Insurance: Mathematics and Economics 49 (2011) 310–324

Contents lists available at SciVerse ScienceDirect

Insurance: Mathematics and Economics

journal homepage: [Link]/locate/ime

Asymptotics for risk capital allocations based on Conditional Tail Expectation

Alexandru V. Asimit a , Edward Furman b , Qihe Tang c,∗ , Raluca Vernic d,e
a
Cass Business School, City University, London EC1Y 8TZ, United Kingdom
b
Department of Mathematics and Statistics, York University, Toronto, Ontario M3J 1P3, Canada
c
Department of Statistics and Actuarial Science, University of Iowa, 241 Schaeffer Hall, Iowa City, IA 52242, USA
d
Department of Mathematics and Informatics, Ovidius University of Constanta, Constanta, Romania
e
Institute of Mathematical Statistics and Applied Mathematics, Bucharest, Romania

article info abstract

Article history: An investigation of the limiting behavior of a risk capital allocation rule based on the Conditional Tail
Received February 2011 Expectation (CTE) risk measure is carried out. More specifically, with the help of general notions of
Received in revised form Extreme Value Theory (EVT), the aforementioned risk capital allocation is shown to be asymptotically
May 2011
proportional to the corresponding Value-at-Risk (VaR) risk measure. The existing methodology acquired
Accepted 10 May 2011
for VaR can therefore be applied to a somewhat less well-studied CTE. In the context of interest, the
EVT approach is seemingly well-motivated by modern regulations, which openly strive for the excessive
MSC:
primary 91B30
prudence in determining risk capitals.
secondary 91G10 © 2011 Elsevier B.V. All rights reserved.
62H20
60E05

Keywords:
Asymptotic dependence and independence
Capital allocation
Conditional Tail Expectation
Extreme Value Theory
Heavy-tailed distributions
Value-at-Risk

1. Introduction by regulations. Then

VaRq [X ] := inf{x : F (x) ≥ q}
Let X denote an insurance risk. Speaking more formally, X is
a non-negative random variable defined on the probability space establishes arguably the most popular risk measure, which has
(Ω , F , Pr) and possessing a distribution function F (x) := Pr(X ≤ been a cornerstone of the financial risk measurement of the last
x) and a tail function F (x) := 1 − F (x), x ∈ R. A risk measure century. We note in passing that the Solvency II Accord designed by
is generally formulated as a functional, Q , from the space of the EU Commission sets q = 0.995 over a one-year time horizon.
distribution functions to [0, ∞]. Similarly (see, e.g. Bühlmann, Noticeably, the recent financial instability and, as a result,
1980), we can consider the functional Q as from X, the space of regulators’ inclination to excessive prudence in determining risk
capital requirements have to a certain extent enfeebled VaR’s
insurance risks, to [0, ∞], which we indeed often do in the sequel.
status. In this respect, the so-called tail-based risk measurement
Certainly, F establishes a meaningful ordering of X and, hence,
has emerged as a natural tool for quantifying insurance risks
it can be interpreted as a risk measure. However, for the sake of risk
while emphasizing the adverse effect of low probability but high
capital determination, it is desirable for the risk measure Q [F ] to
severity tail events. Thereby, a more pessimistic Conditional Tail
take on monetary units. Thus, Q [F ] = F (x) is naturally replaced
Expectation (CTE) risk measure is defined, for q ∈ (0, 1), as
with, e.g., its inverse, bringing us to the notion of Value-at-Risk
(VaR). Namely, let q ∈ (0, 1) denote the confidence level required CTEq [X ] := E[X |X > VaRq [X ]].
CTE is known as a coherent risk measure over the space of
continuous random variables. It belongs to both the distorted
∗Corresponding author. Tel.: +1 319 335 0730; fax: +1 319 335 3017. and weighted risk measures (see Wang, 1996; Dhaene et al.,
E-mail addresses: asimit@[Link] (A.V. Asimit), efurman@[Link] 2006; Furman and Zitikis, 2008a). Practically, CTE has already
(E. Furman), qihe-tang@[Link] (Q. Tang), rvernic@[Link] (R. Vernic). replaced VaR in regulatory requirements of, e.g., Canada, Israel
0167-6687/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/[Link].2011.05.002
 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324 311

and Switzerland. We note in passing that the current practice in 2. Main results under asymptotic dependence
the aforementioned countries is q = 0.99 over a one-year time
horizon. From now on, we consider a multi-line insurance business
Let Xi ∈ X, i = 1, . . . , d, denote d ∈ N insurance risks consisting of d non-negative risk variables X1 , . . . , Xd . Denote X =
and let Sd := X1 + · · · + Xd denote the aggregate risk. Then (X1 , . . . , Xd ) and Sd = di=1 Xi . Unless otherwise stated, all limit
∑
evaluating VaRq [Sd ] and CTEq [Sd ] is a somewhat basic phase of relationships hold as t → ∞ or q ↑ 1, letting the relations speak
the modern risk capital framework. Indeed, while it is of pivotal
for themselves. For two positive functions a(·) and b(·), we write
importance to determine the overall risk capital requirement for
a(·) ∼ cb(·) to mean strong equivalence, i.e., lim a(·)/b(·) = c for
an insurance company, it is of consequent interest to decompose
some positive constant c, and we write a(·) ≍ b(·) to mean weak
the aforementioned capital into the associated risk sources.
equivalence, i.e., 0 < lim inf a(·)/b(·) ≤ lim sup a(·)/b(·) < ∞.
To this end, the functional Q is naturally generalized beyond
We also denote lim inf a(·)/b(·) ≥ 1 and lim sup a(·)/b(·) ≤ 1 by
the conditional state independence, to a risk capital allocation
functional, A, from the space of the Cartesian product of X with a(·) & b(·) and a(·) . b(·), respectively.
itself to [0, ∞], and such that A[Xi , Xi ] = Q [Xi ], i = 1, . . . , d; see, To develop the main results of this paper we extensively employ
e.g., Furman and Zitikis (2008b). It should be noted that apart from the EVT techniques. A distribution function F is said to belong
purely regulatory interest, the functional A is often employed for, to the Maximum Domain of Attraction (MDA) of a non-degenerate
e.g., profitability analysis, pricing and quality control. distribution function G, written as F ∈ MDA(G), if there are some
Various functional forms of A have been proposed in the an > 0 and bn ∈ R for n ∈ N such that
literature, with the allocation based on the CTE risk measure,
lim F n (an x + bn ) = G(x).
formulated as n→∞

E[Xi |Sd > VaRq [Sd ]], i = 1, . . . , d, (1.1) Due to the Fisher–Tippett theorem (see Fisher and Tippett, 1928;
being arguably the most popular. See Section 6.3 of McNeil et al. Gnedenko, 1943), G is of one of the following three types:
(2005) for related discussions on this allocation as a consequence Fréchet type: Φα (x) = exp −x−α ,
 
x > 0, α > 0,
of the Euler principle, as well as Dhaene et al. (2011) for optimality
Gumbel type: Λ(x) = 1 − exp −e−x , −∞ < x < ∞,
 
studies of interest. Although (1.1) is quite elegant and satisfies
α
many desirable properties, its analytic tractability for generally Weibull type: Ψα (x) = exp {−(−x) } , x ≤ 0, α > 0.
distributed and possibly dependent X1 , . . . , Xd remains seldom
feasible. To emphasize the point, we refer the reader to Panjer Since distributions from MDA(Ψα ) have finite upper endpoints
and Jia (2001), Landsman and Valdez (2003), Valdez and Chernih while we are interested in risk variables with unbounded supports,
(2003), Cai and Li (2005), Furman and Landsman (2005, 2006, in this paper we shall consider the Fréchet and Gumbel cases only.
2008), Chiragiev and Landsman (2007), Vernic (2006, 2011) and Another important notion that is crucial for establishing our
Dhaene et al. (2008) for analytic expressions for (1.1) under main results is vague convergence. Let {µn , n ≥ 1} be a sequence of
specific multivariate distributions of a multi-line business and/or a measures on a locally compact Hausdorff space B with countable
portfolio of risks. base. Then µn converges vaguely to some measure µ, written as
In this paper, we follow a different route. Namely, as the exces- v
µn → µ, if for all continuous functions f with compact support
sive prudence of the current regulatory framework requires a con- we have
fidence level close to 1, the notion of Extreme Value Theory (EVT) ∫ ∫
becomes appropriate. We therefore study the asymptotic behav- lim f d µn = f dµ.
ior of capital allocations defined in (1.1) as q ↑ 1, when X1 , . . . , Xd n→∞ B B
are asymptotically dependent or asymptotically independent. Fol-
A thorough background on vague convergence is given by Kallen-
lowing Section 5.2 of McNeil et al. (2005), the asymptotic indepen-
berg (1983) and Resnick (1987).
dence between two random variables Xi and Xj with distribution
functions Fi and Fj is defined as
2.1. Fréchet case
lim Pr Fj (Xj ) > q|Fi (Xi ) > q = 0,
 
q ↑1
The next assumption is sufficient for our first main result.
while the asymptotic dependence is defined via this relation with
a positive limit. However, in this paper we slightly relax the notion
Assumption 2.1. Let X be a non-negative random vector with
of asymptotic dependence and define it as
marginal distributions F1 , . . . , Fd such that
lim inf Pr Fj (Xj ) > q|Fi (Xi ) > q > 0.
 
(1.2)
q ↑1 Pr(X1 > tx1 , . . . , Xd > txd )
lim := HF (x)
The notion of asymptotic dependence in higher dimensions is an t →∞ F¯1 (t )
obvious generalization of the two-dimensional definition above.
It is known that, for both Fréchet and Gumbel cases, the CTE and exists for all x = (x1 , . . . , xd ) ∈ [0, ∞]d \ {0}, where HF (·) is
VaR of a single risk are proportional for a high confidence level assumed to be a non-degenerate function and 0 is the vector of
(see Asimit and Badescu, 2010). Therefore, not surprisingly our zeroes.
main results show that capital allocations, as described by (1.1), This assumption implies that the marginal distribution func-
are asymptotically proportional to VaRq [Sd ] with a readily calcu-
tions are tail equivalent. That is,
lable coefficient of proportionality. This allows for the utilization
of the VaR-related machinery when dealing with the risk capital F̄j (t )
allocation based on CTE. 0 < lim < ∞, (2.1)
The remainder of the paper is organized as follows. The
t →∞ F̄i (t )
main results under asymptotic dependence and asymptotic for all 1 ≤ i, j ≤ d. In addition, there exists some α > 0 such
independence are formulated and proved in Sections 2 and 3, that, for each 1 ≤ i ≤ d, the distribution function of Xi is regularly
respectively. Relevant examples are discussed in Section 4, while varying with index α , written as Xi ∈ R−α ,
certain simulation studies verifying the accuracy of the main
results are carried out in Section 5. Finally, Section 6 concludes the F̄i (tx)
paper. lim = x−α , x > 0. (2.2)
t →∞ F̄i (t )
312 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324

Note that Xi ∈ R−α is equivalent to Xi ∈ MDA(Φα ) (see Resnick, Due to Assumption 2.1, it follows that
1987). Moreover,  
d
−
Pr ((X1 /t , . . . , Xd /t ) ∈ ·) v Pr (Sd > t ) ∼ µ x : xi > 1 F¯1 (t ); (2.5)
→ µ(·) (2.3) i =1
F¯1 (t )
for details, see Proposition 7.3 of Resnick (2007). Thus, Sd ∈ R−α ,
holds on [0, ∞]d \ {0}, where the measure µ is given by ∑d α
which gives that i=1 Ck = α− 1
; see Balkema and de Haan (1974),
µ ((x1 , ∞] × · · · × (xd , ∞]) := HF (x). (2.4) Alink et al. (2005), or Asimit and Badescu (2010).

The function HF satisfies certain properties (see Resnick, 1987), and Note 2.1. After the majority of this work had been done we
one of the most important is continuity on the set (0, ∞)d , but became aware of a forthcoming paper by Joe and Li (in press). Their
not necessarily on the boundary of its domain. The non-degeneracy Remark 2.3(2) suggests that our Theorem 2.1 is a consequence of
assumption ensures that the measure µ(·) does not put any mass their Theorem 2.2. However, this statement is not true because
on the boundary of the domain. their proof is based on the vague convergence property (see
Recall Proposition 0.8(vi) of Resnick (1987), which in our relation (2.3)), which can be used only for µ-continuous sets.
current context can be easily restated as: The reasoning behind applying the vague convergence property is
given in the proof of Theorem 2.1, for completeness. It is conveyed
Lemma 2.1. Let X1 and X2 be two random variables, both belonging that the asymptotic independence case (see Theorem 3.1) requires
to R−α for some α > 0. Then, for 0 < c < ∞, we have Pr (X2 > t ) ∼ an altered argumentation.
c Pr (X1 > t ) if and only if VaRq [X2 ] ∼ c 1/α VaRq [X1 ].
Proof. We first note that
With the help of Lemma 2.1 we can easily verify that ∫ ∞
Assumption 2.1 describes an asymptotic dependence case. E [Xk |Sd > t] = Pr (Xk > z |Sd > t ) dz
0

Lemma 2.2. Under Assumption 2.1 the components of X = (X1 , t

Pr (Xk > z , Sd > t ) ∞
Pr(Xk > z )
∫ ∫
. . . , Xd ) are pairwise asymptotically dependent. = dz + dz
0 Pr (Sd > t ) t Pr (Sd > t )
Proof. As an illustration we consider relation (1.2) for (i, j) = = I1 (t ) + I2 (t ). (2.6)
(1, 2). By relations (2.1) and (2.2) and Lemma 2.1, there is some
1/α For the first part of (2.6), I1 (t ), we have
c2 > 0 such that VaRq [X2 ] ∼ c2 VaRq [X1 ]. As q ↑ 1, or, equiva-
lently, as t = VaRq [X1 ] → ∞, we have 1
Pr(Xk > tz , Sd > t )
∫
I1 (t ) = t dz
Pr (F1 (X1 ) > q, F2 (X2 ) > q) 0 Pr(Sd > t )
Pr (F2 (X2 ) > q|F1 (X1 ) > q) =
Pr (F1 (X1 ) > q)
 d

µ x : xk > z , xi > 1
∑
Pr X1 > VaRq [X1 ], X2 > VaRq [X2 ] 1
  ∫
i =1
≥   ∼t   dz , (2.7)
d
Pr X1 ≥ VaRq [X1 ] 0
µ x: xi > 1
∑
1/α
 
Pr X1 > t , X2 > 2c2 t i=1

& which is a consequence of relations (2.3) and (2.5), the Dominated

Pr (X1 > t ) Convergence Theorem and Proposition A2.12 of Embrechts et al.
1/α
 
→ HF 1, 2c2 , 0, . . . , 0 > 0. (1997).
  Note that the latter proposition can be applied since
µ ∂ x : xk > z , xi > 1
∑d
i=1 = 0 for all z ≥ 0. In fact,
This proves relation (1.2) for (i, j) = (1, 2). 
Assumption 2.1 implies that
We are now able to provide asymptotic expressions for the risk
µ {xi = z , (x1 , . . . , xi−1 , xi+1 . . . , xd ) ∈ A} = 0
capital allocation for a multi-line insurance business. Noticeably,
switching the context to a portfolio consisting of ai units of Xi for for all z > 0 and any relatively compact set A in [0, ∞]d−1 \{0}. This
all 1 ≤ i ≤ d, similar results as given in the next theorem can be is still true for z = 0 under the additional condition that the set A
obtained by replacing the measure µ from (2.4) with is bounded away from {0}. Moreover, the proof of Theorem 3.2 of
Kortschak and Albrecher (2009) justifies the fact that no mass is put
µa ((x1 , ∞] × · · · × (xd , ∞])
by the measure µ over the line i=1 xi = 1 and any neighborhood
∑d
Pr(a1 X1 > tx1 , . . . , ad Xd > txd ) around ∞.
= lim
t →∞ Pr(a1 X1 > t ) For the second part of relation (2.6), I2 (t ), we have
= HF (x1 /a1 , . . . , xd /ad )a−α
1 . Pr(Xk > t ) Pr(X1 > t ) ∞
Pr(Xk > tz )
∫
I2 (t ) = t dz
Pr(X1 > t ) Pr (Sd > t ) 1 Pr(Xk > t )
Theorem 2.1. Let X be a random vector satisfying Assumption 2.1. If t µ(x : xk > 1)
X1 ∈ R−α with α > 1 then we have, for all 1 ≤ k ≤ d, ∼ , (2.8)
α−1
 d
µ x: xi > 1
∑
E Xk |Sd > VaRq [Sd ] ∼ Ck VaRq [Sd ],
 
i=1

where, with µ defined by (2.4), where in the last step above we have applied the relations (2.5) and
Pr(Xk > t ) ∼ µ(x : xk > 1) Pr(X1 > t )
 d

1
1
µ(x : xk > 1) + 0 µ x : xk > z , xi > 1 dz
∑
α−1
i =1 due to (2.3), and made use of the Dominated Convergence Theorem
Ck =  d
 .
as justified by Proposition 0.8 of Resnick (1987).
µ x: xi > 1
∑
Plugging (2.7) and (2.8) into (2.6) yields E [Xk |Sd > t] ∼ Ck t. The
i =1
proof is complete. 
 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324 313

As a consequence of Theorem 2.1, we now express the capital al- Lemma 2.3. Under Assumption 2.2 the components of X = (X1 ,
locations in terms of the VaR of the reference risk. The proof simply . . . , Xd ) are pairwise asymptotically dependent.
uses Lemma 2.1 and relation (2.5) and for this reason is omitted.
Proof. As in the proof of Lemma 2.2, we consider relation (1.2) for
Corollary 2.1. Under the conditions of Theorem 2.1, it holds for all (i, j) = (1, 2). By (2.12) and (2.9) with F replaced by F1 , we have
1 ≤ k ≤ d that
F¯2 (t + a(t )s)
= ν(x : x2 > 0)e−s .
 
d lim
1/α
F¯1 (t )
−
E Xk |Sd > VaRq [Sd ] ∼ Ck µ xi > 1 VaRq [X1 ],
 
x: t →∞

i=1
Fix some s large enough such that the right-hand side above is
where the constants Ck are defined as in Theorem 2.1. smaller than 1. Then it holds for all large t that F¯2 (t + a(t )s) < F¯1 (t ),
from which it follows that
2.2. Gumbel case
VaRF1 (t ) [X2 ] ≤ t + a(t )s.
The Gumbel case is further investigated in the presence of Analogously to the proof of Lemma 2.2, as q ↑ 1, or, equivalently,
asymptotic dependence. Since we are only interested in risks with as t = VaRq [X1 ] → ∞,
unbounded supports, all individual risks are assumed to have an
infinite upper endpoint. It is well known (see Embrechts et al., Pr (F2 (X2 ) > q|F1 (X1 ) > q)
1997) that if F ∈ MDA(Λ), then there exists a positive, measurable
Pr X1 > VaRq [X1 ], X2 > VaRq [X2 ]
 
function a(·) such that ≥  
Pr X1 ≥ VaRq [X1 ]
F̄ (t + a(t )s)
= e− s Pr X1 > t , X2 > VaRF1 (t ) [X2 ]
 
lim (2.9)
t →∞ F̄ (t ) ≥
Pr (X1 ≥ t )
for any s ∈ R. In addition, the latter holds locally uniformly in s
(see Resnick, 1987). Recall that the auxiliary function a(·) satisfies Pr (X1 > t , X2 > t + a(t )s)
≥
a(t ) = o(t ) and is such that the relation Pr (X1 ≥ t )
a (t + a(t )x) → ν(x : x1 > 0, x2 > s) > 0,
lim =1 (2.10)
t →∞ a( t ) where in the last step we used the fact Pr (X1 ≥ t ) ∼ Pr (X1 > t )
holds locally uniformly in x. Moreover, this auxiliary function can due to Corollary 1.6 of Resnick (1987). This proves relation (1.2) for
be chosen as the mean excess function of F , i.e., (i, j) = (1, 2). 
∞
F̄ (s)
∫
a(t ) = ds. Note 2.2. Due to Assumption 2.2, it holds that
t F̄ (t )  
d
See Section 3.3 of Embrechts et al. (1997) for more details. −
The next assumption is sufficient for our main results of this Pr (Sd > dt ) ∼ ν x: xi > 0 F¯1 (t ), (2.13)
subsection. i =1

Assumption 2.2. Let X = (X1 , . . . , Xd ) be a non-negative random which is a consequence of the vague convergence in (2.11)
vector with marginal distributions F1 , . . . , Fd such that and Proposition 4.1 of Klüppelberg and Resnick (2008). Note
that Assumption 2.2 allows us to withdraw the equal marginal
Pr(X1 > t + a(t )x1 , . . . , Xd > t + a(t )xd ) distributions assumption from their proposition. Thus,
lim := HG (x)
t →∞ F¯1 (t )
Pr (Sd > t + a (t /d) ds)
exists for all x ∈ Rd , where HG (·) is assumed to be a non-degenerate lim = e− s (2.14)
t →∞ Pr (Sd > t )
function.
holds for any s ∈ R and relation (2.9) is satisfied by Sd with an
Assumption 2.2 implies that
auxiliary function ã(t ) := a(t /d)d. The latter implies that Sd ∈
MDA(Λ), which gives that
  
Pr
X1 −t
a(t )
, . . . , Xad(−t )t ∈ · v
→ ν(·) (2.11) E [Sd |Sd > t] ∼ t . (2.15)
F¯1 (t )
holds on [−∞, ∞]d \ {−∞} for the measure ν given by
Relation (2.15) agrees with the next theorem, which provides
ν ((x1 , ∞] × · · · × (xd , ∞]) := HG (x), the asymptotic expressions for the capital allocations for a multi-
where {−∞} = (−∞, . . . , −∞). Moreover, the distribution line insurance business of d asymptotically dependent risks that
functions are tail equivalent with belong to MDA(Λ).

F̄i (t )
lim = ν(x : xi > 0), 1 ≤ i ≤ d, (2.12) Theorem 2.2. Let X be a random vector satisfying Assumption 2.2.
t →∞ F¯1 (t ) Then it holds for all 1 ≤ k ≤ d that
and, therefore, all belong to MDA(Λ). One choice for the function 1
a(·) from Assumption 2.2 is given by the auxiliary function E Xk |Sd > VaRq [Sd ] ∼
 
VaRq [Sd ].
corresponding to the random variable X1 as described in relation d
(2.9). In fact, all marginal distributions admit the same auxiliary
function. Proof. The first step of the proof is developed as
We now show that the non-degeneracy of HG in Assumption 2.2 ∫ t ∫ ∞
ensures that the random vector X has the asymptotic dependence E [Xk |Sd > dt] = + Pr (Xk > z |Sd > dt ) dz
property, which is a key property in describing the tail probability 0 t
of Sd (see Section 4.1 of Klüppelberg and Resnick, 2008). = I1 (t ) + I2 (t ). (2.16)
314 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324

The second integral can be easily reduced to 1997, page 148), written as F ∈ R−∞ , characterized by
∞
Pr(Xk > z ) F̄ (tx)
∫
0, x > 1,

I2 (t ) ≤ dz . (2.17) lim =
t Pr (Sd > dt ) t →∞ F̄ (t ) ∞, 0 < x < 1.
The change of variables z = t + a(t )v leads to
The following result is similar to Lemma 2.1 but for the rapid
∞
Pr(Xk > z ) ∞
Pr (Xk > t + a(t )v)
∫ ∫
variation case:
dz = a(t ) dv = o(t )
t Pr(Xk > t ) 0 Pr(Xk > t )
Lemma 2.4. Let X1 and X2 be two random variables with distribution
due to the Dominated Convergence Theorem, relations (2.9) and
functions F1 and F2 in R−∞ . If F¯1 (lt ) ≍ F¯2 (t ) for some l > 0 then, for
a(t ) = o(t ). Thus, the latter relation, together with (2.13) and
every m > 0, as q = 1 − p ↑ 1,
(2.17), concludes that

I2 (t ) = o(t ). (2.18) VaRq [X1 ] ∼ lVaR1−mp [X2 ]. (2.22)

It only remains to investigate the first term of (2.16), which is

equal to Proof. Let 0 < ε < 1 be arbitrarily fixed. As p ↓ 0, there is some
∫ t +(d−1)a(t )s ∫ t
 function b(p) → 0 such that
I1 (t ) = t − + Pr (Xk ≤ z |Sd > dt ) dz
F¯1 (1 + ε)lVaR1−mp [X2 ]
 
0 t +(d−1)a(t )s

= b(p)F¯1 (1 + ε/2)lVaR1−mp [X2 ]

 
= t − I11 (t , s) − I12 (t , s), (2.19)
≍ b(p)F¯2 (1 + ε/2)VaR1−mp [X2 ] ≤ b(p)mp.
 
where s is a negative number.
Recall that a(·) is a positive function. It holds for every s < 0
Similarly, as p ↓ 0, there is some function c (p) → ∞ such that
that
I11 (t , s) F¯1 (1 − ε)lVaR1−mp [X2 ]
 
≤ Pr (Xk ≤ t + (d − 1)a(t )s|Sd > dt )
t = c (p)F¯1 (1 − ε/2)lVaR1−mp [X2 ]
 
 
d
≍ c (p)F¯2 (1 − ε/2)VaR1−mp [X2 ] ≥ c (p)mp.
 
Xi > (d − 1)t − (d − 1)a(t )s
∑
Pr
i=1,i̸=k
≤ Thus, for all 0 < p < 1 sufficiently close to 0, we have
Pr (Sd > dt )
F¯1 (1 + ε)lVaR1−mp [X2 ] < p < F¯1 (1 − ε)lVaR1−mp [X2 ] .
     
d
ν x: xi > 0
∑
i=1,i̸=k It follows that
∼  d
 es , (2.20)
ν x:
∑
xi > 0 (1 − ε)lVaR1−mp [X2 ] ≤ VaRq [X1 ] ≤ (1 + ε)lVaR1−mp [X2 ].
i =1
By the arbitrariness of ε , this leads to relation (2.22). 
as a result of (2.13) and (2.14). Thus,
I11 (t , s) Similar to the Fréchet case, Theorem 2.2 allows us to express the
lim lim sup = 0. capital allocations in terms of the reference risk measure, VaRq [X1 ].
s↓−∞ t →∞ t
Finally, relation (2.13), Lemma 2.4 and the fact that Sd ∈ R−∞
As before, the change of variables z = t + (d − 1)a(t )v yields imply the following result:
∫ 0
I12 (t , s) = (d − 1)a(t ) Pr (Xk ≤ t + (d − 1)a(t )v|Sd > dt ) dv
s
Corollary 2.2. Under the conditions of Theorem 2.2, as q = 1−p ↑ 1

d
 it holds that
0
Xi > (d − 1)t − (d − 1)a(t )v dv
∑
Pr
E Xk |Sd > VaRq [Sd ] ∼ VaR1−p/K [X1 ],
s
 
i=1,i̸=k (2.23)
≤ (d − 1)a(t )
Pr (Sd > dt )  
where K = ν x : xi > 0 .
 
∑d
d i =1
Xi > (d − 1)t
∑
Pr
i=1,i̸=k
≤ (d − 1)a(t ) |s| Lemma 2.4 shows that VaRq [X1 ] as a function of p is slowly vary-
Pr (Sd > dt ) ing as p ↓ 0 (see also Proposition 0.8(v) of Resnick, 1987). There-

d
 fore, the right-hand side of (2.23) can be changed to VaR1−mp [X1 ]
ν x: xi > 0 for every constant m > 0. However, in view of relation (2.13), the
∑
i=1,i̸=k most rational choice for m should be m = 1/K .
∼ (d − 1)a(t ) |s|  d
 = o(t ), (2.21)
We have thus obtained an appealing expression for the
ν x: xi > 0
∑
i =1
CTE-based risk capital allocations for a multi-line insurance
business consisting of asymptotically dependent risks that are not
due to a(t ) = o(t ) and relations (2.13) and (2.14). extremely heavy tailed. The fact that all risks belong to MDA(Λ)
A combination of relations (2.19)–(2.21) yields that I1 (t ) ∼ t. allows us to conclude that the marginal risk capitals under CTE are
Plugging this and (2.18) into (2.16) concludes the proof. 
equal for conservative scenarios. It is interesting that this remains
Recall that every distribution F from MDA(Λ) with an infinite true without assuming exchangeable risks, a situation in which the
upper endpoint also has a rapidly varying tail (see Embrechts et al., allocations are obviously equal at any degree of safeness.
 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324 315

3. Main results under asymptotic independence Then for all positive x1 and x2 we have
Pr(X1 > tx1 , X2 > tx2 ) Pr(X1 ∧ X2 > tx1 ∧ tx2 )
There has been a particular interest in understanding the ≤ = o(1).
tail behavior of the sum of asymptotically independent random Pr(X1 > t ) Pr(X1 > t )
variables with heavy tails (see Albrecher et al., 2006; Ko and Tang, Thus,
2008; Asmussen and Rojas-Nandayapa, 2008; Geluk and Tang,  −α
2009; Mitra and Resnick, 2009). A similar pattern for the CTE- Pr(X1 > tx1 , X2 > tx2 )
x1 , x1 > 0, x2 = 0,
capital allocations is expected. The results are less homogeneous lim = c x−α , x1 = 0, x2 > 0,
t →∞ Pr(X1 > t ) 02, 2 otherwise.
under asymptotic independence, and various assumptions are
provided in this case. The Fréchet and Gumbel cases exhibit
Assumption 3.1 describes a situation in which
fundamentally different behaviors for the capital allocations and,
therefore, analyses are made separately. d
−
Pr (Sd > t ) ∼ F¯1 (t ) ci ; (3.3)
i =1
3.1. Fréchet case
see Lemma 2.1 of Davis and Resnick (1996) or Proposition 7.3 of
We first restrict our attention to a portfolio of risks with Resnick (2007).
regularly varying tails. Two overlapping sets of assumptions are The first main result of this subsection is now given for a multi-
made, but neither of them is a consequence of the other. The first line insurance business consisting of asymptotically independent
assumption is a natural extension of Assumption 2.1 under the risks with regularly varying tails.
asymptotic independence setting.
Theorem 3.1. Let X be a random vector satisfying Assumption 3.1. If
Assumption 3.1. Let X be a non-negative random vector with X1 ∈ R−α with α > 1, then for all 1 ≤ k ≤ d we have
marginal distributions F1 , . . . , Fd such that α ck
E Xk |Sd > VaRq [Sd ] ∼
 
VaRq [Sd ],
Pr(X1 > tx1 , . . . , Xd > txd ) α−1∑
d
lim := µI ((x1 , ∞] × · · · × (xd , ∞]) ci
t →∞ F¯1 (t ) i =1

exists for all x ∈ [0, ∞]d \ {0}, where µI is a Radon measure such where the constants ci are given by (3.1).
that µI ((0, ∞]d ) = 0.
Proof. Similar to the proof of Theorem 2.1,
Since the limiting measure is Radon, Assumption 3.1 suggests t
Pr (Xk > z , Sd > t ) ∞
Pr(Xk > z )
∫ ∫
that all marginal distributions are regularly varying tailed with E [Xk |Sd > t] = dz + dz
possibly different indexes, αi , such that α1 = ∧1≤i≤n αi . If these 0 Pr (Sd > t ) t Pr (Sd > t )
indexes are not equal, then the measure µI does not put mass on = I1 (t ) + I2 (t ). (3.4)
the vast majority of the boundary of its domain, [0, ∞]d \ {0}. In
such a case, the problem becomes trivial and only components Now, for any 0 < z ≤ 1 we have
with indexes equal to α1 have contributions to extreme events Pr(Xk > tz , Sd > t ) Pr(Xk > t ) Pr(tz < Xk ≤ t , Sd > t )
related to the aggregate portfolio. Therefore, without loss of = +
generality it is further assumed that all marginal distributions Pr(X1 > t ) Pr(X1 > t ) Pr(X1 > t )
belong to MDA(Φα ), since otherwise the main results remain
 
d
−
unchanged. Then, for each 1 ≤ i ≤ d, there exists a positive ∼ ck + µI x : z < xk ≤ 1, xi > 1
i=1
constant ci such that
= ck ,
F̄i (t ) ∼ ci F¯1 (t ) (3.1)
where Proposition A2.12 of Embrechts et al. (1997) and the vague
and convergence
  applied over the µI -negligible
property in (3.2) are
set x : z < xk ≤ 1, xi > 1 . The latter, relation (3.3) and the
∑d
Pr ((X1 /t , . . . , Xd /t ) ∈ ·) v i=1
→ µI (·) (3.2) Dominated Convergence Theorem yield that
Pr(X1 > t )

holds on [0, ∞]d \{0}. Furthermore, the measure µI puts mass only
1
Pr(Xk > tz , Sd > t )
∫
ck
I1 ( t ) = t dz ∼ t. (3.5)
on the coordinate axes due to the fact that µI ((0, ∞]d ) = 0. Thus, 0 Pr(Sd > t ) ∑d
for all z > 0, ci
i=1
µI x : x1 = · · · = xi−1 = xi+1 = · · · = xd = 0, xi ∈ (z , ∞]
 
As in the proof of Theorem 2.1, the second term in (3.4) satisfies
= ci z −α .
1 ck
It is interesting to outline the link between Assumption 3.1 I2 ( t ) ∼ t. (3.6)
α−1∑
d
and asymptotic independence. Under Assumption 3.1, following ci
the proof of Lemma 2.2 we can easily verify that the components i=1

of X are pairwise asymptotically independent. As for the inverse Putting together (3.4)–(3.6), we obtain
statement, for simplicity consider X = (X1 , X2 ). Assume that X1
α ck
and X2 belong to MDA(Φα ), have strongly equivalent tails as in (3.1) E [Xk |Sd > t] ∼ t,
and are asymptotically independent. For this case, the asymptotic α−1∑
d
ci
independence is equivalent to i=1

lim Pr (X2 > t |X1 > t ) = 0. which completes the proof. 

t →∞
316 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324

An even simpler expression for the CTE-capital allocations is The next result is crucial for developing our second main result
given without proof in the next corollary, which is a consequence of this subsection.
of Theorem 3.1, Lemma 2.1 and relation (3.3).
Lemma 3.2. Let X be a random vector satisfying Assumption 3.2 such
Corollary 3.1. Under the conditions of Theorem 3.1, it holds for all
1 ≤ k ≤ d that that each component has a regularly varying tail. Then the relation
  
  α1 −1 d  d
α d
− −
Pr Xi > t  Xk = x ∼ hk (x) F̄i (t )
− 
E Xk |Sd > VaRq [Sd ] ∼
 
ck ci VaRq [X1 ],
α−1 i=1,i̸=k i=1,i̸=k

i =1

where the constants ci are given by (3.1). holds uniformly for 0 ≤ x < ∞ for all 1 ≤ k ≤ d.

Our next assumption is motivated by the work of Asimit and Proof. The proof below proceeds for d ≥ 3 and k = 1, but it can
Badescu (2010).
easily be adjusted so that it is valid for d ≥ 2 and k ̸= 1. We first
Assumption 3.2. Let X be a nonnegative random vector with derive an asymptotic lower bound. Since each Xi is nonnegative, by
marginal distribution functions F1 , . . . , Fd . Assume that there are Bonferroni’s inequality we have, uniformly for 0 ≤ x < ∞,
some measurable and bounded functions hi (·) : (0, ∞) → (0, ∞) 
d
  
d
 
such that, for distinct i, j ∈ {1, . . . , d}, the relation
−   
Pr Xi > t  X1 = x ≥ Pr Xi > t  X1 = x
 

Pr Xj > t |Xi = x ∼ F̄j (t )hi (x)

  i =2
 i =2

(3.7)
d
holds uniformly for 0 ≤ x < ∞. In addition, for d ≥ 3, it is also
− −
Pr (Xi > t |X1 = x) − Pr Xi > t , Xj > t |X1 = x
 
≥
assumed that i =2 2≤i̸=j≤d

Pr Xj > t , Xk > t |Xi = x = o F̄j (t ) + F̄k (t ) hi (x)

   
(3.8) d
−
∼ h1 ( x ) F̄i (t ),
holds uniformly for 0 ≤ x < ∞ for distinct i, j, k ∈ {1, . . . , d}.
i =2
The uniformity of relation (3.7) is understood as
   where in the last step we applied both (3.7) and (3.8). Next,
 Pr X > t |X = x  we derive the corresponding asymptotic upper bound. For an
j i
lim sup  − 1 = 0
 
t →∞ 0≤x<∞  F̄j (t )hi (x)  arbitrarily fixed constant 0 < ε < 1, we have
     
d d
while the uniformity of relation (3.8) as −   
Pr Xi > t  X1 = x ≤ Pr Xi > (1 − ε)t  X1 = x
 
Pr Xj > t , Xk > t |Xi = x
 
= 0. i =2
 i =2

lim sup
F̄j (t ) + F¯k (t ) hi (x)
  
t →∞ 0≤x<∞
 
−d d 
+ Pr Xi > t , Xi ≤ (1 − ε)t  X1 = x

The recent work of Li et al. (2010) discussed the verification of the 
i=2 i=2
uniformity of relation (3.7) and the boundedness of the functions
hi (·). Clearly, the uniformity property implies that E [hi (X )] = 1. = I1 (t , ε) + I2 (t , ε). (3.9)
For the dependence structures discussed by Asimit and Badescu
(2010) and Li et al. (2010) with strongly equivalent tails, both Let αi > 0 be the regularly varying index of F̄i . Then relation (3.7)
Assumptions 3.1 and 3.2 are satisfied. It is not difficult to prove gives that
that a bivariate random vector with strongly equivalent tails and
d
dependence structure given by the Marshall–Olkin copula (see, −
I1 (t , ε) ≤ Pr (Xi > (1 − ε)t |X1 = x)
e.g. Nelsen, 1999, page 46) satisfies only Assumption 3.1. It will
i =2
be later seen in Note 3.1 that within a particular (yet fairly
general) setting, Assumption 3.2 may provide a refinement for d
−
the capital allocations given by Theorem 3.1. Therefore, neither of ∼ h1 (x) (1 − ε)−αi F̄i (t ). (3.10)
Assumptions 3.1 and 3.2 is dominated by the other one. i =2
We first establish that random vectors satisfying Assump-
Similarly, by (3.8),
tion 3.2 have pairwise asymptotically independent components.
I2 (t , ε)
Lemma 3.1. Relation (3.7) implies that Xi and Xj are asymptotically   
independent. d d d
t
−   
= Pr Xi > t , Xj > , Xi ≤ (1 − ε)t  X1 = x

Proof. By definition, i=2 j=2
d − 1 i=2 
Pr Fi (Xi ) > q, Fj (Xj ) ≥ q
    
d d
Pr Fj (Xj ) > q|Fi (Xi ) > q ≤ t
  − − 
Pr (Fi (Xi ) > q) ≤ Pr Xi > t , Xj > , Xj ≤ (1 − ε)t  X1 = x

∞ j =2 i =2
d−1 
Pr Fi (Xi ) > q|Xj = x Pr Xj ∈ dx
   
VaRq [Xj ]−
  
= . − d − d
t 

Pr (Fi (Xi ) > q) ≤ Pr Xi > ε t , Xj >  X1 = x
j =2 i=2,i̸=j
d − 1
Hence, by the uniformity of (3.7), the right-hand side above con-
εt
  
verges to − t 
≤ Pr Xi > , Xj >  X1 = x
∫ ∞
2≤i̸=j≤d
d−2 d − 1
hj (x) Pr Xj ∈ dx = 0.
 
lim
q ↑1
 
VaRq [Xj ]− − d

Therefore, Xi and Xj are asymptotically independent. 

= o h1 ( x ) F̄i (t ) . (3.11)
i =2
 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324 317

Substituting (3.10) and (3.11) into (3.9) and noticing the arbitrari- We first investigate the numerator from (3.15). By Lemma 3.2, for
ness of ε , we obtain every ε > 0, there is some t0 > 0 such that, for all t ≥ t0 and

d



d
x > 0,
− −
Pr Xi > t  X1 = x . h1 (x) F̄i (t ).

d
−
i=2
 i=2 (1 − ε)h1 (x) F̄i (t ) ≤ Pr (Sd − X1 > t |X1 = x)
i=2
The proof is complete. 
d
Now we are ready to state the second main result of this
−
≤ (1 + ε)h1 (x) F̄i (t ). (3.16)
subsection. i=2

Theorem 3.2. Let X be a random vector satisfying Assumption 3.2 According to this constant t0 , for all t ≥ t0 we split the integral into
such that each component Xi has a regularly varying tail with index three parts as
αi > 1. In addition, if each hi (·) is a regularly varying function, then 3 t − t0 t ∞
− ∫ ∫ ∫
− d Ii (t ) = + +
Pr (Sd > t ) ∼ F̄i (t ) (3.12) i =1 0− t − t0 t

i =1 x Pr (Sd − X1 > t − x|X1 = x) F1 (dx). (3.17)

and
Clearly,
d
α ∞
E [Xk hk (Xk )] F̄i (t ) + t F¯k (t ) α −k 1
∑ ∫
i=1,i̸=k
k I3 ( t ) = xF1 (dx) (3.18)
E [Xk |Sd > t] ∼ . (3.13) t
d
F̄i (t ) and
∑
i =1 ∫ t
I2 ( t ) ≤ xF1 (dx) = o (I3 (t )) . (3.19)
All examples presented in the work of Li et al. (2010) give the t − t0
regular variation property of the functions hi (·). If relation (3.1)
holds for each 1 ≤ i ≤ d, then one may use Theorem 3.2 to recover By (3.16),
Corollary 3.1. I1 (x)
1−ε ≤ ≤ 1 + ε. (3.20)
d 
Note 3.1. Let us assume that there exist finite constants ci , 1 ≤ ∑ t − t0
xh1 (x)F̄i (t − x)F1 (dx)
i ≤ d, such that F̄i (t ) ∼ ci F¯1 (t ) for all 1 ≤ i ≤ d − 1 and i=2
0−

F¯d (t ) ∼ t F¯1 (t )cd . Thus, αi = α1 for all 1 ≤ i ≤ d − 1 and

αd = α1 − 1. Suppose α1 > 2 so that Xd has a finite mean as well. Define a distribution F∗ by
In addition, Pr (Sd > t ) ∼ F¯d (t ), which together with Lemma 2.1, xh1 (x)
implies that VaRq [Sd ] ∼ VaRq [Xd ]. Some algebraic manipulations F∗ (dx) = F1 (dx),
E [X1 h1 (X1 )]
lead to
α1 ck where E [X1 h1 (X1 )] is a finite positive normalizing constant.
lim E [Xk |Sd > t] = E [Xk hk (Xk )] + Clearly, F∗ has a regularly varying tail as well. Thus, by Lemma 1.3.1
t →∞ α1 − 1 cd
of Embrechts et al. (1997) (see also Proposition 1.2 or Theorem 2.1
for all 1 ≤ k ≤ d − 1 and of Cai and Tang, 2004),
αd t − t0
E [Xd |Sd > t] ∼ t .
∫
αd − 1 xh1 (x)F̄i (t − x)F1 (dx)
0−
Obviously, these relations generate the asymptotic capital alloca- ∫ ∞ ∫ ∞ ∫ t

tion estimates = E [X1 h1 (X1 )] − − F̄i (t − x)F∗ (dx)
α1 ck 0− t t − t0
lim E Xk |Sd > VaRq [Sd ] = E [Xk hk (Xk )] +
 
= (1 + o(1)) E [X1 h1 (X1 )] F̄i (t ) + F¯∗ (t )
 
t →∞ α1 − 1 cd
for all 1 ≤ k ≤ d − 1 and − E [X1 h1 (X1 )] F¯∗ (t ) − o(1)F¯∗ (t )
αd = (1 + o(1)) E [X1 h1 (X1 )] F̄i (t ) − o(1)F¯∗ (t ).
E Xd |Sd > VaRq [Sd ] ∼ VaRq (Xd ) .
 
(3.14)
αd − 1 Substituting this into (3.20), then substituting (3.18)–(3.20) into
On the contrary, in this particular setting, Theorem 3.1 gives that (3.17) and noticing the arbitrariness of ε , we obtain
3 d
E Xk |Sd > VaRq [Sd ] = o VaRq [Xd ] , 1 ≤ k ≤ d − 1, ∞
    ∫
− −
Ii (t ) ∼ E [X1 h1 (X1 )] F̄i (t ) + xF1 (dx)
while the dth capital allocation is given as in relation (3.14). i =1 i=2 t
The advantage of using Theorem 3.2 over Theorem 3.1 becomes
transparent.
d
− α1 ¯
∼ E [X1 h1 (X1 )] F̄i (t ) + t F1 (t ). (3.21)
α1 − 1
Proof. The proof is provided only for k = 1 as the extensions to all i=2

other values of k are obvious. Similar to the proof of Theorem 2.1, Using a similar argumentation and keeping in mind E [h1 (X1 )]
∫ ∞ = 1, one sees that the denominator on the right-hand side of (3.15)
E [X1 |Sd > t] = x Pr (X1 ∈ dx|Sd > t ) can be approximated as
0−
∞ d
x Pr (Sd − X1 > t − x|X1 = x) F1 (dx) ∞
∫ −
Pr (Sd − X1 > t − x|X1 = x) F1 (dx) ∼ F̄i (t ),
−
= 0 ∞ . (3.15) (3.22)
0−
Pr (Sd − X1 > t − x|X1 = x) F1 (dx) 0− i=1
318 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324

which is equivalent to (3.12). By substituting (3.21) and (3.22) into Proof. Note that both F1 and F2 belong to MDA(Λ) since c2 > 0.
(3.15), we obtain (3.13) and the proof is complete.  Thus, for every s ∈ R,

F¯2 (t + a(t )s)

3.2. Gumbel case lim = c2 e−s .
t →∞ F¯1 (t )
Fix some s small enough such that the right-hand side above is
In the second part of this section, some asymptotic results
larger than 1. Then it holds for arbitrarily fixed y ∈ (x, 1) and for
are obtained for the case in which the risks belong to MDA(Λ)
all large t that
and are unbounded and asymptotically independent. The class
MDA(Λ) essentially contains all distributions with rapidly varying F¯2 (tx) ≥ F¯2 (ty + a(ty)s) > F¯1 (ty).
tails. Moderately heavy-tailed distributions such as lognormal and
Weibull as well as light-tailed distributions such as exponential It follows that VaRF1 (ty) [X2 ] > tx. Analogously to the proof of
and gamma are members of MDA(Λ). For more information, see Lemma 2.2, as q ↑ 1, or, equivalently, as t = VaRq [X1 ] → ∞,
Embrechts et al. (1997). Pr (F2 (X2 ) > q|F1 (X1 ) > q)
It has been seen that asymptotic capital allocations are closely  
Pr X1 ≥ VaRq [X1 ], X2 ≥ VaRq [X2 ]
related to the tail behavior of the aggregate risk. A moderately ≤
Pr X1 > VaRq [X1 ]
 
heavy-tailed distribution from the Gumbel family is also known
as a distribution from the subexponential class S for which two Pr X1 ≥ t , X2 ≥ VaRF1 (ty) [X2 ]
 
independent, identically distributed and nonnegative copies, X1 ≤
and X2 , satisfy F¯1 (t )
Pr(X1 = t ) + Pr (X1 > t , X2 > tx)
Pr(X1 + X2 > t ) ∼ 2 Pr(X1 > t ). ≤ → 0,
F¯1 (t )
For a distribution function F ∈ MDA(Λ), conditions on its auxiliary where in the last step we used the fact Pr (X1 = t ) = o F¯1 (t )
 
function a(·) under which F ∈ S are available in Goldie and due to Corollary 1.6 of Resnick (1987). Therefore, X1 and X2 are
Resnick (1988) and Hashorva et al. (2010). The mainstream study asymptotically independent. 
of the tail probability of the sum of asymptotically dependent
and asymptotically independent random variables has focused on Corollary 2.2 of Mitra and Resnick (2009) shows that, under
the subexponential case. Mitra and Resnick (2009) investigated Assumption 3.3,
the problem from a different perspective. They considered
d
dependent random variables from MDA(Λ) but not necessarily −
Pr (Sd > t ) ∼ F¯1 (t ) ci . (3.26)
subexponential.
i=1
Following the work of Mitra and Resnick (2009) we propose a
set of conditions below: This result aligns with other asymptotic results for the sum of
asymptotically independent subexponential random variables and
Assumption 3.3. Let X be a positive random vector with marginal provides a conspicuous step ahead in understanding the extreme
distributions F1 , . . . , Fd . Assume that F1 ∈ MDA(Λ) with an auxil- behavior of the sum of dependent random variables for the Gumbel
iary function a(·) as defined in (2.9). In addition, for each 1 ≤ i ≤ d, case.
there exists a non-negative constant ci such that F̄i (t )/F¯1 (t ) → ci . The next lemma provides useful information for the proof of
Furthermore, assume that, for all 1 ≤ i ̸= j ≤ d, Theorem 3.3 below.

Pr Xi > t , Xj > a(t )x Lemma 3.4. Let X be a random vector satisfying Assumption 3.3.
 
lim = 0 for all x > 0, (3.23) Then, for every y ∈ R,
t →∞ F¯1 (t )
Pr (Xk > a(t )x, Sd − Xk > t + a(t )y)
and lim
t →∞ F¯1 (t )
Pr Xi > Lij t , Xj > Lij t
  
d
lim = 0 for some Lij > 0. (3.24) −
ci , x ≤ 0,

F¯1 (t ) e− y

t →∞
= i=1,i̸=k
(3.27)
0, x > 0,


If ci > 0 for all 1 ≤ i ≤ d, then Assumption 3.3 indeed describes
an asymptotic independence case. For simplicity, we only verify and
this for d = 2. Note that relation (3.23) with (i, j) = (1, 2) trivially Pr(Xk ≤ a(t )x, Sd > t + a(t )y)
implies relation (3.25) below since a(t ) = o(t ). lim
t →∞ F¯1 (t )

d
Lemma 3.3. Let X = (X1 , X2 ) be a random vector with marginal −
ci , x > 0,

e− y

distributions F1 and F2 . Assume that F1 ∈ MDA(Λ) with an auxiliary = (3.28)
i=1,i̸=k
function a(·), that F¯2 (t )/F¯1 (t ) → c2 for some constant c2 > 0 and
0, x ≤ 0.


that the relation

Pr (X1 > t , X2 > tx) Proof. Our first remark is that relation (3.27) holds for non-
lim =0 (3.25)
t →∞ F¯1 (t ) positive values of x due to (3.26). In addition, the proof of Corollary
2.2 of Mitra and Resnick (2009) shows that, for every x > 0,
holds for some 0 < x < 1. Then X1 and X2 are asymptotically
Pr (Xk > a(t )x, Sd − Xk > t ) = o F̄1 (t ) .
 
independent. (3.29)
 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324 319

Now, relation (2.10) implies that, for arbitrarily fixed 0 < ε < 1 Now, I1 (t ) ≤ a(t ) = o(t ). The change of variables z = t + a(t )s
and all large t, yields that
∞
(1 − ε)a(t ) ≤ a (t + a(t )y) ≤ (1 + ε)a(t ).
∫
I3 (t ) = a(t ) Pr (Xk > t + a(t )s|Sd > t ) ds
Thus, for any x > 0 and y ∈ R, 0
a(t ) ∞
∫
Pr (Xk > a(t )x, Sd − Xk > t + a(t )y) ≤ Pr (Xk > t + a(t )s) ds
Pr (Sd > t ) 0
F¯1 (t )   −1
d ∫ ∞
Pr Xk > a (t + a(t )y) 1+ε , Sd − Xk > t + a(t )y x
  −
∼ ci a( t ) e−s ds
≤
F¯1 (t ) i =1 0

Pr Xk > a (t + a(t )y) 1+ε , Sd − Xk > t + a(t )y −y

 x
 = o(t ),
∼ e → 0,
F¯1 (t + a(t )y) as a result of relations (2.9) and (3.26), the Dominated Convergence
Theorem and the fact that a(t ) = o(t ). It remains to justify
which is a consequence of relations (2.9) and (3.29). Therefore,
(3.27) is proved. I2 (t ) = (Ck + o(1)) t . (3.31)
The proof of the second constituent is facilitated by some vague Some useful bounds for I2 (t ) are as follows:
convergence properties. Let M < y be fixed. Then, by relation
(3.27), (t − a(t )) Pr (Xk > t |Sd > t ) ≤ I2 (t )
   ≤ t Pr (Xk > a(t )|Sd > t ) . (3.32)
Pr
Xk
a(t )
, Sd −a(Xtk)−t ∈· v
→ µk (·) The left-hand side above is equal to
F¯1 (t )
(t − a(t )) Pr (Xk > t |Sd > t )
holds on [−∞, ∞] × [M , ∞] for the measure µk given by
Pr(Xk > t )
= (t − a(t )) = (Ck + o(1)) t ,
d
− Pr(Sd > t )
µk (dx, dy) := ci e−y ϵ0 (dx)dy,
i=1,i̸=k
due to relation (3.26). A similar argumentation and relation (3.28)
help us to find the asymptotic behavior of the right-hand side of
where ϵ0 (·) denotes the Dirac measure. It is useful to note that in (3.32), as
order to fully justify the latter vague convergence property, one
should obtain similar results to (3.27) for other compact sets, which lim Pr (Xk > a(t )|Sd > t ) = Ck .
t →∞
can be simply verified since (3.26) holds.
Note that (3.28) is trivial for non-positive values of x and for Thus, relation (3.31) holds, which completes the proof. 
this reason we assume x > 0. Denote A := {(x1 , x2 ) : x1 ≤ As before, Theorem 3.3 indicates that capital allocations can
x, x1 + x2 > y}. The measure µk puts mass over the set A only be related only to the reference risk measure, VaRq [X1 ]. Finally,
on the line {x1 = 0, x2 > y} and, therefore, µk (∂ A) = 0. Thus, relation (3.26), Lemma 2.4 and the fact that Sd ∈ R−∞ conclude
Proposition A2.12 of Embrechts et al. (1997) allows us to generate the following result:
the conclusion
Pr(Xk ≤ a(t )x, Sd > t + a(t )y) Corollary 3.2. Let X be a random vector satisfying Assumption 3.3.
lim Then, as q = 1 − p ↑ 1,
t →∞ F¯1 (t )
E Xk |Sd > VaRq [Sd ] = (Ck + o(1)) VaR
 
  
d [X1 ].
Pr
Xk
a(t )
, Sd −a(Xtk)−t ∈ A 1−p/
∑
ci
i=1
= lim
t →∞ F¯1 (t )
d The same as in Corollary 2.2, the right-hand side above can be
changed to VaR1−mp [X1 ] for every constant m > 0, but the most
−
= µk (A) = e−y ci .
rational choice for m should be m = 1/
∑d
i=1,i̸=k i=1 ci .
The proof is complete. 
4. Examples
Now, we are able to provide the last main result of this section.
We first provide some examples under which Assumptions 2.1,
Theorem 3.3. Let X be a random vector satisfying Assumption 3.3. 2.2 and 3.1 are satisfied. We start out by giving some information
Then the relation regarding the copula concept (further details can be found in
Nelsen, 1999). It is well known that the dependence structure
E Xk |Sd > VaRq [Sd ] = (Ck + o(1)) VaRq [Sd ]
 
(3.30)
associated with the distribution of a random vector can be
∑
d
 −1 characterized in terms of a copula. A bivariate copula is a
holds for all 1 ≤ k ≤ d, where Ck = ck i=1 ci . two-dimensional distribution function defined on [0, 1]2 with
uniformly distributed marginals. Due to Sklar’s theorem (see Sklar,
Note that the constant Ck above can take value 0 for which 1959), if (X1 , X2 ) has continuous marginal distributions, then there
reason we did not write relation (3.30) as an equivalence. exists a unique copula, C , such that
Proof. Similar to the proof of Theorem 2.2, Pr(X1 ≤ x1 , X2 ≤ x2 ) = C (Pr(X1 ≤ x1 ), Pr(X2 ≤ x2 )) .
a(t ) t ∞
∫ ∫ ∫
Similarly, the survival copula,  C , is defined as the copula
E [Xk |Sd > t] = + + Pr (Xk > z |Sd > t ) dz
0 a(t ) t
corresponding to the joint tail function satisfying
= I1 (t ) + I2 (t ) + I3 (t ). Pr(X1 > x1 , X2 > x2 ) = 
C (Pr(X1 > x1 ), Pr(X2 > x2 )) .
320 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324

Note that every copula satisfies The work of Juri and Wüthrich (2003) justifies the above for 0 <
(u1 + u2 − 1) ∨ 0 ≤ C (u1 , u2 ) ≤ u1 ∧ u2 . β < ∞, while Lemma 4.1 below proves (4.5) for β = 0 and
β = ∞. The same paper shows that Assumptions 2.1, 2.2 and 3.1
Recall that W (u1 , u2 ) := (u1 + u2 − 1)∨ 0 and M (u1 , u2 ) := u1 ∧ u2 are satisfied for finite positive values of β , but one can easily extend
are known as the comonotonic and counter-monotonic copulae, this to the remaining cases. Specifically, if relation (4.2) holds with
which respectively correspond to the strongest and weakest pos- c = cF > 0 and X1 ∈ MDA(Φα ), then
sible dependence structures that may occur between two random
Pr(X1 > tx1 , X2 > tx2 )
variables. The comonotonic (respectively, counter-monotonic) de- lim
pendence structure arises when one random variable is a non- t →∞ Pr(X1 > t )
decreasing (respectively, non-increasing) function of the other.
 −1/β
αβ β αβ
An appealing class of copulae is the Archimedean one. By = x1 + cF x2 , if 0 < β < ∞,

1 ∧ cF x2 ,
x−α if β = ∞,
−α
definition, an Archimedean copula C is given by
C (u1 , u2 ) = ϕ −1 (ϕ(u1 ) + ϕ(u2 )) , and
where ϕ : [0, 1] −→ [0, ∞], called the generator of C (u1 , u2 ), is a Pr(X1 > tx1 , X2 > tx2 )
lim
strictly decreasing and convex function with 0 < ϕ(0) ≤ ∞ and t →∞ Pr(X1 > t )
ϕ(1) = 0. The function ϕ −1 (·) is the pseudo-inverse of ϕ(·), and  −α
by convention ϕ −1 (t ) = 0 if t > ϕ(0). A strict generator satisfies  x1 , if β = 0, x1 > 0, x2 = 0,
ϕ(0) = ∞. = c x−α , if β = 0, x1 = 0, x2 > 0, (4.6)
0F, 2 otherwise.
Juri and Wüthrich (2003) developed a set of sufficient
conditions on the generator of an Archimedean copula under In addition, if relation (4.2) holds with c = cG > 0 and X1 ∈
which the joint concomitant extreme events are characterized. MDA(Λ), then
That is, if
Pr(X1 > t + a(t )x1 , X2 > t + a(t )x2 )
ϕ(1 − xu) lim
lim = xβ , 1 < β < ∞, (4.1) t →∞ Pr(X1 > t )
u↓0 ϕ(1 − u)  −1/β
β x1 β β x2
then = e + cG e , if 0 < β < ∞,

C (ux1 , ux2 ) 1/β e −x 1

∧ cG e−x2 , if β = ∞.
β β
 
lim = x1 + x2 − x1 + x2 . Finally, note that a non-strict generator for a survival Archimedean
u ↓0 u
copula gives (4.6) since it is certain that concomitant extreme
Now, under the assumption that both risks belong to MDA(Φα ) and events are impossible to occur.
have strongly equivalent tail probabilities, A comprehensive list of copulae that satisfy the conditions
Pr(X2 > t ) explained in (4.1) and (4.4) is included in Charpentier and Segers
lim =c (4.2) (2009). The vast majority of Archimedean copulae that satisfy (4.4)
t →∞ Pr(X1 > t )
with β = 1, together with relation (4.2) with c = cF > 0 and
with c = cF > 0, we have X1 ∈ MDA(Φα ), gives (4.6), which designs the framework defined
Pr(X1 > tx1 , X2 > tx2 ) in Assumption 3.1. Additional examples outside the Archimedean
lim world can be found in Asimit and Jones (2008) and Kortschak and
t →∞ Pr(X1 > t )
1/β Albrecher (2009).
β −αβ

−αβ The next lemma develops some asymptotic results that are
= x−α −α
1 + cF x2 − x1 + cF x2 . (4.3)
useful in justifying the extreme behavior of a bivariate random
As expected, if condition (4.1) is satisfied and both risks belong to vector with an Archimedean survival copula.
MDA(Λ) such that relation (4.2) holds with c = cG > 0, then
Lemma 4.1. Let C (·, ·) be a copula such that the survival copula is
Pr(X1 > t + a(t )x1 , X2 > t + a(t )x2 ) Archimedean with a strict generator satisfying (4.4). Then
lim
t →∞ Pr(X1 > t )
C (ux1 , ux2 ) 0, β = 0,

1/β 
β −β x2

−x1 −x 2 −β x1
, lim = (4.7)
=e + cG e − e + cG e u ↓0 u x1 ∧ x2 , β = ∞,
where the auxiliary function a(·) is defined as in (2.9) correspond- holds for any positive x1 and x2 , and
ing to X1 .
Similarly, assume now that the survival copula is Archimedean C (u, f (u))

lim =1 (4.8)
with a strict generator satisfying u ↓0 f (u)
ϕ(xu) is true for any 0 < β ≤ ∞ and any positive measurable function
lim = x−β , 0 ≤ β ≤ ∞, (4.4)
f (u) = o(u).
u ↓0 ϕ(u)
where β = ∞ implies that Proof. We first prove relation (4.7) for β = 0. Recall that (4.4)
holds locally uniformly. Therefore, for any 0 < ε < 1/2, x1 , x2 > 0
ϕ(xu) ∞, x < 1,

lim = and u sufficiently small, we have
u↓0 ϕ(u) 0, x > 1.
ϕ(ux1 ) + ϕ(ux2 ) > 2(1 − ε)ϕ(u).
Now,
Now, ϕ(0) = ∞ implies that ϕ −1 (·) is rapidly varying at ∞ (see
C (ux1 , ux2 )

Proposition 0.8(v), Resnick, 1987). Thus,
lim
u ↓0 u
C (ux1 , ux2 )
 ϕ −1 (ϕ(ux1 ) + ϕ(ux2 ))
0, β = 0,

=
u u

 −1/β
−β −β
= x1 + x2 , 0 < β < ∞, for all x1 , x2 > 0. (4.5) ϕ −1 (2(1 − ε)ϕ(u))
≤ → 0, u ↓ 0.
x1 ∧ x2 , β = ∞.

ϕ −1 (ϕ(u))

 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324 321

The case in which β = ∞ follows in a similar manner. Proof. A non-strict generator excludes joint extreme events
Proposition 0.8(v) of Resnick (1987) implies that ϕ −1 (·) is slowly with probability one, which together with the fact that a(·) is
varying at ∞. As before, we obtain that unbounded, indicates that, for any x1 , x2 , L > 0 and all large t,

C (ux1 , ux2 )
 ϕ −1 (2ϕ (ux1 ∧ ux2 )) Pr (X1 > t , X2 > x2 a(t )) = Pr (X1 > x1 a(t ), X2 > t )
≥ (x1 ∧ x2 ) = Pr (X1 > La(t ), X2 > La(t )) = 0.
u ϕ −1 (ϕ (ux1 ∧ ux2 ))
→ x1 ∧ x2 , u ↓ 0, (4.9) Evidently, conditions (3.23) and (3.24) are satisfied in this case.
It is further assumed that ϕ is a strict generator. We start out
since ϕ(ux1 ) ∨ ϕ(ux2 ) ≤ ϕ (ux1 ∧ ux2 ) holds for all u, x1 , x2 > 0. the verification of condition (3.23) by noticing that
The upper bound is straightforward since C (x1 , x2 ) ≤ M (x1 , x2 ),
Pr(X1 > x1 , X2 > x2 ) = F̄ 1−k (x1 )F̄ 1−l (x2 )
C F̄ k (x1 ), F̄ l (x2 )
 
which concludes the first part of this lemma.
To verify relation (4.8), notice that f (u) = o(u) yields and that
F̄ (xa(t )) , l > k,
 l
C (u, f (u)) u ∧ f (u) C F̄ k (t ), F̄ l (xa(t )) ∼
  
≤ → 1. F̄ k (t ), l ≤ k,

f ( u) f (u)
First assume that 0 < β < ∞. For any ε > 0 and u sufficiently where the latter is a direct implication of relations (4.8) and (4.11).
small, we have ϕ (f (u)/ε) > ϕ(u). Thus, Thus,
F̄ 1−k (t )F̄ (xa(t )) = o F̄ (t ) , l > k,
  
Pr (X1 > t , X2 > xa(t )) ∼
   
C (u, f (u))
 ϕ −1
ϕ f (εu) + ϕ (f (u)) F̄ (t )F̄ 1−l (xa(t )) = o F̄ (t ) , l ≤ k,
≥
f ( u) f ( u) which implies (3.23).
 β −1/β To verify condition (3.24), without loss of generality we assume
→ ε +1 , u ↓ 0,
k ≤ l. If k < l then
due to (4.5). By taking ε ↓ 0, the lower and upper bounds coincide
F̄ l (La(t )) = o F̄ k (La(t )) for any L > 0,
 
in this setting. For β = ∞, the lower bound follows by a similar
reasoning as used in relation (4.9), which concludes (4.8). The proof which, together with (4.8) and (4.11), gives
is complete.  Pr (X1 > La(t ), X2 > La(t ))
= F̄ 2−k−l (La(t )) 
C F̄ k (La(t )) , F̄ l (La(t ))
 
Examples regarding Assumption 3.2 have been already dis-
cussed. Multiple examples regarding Assumption 3.3 have been
provided by Mitra and Resnick (2009). We now indicate a wide ∼ F̄ 2−k (La(t ))
= o F̄ (t ) .
 
class of distributions under which conditions required by Assump-
tion 3.3 are verified. If l = k then
We consider an asymmetric class of copulae studied by
Khoudraji (1995) (see also Genest et al., 1998). If C (·, ·) is an Pr (X1 > La(t ), X2 > La(t ))
= F̄ 2−2k (La(t )) 
C F̄ k (La(t )) , F̄ k (La(t ))
 
Archimedean copula then

Ck,l (u1 , u2 ) := u11−k u21−l C (uk1 , ul2 ), k, l ∈ (0, 1), (4.10) ∼ 2−1/β F̄ 2−k (La(t ))
= o F̄ (t ) ,
 
defines another copula, which we call a transformed asymmetric
Archimedean copula. It can be easily seen that for any dependence where the second step is due to (4.5) and the last step due to (4.11).
structure, the copula Ck,l (·, ·) describes an asymptotically indepen- Thus, the proof is complete. 
dent scenario in the upper tail. Two moderately heavy-tailed distributions satisfy the sufficient
condition defined in (4.11). The first example is the lognormal
Proposition 4.1. Let (X1 , X2 ) be a bivariate non-negative random
distribution with parameters µ ∈ R and σ > 0,
vector whose marginal distributions are identical to F ∈ MDA(Λ)
log x − µ
 
with auxiliary function a(·) as described in (2.9). The survival copula F̄ (x) = F̄SN , x > 0,
is assumed to be given by a transformed asymmetric Archimedean σ
copula Ck,l (·, ·) as defined in (4.10). where FSN (·) denotes the distribution function of the standard
(i) If ϕ is a strict generator satisfying (4.4) with 0 < β ≤ ∞ and normal distribution. The auxiliary function is given by a(x) =
σ 2x
the distribution F is such that log x−µ
(see Embrechts et al., 1997, page 150). The second example
is a distribution function F with a tail
F̄ b (xa(t )) 0, b > c,

γ
e−(log x) ,

lim =
b ≤ c,
(4.11) x > 1,
t →∞ F̄ c (t ) ∞, F̄ (x) =
0, x ≤ 1,
holds for all b, c , x > 0, then conditions (3.23) and (3.24) hold.
where γ > 1. Its auxiliary function is given by a(x) = γ (logxx)γ −1 .
(ii) If ϕ is a non-strict generator, then conditions (3.23) and
(3.24) hold for any distribution function F ∈ MDA(Λ) with For these two examples, the verification of (4.11) is straightforward
a(t ) → ∞. and therefore is omitted.
We cannot draw a conclusion that all distributions from
The assumption of a non-strict generator can be easily re- the intersection MDA(Λ) ∩ S satisfy the requirements of
laxed. Specifically, conditions (3.23) and (3.24) still hold whenever Proposition 4.1. Recall that a Weibull distribution has a tail
Pr(X1 > t0 , X2 > t0 ) = 0 for some t0 > 0 together with the com- τ
F̄ (x) = e−cx , x ≥ 0, c > 0, 0 < τ < 1,
pulsory condition of strongly equivalent tails with an unbounded
auxiliary function. This is the case if the underlying survival with an auxiliary function a(x) = c −1 τ −1 x1−τ (see Embrechts et al.,
copula is given by an Archimedean copula or a transformed 1997, page 150). Hence, it holds for all k, l, x > 0 that
asymmetric Archimedean copula with non-strict generators. The F̄ k (xa(t ))
counter-monotonic dependence structure reflects a similar ex- lim = ∞.
t →∞ F̄ l (t )
treme behavior.
322 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324

Table 1
Capital allocation ratio estimates with α = 2.5.
q Ratio for X1 Ratio for X2
β=2 β=3 β=5 β=2 β=3 β=5
0.99 1.0705 1.0733 1.0748 1.0730 1.0747 1.0751
(0.0059) (0.0045) (0.0045) (0.0053) (0.0049) (0.0042)

0.995 1.0500 1.0536 1.0548 1.0522 1.0543 1.0551

(0.0074) (0.0067) (0.0066) (0.0076) (0.0067) (0.0065)

0.999 1.0231 1.0243 1.0263 1.0261 1.0263 1.0264

(0.0127) (0.0130) (0.0126) (0.0136) (0.0130) (0.0129)

0.9995 1.0195 1.0198 1.0194 1.0210 1.0194 1.0194

(0.0232) (0.0232) (0.0193) (0.0229) (0.0219) (0.0191)

Table 2
Capital allocation ratio estimates with α = 3.
q Ratio for X1 Ratio for X2
β=2 β=3 β=5 β=2 β=3 β=5
0.99 1.0826 1.0879 1.0905 1.0881 1.0904 1.0917
(0.0040) (0.0038) (0.0038) (0.0040) (0.0038) (0.0039)

0.995 1.0618 1.0660 1.0675 1.0650 1.0681 1.0686

(0.0044) (0.0043) (0.0044) (0.0049) (0.0047) (0.0044)

0.999 1.0320 1.0360 1.0357 1.0357 1.0374 1.0365

(0.0086) (0.0108) (0.0094) (0.0087) (0.0102) (0.0097)

0.9995 1.0246 1.0276 1.0274 1.0254 1.0271 1.0280

(0.0135) (0.0138) (0.0130) (0.0143) (0.0134) (0.0134)

5. Simulation study and numerical results values of the parameters α and β , as discussed previously, as well
as four different confidence levels: q = 99%, 99.5%, 99.9% and
We perform a simulation study on the results derived in 99.95%. These results are presented in Tables 1 and 2.
Corollary 2.1. A portfolio of two risks is considered and the The speed of convergence increases, as the strength of depen-
individual loss random variables, X1 and X2 , are Pareto distributed dence relaxes. Heavier tails, i.e., smaller values for α , entail faster
with distribution function convergence of our ratios to 1. As expected, the variances of the
 x −α ratios increase for large values of q, but they are still at reasonable
F (x; λ, α) = 1 − 1 + , x ≥ 0,
λ levels.
where λ equals 100,000 for X1 and 150,000 for X2 , while α is the In the last part of this section, we use our results to evaluate
same for both X1 and X2 and it will be assigned values 2.5 and 3. the capital requirements for a Swiss-based insurance company.
The portfolio dependence structure is assumed to be given by the According to the Swiss Solvency Test (SST) guidelines, the
Gumbel copula capital requirement for an insurance company that operates in
  1/β  Switzerland is given by the CTE-based risk capital corresponding
C (u1 , u2 ) = exp − (− log u1 )β + (− log u2 )β , β ≥ 1. to a 99% level of confidence over a one-year horizon. Unlike the
SST, Solvency II, which designs the regulatory requirements for
This copula belongs to the Archimedean family with a strict insurance firms that operate in the European Union, sets out
generator ϕ(u) = (− log u)β . Clearly, the Gumbel copula has qualitative and quantitative requirements for Solvency Capital that
the asymptotic dependence property for β values greater than 1 ensures an insurance firm to be able to meet its obligations over
since relation (4.1) is satisfied. The measure µ in relation (2.4) is the next 12 months with a probability of at least 99.5%. For both
defined with HF (·) calculated in (4.3) and relation (4.2) holds with SST and Solvency II, the risk-based economic capital is defined by
c = cF = (3/2)α . The parameter β is chosen to be 2, 3 and
the excess of the capital given by the chosen risk measure over the
5. Note that the strength of dependence for the Gumbel copula
best estimate of the liabilities or the expected amount of liabilities
increases as β increases. The asymptotic constants appearing in
under usual circumstances.
Corollary 2.1, C1 and C2 , are numerically computed by using the
Let us assume that a Swiss-based insurance company holds
formulae C1 + C2 = α/(α − 1) and
1 a portfolio of two dependent business lines, as assumed in the
1
α−1
+ µ ((x1 , x2 ) : x1 > z , x1 + x2 > 1) dz beginning of this section. By the Swiss Solvency Test guidelines, the
C1 =
0
. total risk capital requirement is set to CTE0.99 [X1 + X2 ]− E[X1 + X2 ],
µ ((x1 , x2 ) : x1 + x2 > 1)
that is
Each analysis is performed for 100 samples consisting of
5,000,000 simulations from (X1 , X2 ). The ratios between the E [X1 + X2 |X1 + X2 > VaR0.99 [X1 + X2 ]] − E[X1 + X2 ],
capital allocations, estimated from the empirical distribution of the
simulated samples of size 5,000,000 and from the approximation while the individual capital requirements are given by
provided by Corollary 2.1, are calculated for all 100 samples. The α
averages and standard deviations are then tabulated for various E [Xi |Xi > VaR0.99 [Xi ]] − E[Xi ] = VaR0.99 [Xi ], i = 1, 2. (5.1)
α−1
 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324 323

Table 3
Individual capital requirements with varying α and q.
Confidence level Required capital for Required capital for Required capital for Required capital for
(%) X1 (α = 2.5) X2 (α = 2.5) X1 (α = 3) X2 (α = 3)

99 884,929 1,327,393 546,238 819,357

99.5 1,220,922 1,831,383 727,205 1,090,808
99.9 2,474,822 3,712,233 1,350,000 2,025,000
99.95 3,318,799 4,978,198 1,739,882 2,609,822

Table 4
Individual capital requirements with α = 2.5.
Confidence level (%) Required capital for X1 Required capital for X2 Diversification effect (%) β
99 769,290 1,195,565 88.81 2
99.5 1,086,689 1,687,470 90.89 2
99.9 2,271,198 3,523,221 93.65 2
99.95 3,068,469 4,758,830 94.34 2

99 799,973 1,215,167 91.09 3

99.5 1,129,022 1,714,514 93.16 3
99.9 2,357,006 3,578,039 95.93 3
99.95 3,183,540 4,832,343 96.61 3

99 812,321 1,223,545 92.02 5

99.5 1,146,058 1,726,073 94.10 5
99.9 2,391,539 3,601,470 96.86 5
99.95 3,229,849 4,863,765 97.55 5

Table 5
Individual capital requirements with α = 3.
Confidence level (%) Required capital for X1 Required capital for X2 Diversification effect (%) β
99 467,939 725,971 87.43 2
99.5 639,531 991,330 89.71 2
99.9 1,230,061 1,904,559 92.88 2
99.95 1,599,744 2,476,258 93.71 2
99 485,866 737,220 89.56 3
99.5 663,397 1,006,305 91.84 3
99.9 1,274,365 1,932,359 95.01 3
99.95 1,656,843 2,512,087 95.84 3
99 493,058 741,936 90.44 5
99.5 672,971 1,012,584 92.71 5
99.9 1,292,139 1,944,015 95.89 5
99.95 1,679,751 2,527,109 96.72 5

Due to the SST, the insurer should allocate the risk capital for the makes intuitive sense since higher β values imply that the risks are
two business lines as follows: more positively dependent, i.e., closer to be comonotonic. We refer
the reader to Dhaene et al. (2009), who investigated the influence
E [Xi |X1 + X2 > VaR0.99 [X1 + X2 ]] − E[Xi ], i = 1, 2. (5.2)
of the dependence between losses on the diversification benefit
Table 3 elucidates relations (5.1) numerically. All tables herein con- that arises from merging these losses.
sider four different confidence levels, i.e., q = 99%, 99.5%, 99.9%
and 99.95%. The 99% level is recommended by the SST, while the re-
6. Conclusions
maining calculations may help in understanding the effect of more
conservative regulatory requirements.
Further, Tables 4 and 5 elucidate relations (5.2), which are to In this paper we considered the problem of allocating the
this end calculated for each business line for various confidence aggregate risk of a multi-line insurance business consisting of
levels q and values of the parameter β . In this respect, it is well dependent risks to the various sources. We fixed the risk measure
known that risk aggregation should reduce the overall risk of a to be Conditional Tail Expectation and we employed the machinery
multi-line business and thus results in the diversification effect. The of the Extreme Value Theory, in general, and vague convergence, in
phenomenon is conveniently quantified by the ratio particular. The allocation phenomenon is of immense importance
in view of the increasing risk awareness, as well as because
CTE0.99 [X1 + X2 ] − E[X1 + X2 ]
, of the indisputable utility of the, e.g., profitability studies and
CTE0.99 [X1 ] + CTE0.99 [X2 ] − E[X1 + X2 ] quality control in insurance context (see, e.g. Valdez and Chernih,
and it is included in both tables. 2003). The proposed approach to tackle the problem is adequately
It can be seen that there is a significant drop in risk capital motivated by the high confidence levels being required by
requirements as α increases, which is due to the reduction in regulations.
the degree of heavy-tailedness. In addition, the change in capital Our main results, under both asymptotic dependence and
requirements is more pronounced as the confidence level becomes asymptotic independence, reduce the problem to calculating the
less severe. The first business line always requires less risk capital Value-at-Risk of the aggregate risk of the multi-line business
than the second business line as expected. Also, an increase in the of interest. More specifically, for risks having similar extreme
strength of dependence reduces the diversification effect, which behaviors, we showed the following:
324 A.V. Asimit et al. / Insurance: Mathematics and Economics 49 (2011) 310–324

(i) Extremely heavy-tailed risks that are asymptotically depen- Fisher, R.A., Tippett, L.H.C., 1928. Limiting forms of the frequency distribution of
dent in the upper tail require allocations asymptotically the largest or smallest member of a sample. Mathematical Proceedings of the
Cambridge Philosophical Society 24 (2), 180–190.
proportional to the VaR-based aggregate capital, where pro- Furman, E., Landsman, Z., 2005. Risk capital decomposition for a multivariate
portions are sensitive to the strength of the upper tail depen- dependent gamma portfolio. Insurance: Mathematics and Economics 37 (3),
dence and the weight of each individual risk. 635–649.
Furman, E., Landsman, Z., 2006. Tail variance premium with applications for
(ii) Moderately heavy-tailed or light-tailed risks that are asymp- elliptical portfolio of risks. ASTIN Bulletin 36 (2), 433–462.
totically dependent in the upper tail require asymptotically Furman, E., Landsman, Z., 2008. Economic capital allocations for non-negative
equal capital allocations. portfolios of dependent risks. ASTIN Bulletin 38 (2), 601–619.
Furman, E., Zitikis, R., 2008a. Weighted premium calculation principles. Insurance:
(iii) Asymptotic independence in the upper tail requires capital Mathematics and Economics 42 (1), 459–465.
allocations that are asymptotically proportional to the weight Furman, E., Zitikis, R., 2008b. Weighted risk capital allocations. Insurance:
of each individual risk in the portfolio. Mathematics and Economics 43 (2), 263–269.
Geluk, J., Tang, Q., 2009. Asymptotic tail probabilities of sums of dependent
subexponential random variables. Journal of Theoretical Probability 22 (4),
Acknowledgments 871–882.
Genest, C., Ghoudi, K., Rivest, L.P., 1998. Discussions of understanding relationships
using copulas. North American Actuarial Journal 2 (3), 143–149.
The authors would like to thank the anonymous referee Gnedenko, B.V., 1943. Sur la distribution limité du terme maximum d’une série
for his/her careful reading and useful comments. Asimit thanks aléatoaire. Annals of Mathematics 44, 423–453.
Christian Genest for providing the Ph.D. thesis of Khoudraji Goldie, C.M., Resnick, S., 1988. Distributions that are both subexponential and in
the domain of attraction of an extreme-value distribution. Advances in Applied
(1995). Furman acknowledges support of the Natural Sciences Probability 20 (4), 706–718.
and Engineering Research Council of Canada. Tang acknowledges Hashorva, E., Pakes, A.G., Tang, Q., 2010. Asymptotics of random contractions.
support of the 2011 Individual Grant of the Society of Actuaries. Insurance: Mathematics and Economics 47 (3), 405–414.
Joe, H., Li, L., 2011. Tail risk of multivariate regular variation, Methodology and
Computing in Applied Probability, in press (doi:10.1007/s11009-010-9183-x).
References Juri, A., Wüthrich, M.V., 2003. Tail dependence from a distributional point of view.
Extremes 6 (3), 213–246.
Kallenberg, O., 1983. Random Measures, 3rd ed.. Akademie-Verlag, Berlin.
Albrecher, H., Asmussen, S., Kortschak, D., 2006. Tail asymptotics for the sum of two Khoudraji, A., 1995. Contributions à l’étude des copules et à la modélasion des
heavy-tailed dependent risks. Extremes 9 (2), 107–130. valeurs extrêmes bivariées, Ph.D. thesis, Université Laval, Québec, Canada.
Alink, S., Löwe, M., Wüthrich, M.V., 2005. Analysis of the expected shortfall of Klüppelberg, C., Resnick, S.I., 2008. The Pareto copula, aggregation of risks, and the
aggregate dependent risks. ASTIN Bulletin 35 (1), 25–43. emperor’s socks. Journal of Applied Probability 45 (1), 67–84.
Asimit, A.V., Badescu, A.L., 2010. Extremes on the discounted aggregate claims in a Ko, B., Tang, Q., 2008. Sums of dependent nonnegative random variables with
time dependent risk model. Scandinavian Actuarial Journal 2, 93–104. subexponential tails. Journal of Applied Probability 45 (1), 85–94.
Asimit, A.V., Jones, B.L., 2008. Asymptotic tail probabilities for large claims Kortschak, D., Albrecher, H., 2009. Asymptotic results for the sum of dependent
reinsurance of a portfolio of dependent risks. ASTIN Bulletin 38 (1), 147–159. non-identically distributed random variables. Methodology and Computing in
Asmussen, S., Rojas-Nandayapa, L., 2008. Asymptotics of sums of lognormal random Applied Probability 11 (3), 279–306.
variables with Gaussian copula. Statistics and Probability Letters 78 (16), Landsman, Z., Valdez, E., 2003. Tail conditional expectation for elliptical distribu-
2709–2714. tions. North American Actuarial Journal 7 (4), 55–71.
Balkema, A.A., de Haan, L., 1974. Residual life time at great age. The Annals of Li, J., Tang, Q., Wu, R., 2010. Subexponential tails of discounted aggregate claims in
Probability 2 (5), 792–804. a time-dependent renewal risk model. Advances in Applied Probability 42 (4),
Bühlmann, H., 1980. An economic premium principle. ASTIN Bulletin 11 (1), 52–60. 1126–1146.
Cai, J., Li, H., 2005. Conditional tail expectations for multivariate phase-type McNeil, A.J., Frey, R., Embrechts, P., 2005. Quantitative Risk Management. Princeton
distributions. Journal of Applied Probability 42 (3), 810–825. University Press, Princeton, NJ.
Cai, J., Tang, Q., 2004. On max–sum equivalence and convolution closure of heavy- Mitra, A., Resnick, S.I., 2009. Aggregation of rapidly varying risks and asymptotic
tailed distributions and their applications. Journal of Applied Probability 41 (1), independence. Advances in Applied Probability 41 (3), 797–828.
117–130. Nelsen, R.B., 1999. An Introduction to Copulas. Springer-Verlag, New York.
Charpentier, A., Segers, J., 2009. Tails of multivariate Archimedean copulas. Journal Panjer, H.H., Jia, J., 2001. Solvency and capital allocation. Institute of Insurance and
of Multivariate Analysis 100 (7), 1521–1537. Pension Research Report 01–14. Waterloo, Ontario: University of Waterloo.
Chiragiev, A., Landsman, Z., 2007. Multivariate Pareto portfolios: TCE-based capital Resnick, S.I., 1987. Extreme Values, Regular Variation and Point Processes. Springer-
allocation and divided differences. Scandinavian Actuarial Journal 2007 (4), Verlag, New York.
261–280. Resnick, S.I., 2007. Heavy-Tail Phenomena: Probabilistic and Statistical Modeling.
Davis, R.A., Resnick, S.I., 1996. Limit theory for bilinear processes with heavy-tailed Springer-Verlag, New York.
noise. The Annals of Applied Probability 6 (4), 1191–1210. Sklar, A., 1959. Fonctions de répartion à n dimensions et leurs marges, Publications
Dhaene, J., Denuit, M., Vanduffel, S., 2009. Correlation order, merging and de l’Institut de Statistique de l’Université de Paris, 8, p. 229–231.
diversification. Insurance: Mathematics and Economics 45 (3), 325–332. Valdez, E., Chernih, A., 2003. Wang’s capital allocation formula for elliptically
Dhaene, J., Henrard, L., Landsman, Z., Vandendorpe, A., Vanduffel, S., 2008. Some contoured distributions. Insurance: Mathematics and Economics 33 (3),
results on the CTE-based capital allocation rule. Insurance: Mathematics and 517–532.
Economics 42 (2), 855–863. Vernic, R., 2006. Multivariate skew-normal distributions with applications in
Dhaene, J., Tsanakas, A., Valdez, E., Vanduffel, E., 2011. Optimal capital allocation insurance. Insurance: Mathematics and Economics 38 (2), 413–426.
principles. Journal of Risk and Insurance 78. Vernic, R., 2011. Tail conditional expectation for the multivariate Pareto distribution
Dhaene, J., Vanduffel, S., Goovaerts, M.J., Kaas, R., Tang, Q., Vyncke, D., 2006. Risk of the second kind: another approach. Methodology and Computing in Applied
measures and comonotonicity: a review. Stochastic Models 22, 573–606. Probability 13 (1), 121–137.
Embrechts, P., Klüppelberg, C., Mikosch, T., 1997. Modeling Extremal Events for Wang, S.S., 1996. Premium calculation by transforming the layer premium density.
Insurance and Finance. Springer-Verlag, Berlin. ASTIN Bulletin 26, 71–92.