.3559, 859.8439) (35.2579, 1242.4300)
0.99 (46.1428, 2360.9330) (46.0628, 4036.78)
0.995 (50.7628, 3578.9410) (50.6861, 6861.1220)

123
356 H. Cossette et al.

Table 11 Upper bound of


h Upper bound
VaR.95(S1 ? S2)
2/3 56.38
2 56.34
6 56.25

to establish the upper bound, using the bivariate lower orthant VaR, considering the
dependence model of S1 and S2 (Table 10). h

Acknowledgments The authors wish to thank the anonymous reviewers for their detailed and helpful
comments. This work was partially supported by the Natural Sciences and Engineering Research Council
of Canada, the Fonds qubcois de la recherche sur la nature et les technologies, the Chaire en actuariat de
l’Université Laval, and the Faculty of Arts and Science of Concordia University.

References

1. Balakrishnan N, Lai CD (2009) Continuous bivariate distributions. Springer


2. Cai J, Li H (2005) Multivariate risk model of phase type. Insur Math Econ 36:137–152
3. Cai J, Li H (2007) Dependence properties and bounds for ruin probabilities in multivariate compound
risk models. J Multivar Anal 98:757–773
4. Chan WS, Yang H, Zhang L (2003) Some results on ruin probabilities in a two-dimensional risk
model. Insur Math Econ 32:345–358
5. Cherubini U, Luciano E (2001) Value at risk trade-off and capital allocation with copulas. Econ
Notes 30(2):235–256
6. Cherubini U, Luciano E, Vecchiato W (2004) Copula Methods in Finance. Wiley, London
7. Denuit M, Genest C, Marceau E (1999) Stochastic bounds on sums of dependent risks. Sci Agric
25:85–104
8. Di Bernardino E, Lalo T, Maume-Deschamps V, Prieur C (2011) Plug-in estimation of level sets in a
non-compact setting with application in multivariate risk theory. ESAIM Prob Stat
9. Embrechts P, Hoeing A, Juri A (2003) Using Copulae to bound the value-at-risk for functions of
dependent risks. Finance Stoch 7(2):145–167
10. Embrechts P, Puccetti G (2006a) Bounds for functions of multivariate risks. J Multivar Anal
97(2):526–547
11. Embrechts P, Puccetti G (2006b) Bounds for functions of dependent risks. Finance Stoch 10:341–352
12. Frank MJ, Schweizer B (1979) On the duality of generalized infimal and supremal convolutions.
Rendiconti di Matematica 12:1–23
13. Frees EW, Valdez EA (1998) Understanding relationships using copulas. N Am Actuar J 2(1):1–25
14. Guégan D, Hassani B (2012) Multivariate VaRs for operational risk capital computation : a vine
structure approach. Working paper
15. Johnson N, Kotz S, Balakrishnan N (1997) Discrete multivariate distributions. Wiley, London
16. Jouini E, Meddeb M, Touzi N (2004) Vector-valued coherent risk measures. Finance Stoch
8(4):531–552
17. Lee SCK, Lin XS (2012) Modeling dependent risks with multivariate Erlang mixtures. ASTIN Bull
42(1):153–180
18. Lehmann EL (1966) Some concepts of dependence. Ann Math Stat 37:1137–1153
19. Lindskog F, McNeil A (2003) Common poisson shock models: applications to insurance and credit
risk modelling. ASTIN Bull 33(2):209–238
20. Makarov GD (1982) Estimates for the distribution function of the sum of two random variables when
the marginal distributions are fixed. Theory Probab Appl 26:803–806
21. McNeil AJ, Frey R, Embrechts P (2005) Quantitative risk management. Princeton University Press,
Princeton

123
Bivariate lower and upper orthant value-at-risk 357

22. Mesfioui M, Quessy J-F (2005) Bounds on the value-at-risk for the sum of possibly dependent risks.
Insur Math Econ 37:135–151
23. Nelsen RB (2006) An introduction to Copulas. Springer, Berlin
24. Rüschendorf L (1982) Random variables with maximum sums. Adv Appl Prob 14:623–632
25. Sklar A (1959) Fonctions de répartition àn dimensions et leurs marges. Publications de l’Institut de
statistique de l’Université de Paris 8:229–231
26. Williamson RC, Downs T (1990) Probabilistic arithmetic I: numerical methods for calculating
convolutions and dependency bounds. Int J Approx Reason 4:89–158

123

You might also like