Mathematical and Computer Modelling 58 (2013) 1634–1647

Contents lists available at ScienceDirect

Mathematical and Computer Modelling

journal homepage: www.elsevier.com/locate/mcm

Measuring the capital charge for operational risk of a bank

with the large deviation approach
Zhaoyang Lu ∗
School of Science, Engineering University of Chinese Armed Police Forces, Xi’an, Shaanxi 710086, PR China

article info abstract

Article history: In this paper, the large deviation approach for computing the capital charge for operational
Received 21 November 2010 risk of a bank is explored. Firstly, the negatively-associated structure is utilized to measure
Received in revised form 5 July 2013 the dependence between distinct operational loss cells. Secondly, the lower and upper
Accepted 16 July 2013
bounds of the tail distribution function of total aggregated loss processes are determined.
Keywords:
In addition, first order approximations using a value-at-risk measure are derived. Finally,
Operational risk an important example calculating the capital charge for operational risk under the class of
Negatively associated a heavy-tailed distribution is provided.
Extended regular variation © 2013 Elsevier Ltd. All rights reserved.
Large deviation approach
Multivariate dependence
Value at risk

1. Introduction

Measuring the capital charge for operational risk of a bank has become an important area of research in finance in recent
years (e.g., [1–8]). In Refs. [6,7], the Basel committee on banking supervision defines operational risk as the risk of loss
resulting from inadequate or failed internal processes, people and systems or from external events. It is especially difficult
to measure this new type of risk because operational loss events are extremely rare and produce enormously high losses
and because there are relatively few data available on this topic.
One branch of the literature has focused on developing various kinds of risk models for allocating the regulatory capital
that covers yearly operational risk exposure with a confidence interval of 99.9%. Several advanced measurement methods
have been described, including the internal measurement approach, the scored approach, the loss distribution approach
(LDA), and the Bayesian approach (e.g. [1–5,8–29]).
Broadly speaking, measuring the capital charge of a bank is part of risk management. Performance analysis in the
banking industry has become a key indicator for assessing the quality of a bank management practices. In Ref. [30], data
envelopment analysis (DEA) and neural networks (NNs) were integrated to examine the relative branch efficiency of a large
Canadian bank. These authors also provided guidance on how to improve the branch performance and developed a short-
term efficiency prediction model. In Ref. [31], the authors developed a new DEA model to assess the dual impacts of the
operating and business strategies for the Canadian L&H insurance industry. These authors found that the new DEA model can
simultaneously assess the performance of production and investment. Wu [32] developed a new bi-level programming DEA
approach for optimizing the performance of decentralized companies by using multiple inputs to produce multiple outputs
in a cost-effective manner. In Ref. [33], a predictive scorecard model was introduced to assess account credit worthiness in
large banks. These authors pointed out that the internal scorecard was better able to distinguish the ‘bads’ from the ‘goods’
than the Bureau Score. Descriptions of other recent advances in risk analysis and management can be found in reports by
Wu or Olson [34–43].

∗ Tel.: +86 029 84563509.

E-mail address: [email protected].

0895-7177/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mcm.2013.07.001
 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647 1635

Two definitions for enterprise risk management (ERM) have been developed by COSO and by Wu and Olson [37]. The latter
definition provides the most useful way for firms to manage their risks in the sense of an ‘‘integrated’’ risk management.
The concrete definition is as follows — ‘‘ERM is the integrated process of identification, analysis and either acceptance or
mitigation of uncertainty in investment decision making’’. As pointed out by Valle and Giudici [15], one of the shortcomings
of LDA is its lack of qualitative data (such as expert opinions) to define prior probabilities. Scorecards, which were well
developed by [37], can be used to further study LDA, and the internal risk rating system and the measurement of the
performance of capital charges in multiple risk cells can be developed by validating predictive scorecards.
It is often assumed that the loss frequency process Ni (t ), i = 1, 2, . . . , d for every risk cell is a homogeneous Poisson
process. The assumption of a constant rate is unrealistic in many practical examples. As reported by Ref. [1], the variance
of the monthly frequency series is much higher than the mean rate in some risk cells, suggesting that frequency count data
are over-dispersed. Empirical evidence of the over-dispersion phenomena was also documented by Refs. [5,9]. This problem
can be dealt with by using a nonnegative integer-based process, also known as the loss frequency process.
Gaining an accurate understanding of the degrees of dependence among losses in various units of measurement is critical
to estimating the total capital charge under the LDA. In Ref. [4], it was pointed out that the number of external fraud events
has historically been high (or low) when the number of internal fraud events is high (or low). This inter-dependence of
risk cells has been convincingly shown to be an aggregated loss correlation; this dependence is elegantly illustrated by
the underlying correlation between frequencies but not by correlations in severity in the LDA framework. Other recent
advances in describing the dependence of operational risks can be found in Ref. [9]. Moreover, empirical findings show that
the correlation between two aggregate losses is generally weak (typically below 5%); as a result, a suitable tool for measuring
such a weak dependence is needed. This correlation could be completely dependent, co-monotonic, positively/negatively
dependent, independent, etc.
The remainder of this paper is organized as follows. Section 2 outlines the univariate and multivariate LDA models
used in this paper and explores concepts related to these models. In Section 3, the lower and upper bounds about the
tail distribution function of total aggregated loss processes are provided, and the large deviation results describing first
passage time are explored. In Section 4, first order approximations using a value-at-risk measure are derived. In addition,
an important example calculating the capital charge for operational risk under the class of a heavy-tailed distribution is
provided. Finally in Section 5, the complete proofs of the results are listed.

2. Preliminary results

In this section, we provide information on the univariate and multivariate LDA models used in this paper. It is noteworthy
that the loss frequency process N (t ) is assumed to be a nonnegative integer-based process instead of a homogeneous Poisson
process.

Definition 2.1 (Univariate Loss Distribution Approach (LDA) Model).

1. The severity process: the severities are modeled by a sequence of positive independent and identically distributed (i.i.d.)
random variables {Xk , k ∈ N}. Let F be the distribution function (df ) of Xk with µ = EX1 < ∞.
2. The frequency process: the random variable N (t ) representing losses during the time interval [0, t ] is a nonnegative
integer-based process; we also assume that λ(t ) = EN (t ) < ∞ for any t > 0 and λ(t ) → ∞ as t → ∞.
3. The severity process and the frequency process are assumed to be independent.
N ( t )
4. The aggregated loss process is defined as S (t ) := k=1 Xk .

Definition 2.2 (Extended Regular Variation (ERV) Class). F ∈ ERV (−α, −β) for 1 < α ≤ β < ∞ if F satisfies this condition,
then for any y > 1

F (xy) F (xy)
y−β ≤ lim inf ≤ lim sup ≤ y−α . (2.1)
x→∞ F ( x) x→∞ F (x)
This is also applicable for any v > 1

F (x/v) F (x/v)
v α ≤ lim inf ≤ lim sup ≤ vβ . (2.2)
x→∞ F ( x) x→∞ F (x)

Definition 2.3 (Multivariate LDA Model).

1. Let every ith risk cell be a one-dimensional LDA model with an aggregated loss process Si , severity tail distribution F i ,
severity expectation µi and frequency process Ni (t ) with parameters λi (t ).
2. The dependence between the aggregated operational losses {Si , i = 1, 2, . . . , d} of the Basel cells is modeled by a
negatively-associated (NA) structure. More precisely, for any disjoint subsets A1 , A2 of the set {1, 2, . . . , d}, the following
inequality holds:
Cov(f1 (Si , i ∈ A1 ), f2 (Sj , j ∈ A2 )) ≤ 0, (2.3)
1636 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647

where f1 and f2 are both monotonic non-decreasing (or non-increasing) functions and a defined covariance between
these functions exists.
3. For t ≥ 0, the total aggregated loss process is defined as
S + (t ) := S1 (t ) + S2 (t ) + · · · + Sd (t )
N1 (t ) N2 (t ) Nd (t )
(1) (2) (d)
  
= Xk + Xk + ··· + Xk (2.4)
k=1 k=1 k=1

with the partial sum

n1 n2 nd
(1) (2) (d)
  
Sn+ := Xk + Xk + ··· + Xk , (2.5)
k=1 k=1 k=1

where n := n1 + n2 + · · · + nd .

It is noteworthy that S + (t ) is the sum of the aggregated loss processes in different risk cells, but that the relationships
between the aggregated loss processes in different cells are not necessarily mutually independent.

Definition 2.4 (OpVaR, Total OpVaR). Let Gt be the distribution function of the aggregated loss process S (t ) in a univariate
LDA model. The operational value at risk until time t at level k ∈ (0, 1) is the generalized inverse G←
t of Gt :

VaRt (k) := G←
t (k) = inf{x ∈ R : Gt (x) ≥ k}. (2.6)

The total operational value at risk until time t > 0 at level k ∈ [0, 1) is defined as

t (k) := inf{x ∈ R : Gt (x) ≥ k},

VaR+ +
(2.7)

t (x) is the df of S (t ).
where G+ +

Recall that . means that the quotients on the left and right hand sides of the expression remain bounded. And & denotes
that the quotients on the right and left hand sides of the expression remain bounded.
Furthermore, for some positive number M, the first passage times in discrete and continuous times can be defined as
follows:
inf{n : Sn+ > M }

τd (M ) :=
+∞, if Sn+ ≤ M for n = 1, 2, . . .

and
inf{n : S + (t ) > M }

τc (M ) :=
+∞, if S + (t ) ≤ M for t ≥ 0.

In thispaper, we use the basic ideas described by Refs. [44–46] to derive results related to large deviations. It is noteworthy
that the corresponding results in Ref. [46] are extended to the multivariate case, and that this paper is, to our knowledge,
the first study to implement the large deviation approach in the study of operational risk.
Our results differ significantly from those obtained by Ref. [29]. For example, the results of Ref. [29] are mainly obtained
by means of the empirical study method. We explored the loss data collection exercise for operational risk in Chinese
commercial banks from 1999 to the first half of 2006. For every selected cell, we calibrated two truncated distributions to
fit the loss severity — one for ‘normal’ losses and the other for ‘extreme’ losses. Moreover, the multivariate t copula function
is used to measure the relationship among the selected cells. To some extent, this method lacks a strict mathematical
proof. In contrast, the results from the present study are obtained mainly by means of a strict mathematical proof using
the classical large deviation approach. The inter-dependence of the aggregated operational losses of Basel cells is modeled
with a negatively-associated structure. We also assume that the severity distribution functions are almost in the extended
regular variation (ERV) class. In this framework, the lower and upper bounds about the tail df of the total aggregated loss
process are given. In addition, first order approximations using a value-at-risk measure are derived.

3. Main results related to total OpVaR in the multivariate model

Theorem 3.1. Consider a multivariate LDA model with an NA structure and assume that F1 ∈ ERV (−α, −β), 1 < α ≤ β < ∞
and satisfies

F i (z )
lim = ci , ci ∈ (0, ∞), i = 1, 2, . . . , d.
z →∞ F 1 (z )
 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647 1637

Then,
(i) for any fixed γ > 0, the following lower bound holds uniformly for x ≥ γ
d
i=1 ni ci , i.e.

d

P {Sn+ > x} & F 1 (x) ni ci (1 + o(1)); (3.1)
i=1

µ1 e1/K
 
(ii) for any fixed γ > µ1 , there exist a positive constant K and m = m(K , γ ) such that µ1 e1/K < γ and 1
Km γ
− 1 < −1,
and for x ≥ γ
d
i =1 ni ci the following upper bound holds uniformly:
d

P {Sn+ > x} . F 1 (x)m−β ni ci (1 + o(1)). (3.2)
i=1

Theorem 3.2. Consider a multivariate LDA model with an NA structure and assume that F1 ∈ ERV (−α, −β), 1 < α ≤ β < ∞
and satisfies
F i (z )
lim = ci , ci ∈ (0, ∞), i = 1, 2, . . . , d.
z →∞ F 1 (z )
Also assume that for any δ > 0 and a small ϵ = ϵ(δ) > 0, there exists

E ((Ni (t ))β+ϵ I(Ni (t )>(1+δ)λi (t )) ) = O(λi (t )), i = 1, 2, . . . , d. (3.3)

Then,
(i) for any fixed γ > 0, the following lower bound holds:
d

P {S + (t ) > x} & F 1 (x) λi (t )ci (1 + o(1)) (3.4)
i=1

uniformly for x ≥ γ λi (t )ci ;

d
i=1
µ1 e1/K
 
(ii) for any fixed γ > µ1 , there exist a positive constant K and m = m(K , γ /2) such that µ1 e1/K < γ /2 and 1
Km γ /2
−1 <
−1, the following upper bound holds uniformly for x ≥ γ λi (t )ci , i.e.
d
i =1

d

P {S + (t ) > x} . F 1 (x)m−β λi (t )ci (1 + o(1)). (3.5)
i=1

Theorem 3.3. Consider a multivariate LDA model with an NA structure, and assume that F1 ∈ ERV (−α, −β), 1 < α ≤ β < ∞
and satisfies
F i (z )
lim = ci , ci ∈ (0, ∞), i = 1, 2, . . . , d.
z →∞ F 1 (z )
Then,
(i) for any fixed y > 0 and 0 < x ≤ 1, the following lower bound holds:
1
lim inf log P (τd (M ) ≤ yM x ) & x − β; (3.6)
M →∞ log M
(ii) for any fixed y > 0, and 0 < x ≤ 1, the following upper bound holds:
1
lim sup log P (τd (M ) ≤ yM x ) . x − α. (3.7)
M →∞ log M

Theorem 3.4. Under the conditions of Theorem 3.2,

(i) for any fixed y > 0 and 0 < x ≤ 1, the following lower bound holds:
1
lim inf log P (τc (M ) ≤ yM x ) & x − β; (3.8)
M →∞ log M
(ii) for any fixed y > 0, and 0 < x ≤ 1, the following upper bound holds:
1
lim sup log P (τc (M ) ≤ yM x ) . x − α. (3.9)
M →∞ log M
1638 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647

Corollary 3.1. If F1 ∈ R−α , 1 < α < ∞, then for any fixed y > 0 and 0 < x ≤ 1,

1
lim log P (τd (M ) ≤ yM x ) = x − α. (3.10)
M →∞ log M

Corollary 3.2. If F1 ∈ R−α , 1 < α < ∞, we also assume that for any fixed δ > 0 and sufficiently small ϵ = ϵ(δ) > 0, there
exists
E ((Ni (t ))α+ϵ I(Ni (t )>(1+δ)λi (t )) ) = O(λi (t )), i = 1 , 2 , . . . , d.
Then, for any fixed y > 0 and 0 < x ≤ 1,

1
lim log P (τc (M ) ≤ yM x ) = x − α. (3.11)
M →∞ log M

It is noteworthy that the following condition is always assumed to hold from Theorem 3.3 to Corollary 3.2., i.e.
λ(t )
lim = Λ,
t →∞ t
where Λ is a positive constant. This rather weak requirement can be satisfied by many count processes, including the
renewal process. In Ref. [45], the large deviation results related to the sum of i.i.d. random variables with a subexponential
distribution were derived. In Ref. [44], the results obtained by Cline et al. were extended to the class of an extended regular
variation distribution under weaker conditions. Liu [46] went further in this direction under the weak dependence case of
negatively-associated random sums. To the best of our knowledge, this paper is the first study to use the large deviation
approach to compute the capital charge for operational risk of a bank. In contrast to Ref. [46], we target the finite Basel
loss cells, the dependence structure of which is modeled by a negatively-associated structure. However, consistently with
Ref. [46], the common distribution in every loss cell is also considered under the extended regular variation distribution. First
order approximations using a value-at-risk measure, which are important to the capital charge of a bank and can therefore
enhance the competitiveness of the bank, are also derived.

4. Applications to the capital allocation of operational risk

Theorem 4.1. Under the conditions of Theorem 3.1, and assuming that P {Sn+ > x} is monotonically decreasing for a sufficiently
large x, then for any 0 < k < 1, it can be shown that
   

1−k 1−k 
   
F1← 1 − mβ  . VaR+
n (k) . F1 1 −
←
. (4.1)
 
 d
   d
 
ni ci ni ci
i=1 i=1

Theorem 4.2. Under the conditions of Theorem 3.2, and assuming that P {S + > x} is monotonically decreasing for a sufficiently
large x, then for any 0 < k < 1, it can be shown that
   

1−k 1−k
   
F1← 1 − mβ  . VaR+
t (k) . F1 1 −
←
. (4.2)
  
d d
λi (t )ci λi (t )ci
     
i=1 i=1

Example 4.1 (OpVaR Bound for the Pareto Severity Distribution). In Ref. [9], the authors surveyed a range of correlation and
dependence measures describing operational losses from an international consortium of banks (the Operational Riskdata
Exchange Association) and found evidence of tail dependences among quarterly aggregate loss values of business lines,
event types and the Basel cell units of measure. As illustrated in Fig. 1, a negatively-associated relationship exists among
the quarterly aggregate loss values by business lines (such as clearing, retail brokerage, and asset management). As a result,
we consider a three-dimensional compound Poisson LDA model with intensities λ1 , λ2 , λ3 , and a tail df of severities that is
equivalent to the Pareto df, i.e. for i = 1, 2, 3:
 −α
x
Fi (z ) = 1+ , x > 0, α > 1, θi > 0,
θi
 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647 1639

then
α α
F2 (z ) θ2 θ3
 
c1 = 1, c2 = lim = , c3 = .
z →∞ F1 (z ) θ1 θ1
Without loss of generality, if γ = 4µ1 λ1 t, then K > 1
ln 2
. With K = 2, the most suitable value of m is 0.0875.
According to Theorem 4.2
   
1−k 1−k
F1← 1− (0.0875)α . VaR+ (k) . F1← 1− .
(λ1 + λ2 c2 + λ3 c3 )t (λ1 + λ2 c2 + λ3 c3 )t

5. Proofs of main results

5.1. Proof of Theorem 3.1

Lemma 5.1.1. Assume that X is a nonnegative r.v. with the distribution function F ∈ ERV (−α, −β), 0 < α ≤ β < ∞, then,
for any fixed 0 < α ′ < α ≤ β < β ′ < ∞, EX α < ∞; furthermore, and for all x ≥ x0 (β ′ ):
′

′ ′
k1 x−β ≤ F (x) ≤ k2 x−α ,

where k1 = k1 (β ′ ) and k2 = k2 (β ′ ); these values of k therefore do not depend on x.

This Lemma refers to Lemma 3.1 in Ref. [44].

(i) Proof of (3.1). Because of the NA dependence between S1 , S2 , . . . , Sd , for any fixed a > 1,
 
P (Sn > x) ≥ P max Si > ax
1≤i≤d

d 
≥ P (Si > ax) − P (Si > ax, Sj > ax)
i =1 1≤i<j≤d

d
 
≥ P (Si > ax) − P (Si > ax)P (Sj > ax).
i =1 1≤i<j≤d

According to Ref. [45], for every Fi ∈ ERV (−α, −β), the expression

P (Si > ax) ∼ ni F i (ax) (ni → ∞) (5.1)

holds uniformly for x ≥ γ ni and for every fixed γ > 0. Then,
d
 
P (Sn > x) & ni F i (ax) − ni F i (ax)nj F j (ax)
i=1 1≤i<j≤d

d
 
& ni ci F 1 (ax) − ni ci F 1 (ax)nj cj F 1 (ax)
i=1 1≤i<j≤d

d
 
& F 1 (ax) ni ci − [F 1 (ax)]2 ni ci nj cj
i=1 1≤i<j≤d
 

d
ni ci nj cj 
1≤i<j≤d
 
= F 1 (ax) ni ci 1 − F 1 (ax)
 
d

i=1
  
ni ci
i =1
 
d
 d

≥ F 1 (ax) ni ci 1 − F 1 (ax) ni ci .
i=1 i =1

According to Lemma 5.1.1,

d

lim sup sup F 1 (ax) ni ci = 0. (5.2)
n1 →∞ x≥γ d n c
i=1 i i i=1
1640 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647

If a → 1, then
d

P (Sn > x) & F 1 (x) ni ci (1 + o(1)).
i=1

(ii) Proof of (3.2). For any fixed 0 < m < 1,

 

P (Sn > x) ≤ P Sn > x, (Si > mx) + P (S̃n > x), (5.3)
1≤i≤d

S̃i , S̃i = min{Si , mx}. Taking advantage of the fact that

d
where S̃n = i =1
 
 d
 d

P S n > x, (Si > mx) ≤ P (Si > mx) . ni F i (mx)
1≤i≤d i=1 i=1
d

. F 1 (mx) ni ci
i =1

and F1 ∈ ERV (−α, −β), we can deduce that

 
 d

P S n > x, (Si > mx) . F 1 (x)m−β ni ci . (5.4)
1≤i≤d i =1

For any x ≥ γ i=1 ni ci , if a := an (x) = − log F 1 (x)

d  d 
i=1 ni ci , according to Chebyshev’s inequality for any positive
function c := cn (x), we can conclude that
P (S̃n > x)
d

d
≤ e−cx+a Eec S̃i
F 1 (x) i =1

ni ci
i =1
d 
  mx

≤ e−cx+a 1+ (ecy − 1)dFSi (y) + (ecmx − 1)P (Si > mx)
i =1 0

d  mx

≤ e−cx+a exp (ecy − 1)dFSi (y) + (ecmx − 1)P (Si > mx)
i =1 0
 
d
  mx
. exp −cx + a + ni ci (e − 1)dF1 (y) + (e
cy cmx
− 1)F 1 (mx)
i=1 0

mx/a

d
 
= exp −cx + a + ni ci (ecy − 1)dF1 (y)
i=1 0
 mx
 
+ (e − 1)dF1 (y) + (e
cy cmx
− 1)F 1 (mx)
mx/a
 
d

:= exp −cx + a + ni ci [I1 + I2 ] .
i=1

According to ey − 1 6 yey (y > 0), the following expression can be derived:

 mx/a
I1 := (ecy − 1)dF1 (y) ≤ c µ1 ecmx/a . (5.5)
0
According to the characteristics of the ERV class (which can be found in Ref. [47, Proposition 2.2.1]), the positive constants
x0 , ρ and B exist such that
F 1 (ηx)
≤ Bη−ρ , x ≥ ηx ≥ x0 . (5.6)
F 1 (x)
For any fixed K > 1, if
a − K ρ log a
c := cn (x) = ,
Kmx
then limn1 →∞ supx≥γ d n c cn (x) = 0 and cmx/a ≤ 1/K for a sufficiently large m.

i=1 i i
 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647 1641

Hence, for a sufficiently large m,

 mx
I2 := (ecy − 1)dF1 (y) + (ecmx − 1)F 1 (mx)
mx/a

≤ ecmx F 1 (mx/a)
a − K ρ log a
   −ρ
m
≤ exp B F 1 (x)
K a
 −1/K
 d
= Bm−ρ ni ci F 1 (x) F 1 (x).
i=1
 
Because K > 1 and limn1 →∞ supx≥γ d ni ci F 1 (x) = 0, then,
d
i=1 ni ci
i =1
  1−1/K 
 d
 
lim sup exp Bm−ρ ni ci F 1 (x) = 0. (5.7)
n1 →∞ d
x≥γ i=1 ni ci
 i=1

Then,
P (S̃n > x)
lim sup sup
n1 →∞ x≥γ d d
i=1 ni ci F 1 (x)

ni ci
i=1
  1−1/K 
 d d 
ni ci µ1 ecmx/a + Bm−ρ
 
≤ lim sup −cx + a + c
sup exp ni ci F 1 (x)
n1 →∞ x≥γ d n c
 
i=1 i i i=1 i=1

µ1 e1/K
   
1
≤ exp (a − K ρ log a) −1 +a .
Km γ
µ1 e1/K
 
Because γ > µ1 , there exist a positive constant K and m = m(K , γ ) such that µ1 e1/K < γ and L := 1
Km γ
− 1 < −1;
furthermore, as x ≥ γ
d
i=1 ni ci , inequality (3.2) holds.

5.2. Proof of Theorem 3.2

It suffices that Theorem 3.2 will hold if the following Lemmas are proved.

Lemma 5.2.1. For any v > 0 and x > 0,

 v
d

 e ni µi 

d
x
 i=1
P (Sn > x) . F 1 ni ci +   . (5.8)

v i =1
 x 

Lemma 5.2.2. Suppose that {N (t ); t ≥ 0} is defined as in (3.3), then there exists a positive function ϵ(t ) such that
limt →∞ ϵ(t ) = 0 and
lim P (|N (t ) − λ(t )| > ϵ(t )λ(t )) = 0.
t →∞

This lemma is also described as Lemma 4.4 in Ref. [46]. It is noteworthy that, in the following section, ϵ(t ) is a positive
function as defined in Lemma 5.2.2.
Lemma 5.2.3. For any fixed γ > 0, when x ≥ γ λi (t )ci ,
d
i=1

d  d

P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn+ > x) & F 1 (x) λi (t )ci (1 + o(1)).
i=1 |ni −λi (t )|≤ϵ(t )λi (t ) i =1

γ
Lemma 5.2.4. For any fixed γ > µ1 and m = m K , defined in Theorem 3.1, as x ≥ γ λi (t )ci , the following expression
  d
2 i=1
holds:
d
  d

P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn+ > x) . F 1 (x)m−β λi (t )ci .
i=1 |ni −λi (t )|≤ϵ(t )λi (t ) i =1
1642 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647

Lemma 5.2.5. For any fixed γ > µ1 and m = m(K , γ ) defined in Theorem 3.1, as x ≥ γ λi (t )ci , the following expression
d
i =1
holds:
 
d
  d

P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn > x) = o F 1 (x)
+
λi (t )ci .
i=1 ni <(1−ϵ(t ))λi (t ) i =1

γ
Lemma 5.2.6. For any sufficiently small 0 < δ < 1, γ > µ1 and m = m K , 1+δ
 
defined in Theorem 3.1, as x ≥
γ di=1 λi (t )ci , the following expression holds:

 
d
  d

P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn > x) = o F 1 (x)
+
λi (t )ci .
i=1 (1+ϵ(t ))λi (t )<ni ≤(1+δ)λi (t ) i =1

Lemma 5.2.7. For any sufficiently small 0 < δ < 1, γ > 0, as x ≥ γ λi (t )ci , the following expression holds:
d
i=1
 
d  d
P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn > x) = o F 1 (x)
+
λi (t )ci .
i=1 ni >(1+δ)λi (t ) i=1

Proof of Lemma 5.2.1. For any v > 0 and x > 0, the following expression holds:
d
x
P (Sn > x) . F 1 ni ci + P (S̃n > x), (5.9)
v i=1

where S̃i = min Si , vx , S̃n = i=1 S̃i .

  d
According to Markov’s inequality and the NA characteristic, for any fixed positive function h = h(x),
d

P (S̃n > x) ≤ e−hx EehS̃i .
i=1

Taking advantage of the fact that

x
v
 x
Ee hS̃i
:= eht dFSi (t ) + ehx/v F Si
0 v
x
v
 x
= 1+ (eht − 1)dFSi (t ) + (ehx/v − 1)F Si
0 v
 x 
hx/v v
e −1 x  x 
≤ 1+ tdFSi (t ) + F Si
x/v 0 v v
ehx/v − 1
≤ 1+ ni µi
x/v
ehx/v − 1
 
≤ exp ni µ i
x/v
 
and choosing h = h(x) = vx log d
x
+ 1 > 0, inequality (5.8) holds.
i=1 ni µi

Proof of Lemma 5.2.3. For any fixed γ > 0, according to (3.1), when x ≥ γ λi (t )ci ,
d
i=1

d 
P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn+ > x)
i=1 |ni −λi (t )|≤ϵ(t )λi (t )

d
   
≥ P (N1 (t ) = n1 , . . . , Nd (t ) = nd )P S  >x
(1−ϵ(t )) di=1 λi (t )

i=1 |ni −λi (t )|≤ϵ(t )λi (t )

d
  d

& P (N1 (t ) = n1 , . . . , Nd (t ) = nd ) [(1 − ϵ(t ))λi (t )]ci F 1 (x)
i=1 |ni −λi (t )|≤ϵ(t )λi (t ) i =1
 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647 1643

d
 d

= [(1 − ϵ(t ))λi (t )]ci F 1 (x) P (|Ni (t ) − λi (t )| ≤ ϵ(t )λi (t ))
i=1 i =1
d

= F1 (x) λi (t )ci (1 + o(1)).
i=1

γ
Proof of Lemma 5.2.4. For any fixed γ > µ1 , m = m K ,
 
2
defined in Theorem 3.1, then, according to (3.2), as x ≥
γ λi (t )ci ,
d
i=1

d
 
P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn+ > x)
i=1 |ni −λi (t )|≤ϵ(t )λi (t )

d
   
≤ P (N1 (t ) = n1 , . . . , Nd (t ) = nd )P S d  >x
(1+ϵ(t )) i=1 λi (t )
i=1 |ni −λi (t )|≤ϵ(t )λi (t )

d
  d

. P (N1 (t ) = n1 , . . . , Nd (t ) = nd ) [(1 + ϵ(t ))λi (t )]ci m−β F 1 (x)
i=1 |ni −λi (t )|≤ϵ(t )λi (t ) i=1

d
 d

= [(1 + ϵ(t ))λi (t )]ci m−β F 1 (x) P (|Ni (t ) − λi (t )| ≤ ϵ(t )λi (t ))
i=1 i =1

d

= F 1 (x)m−β λi (t )ci .
i=1

Proof of Lemma 5.2.5. For any fixed γ > µ1 and m = m(K , γ ) defined in Theorem 3.1, then, as x ≥ γ λi (t )ci , by
d
i=1
means of (3.2),
d
 
P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn+ > x)
i=1 ni <(1−ϵ(t ))λi (t )

d
   
≤ P (N1 (t ) = n1 , . . . , Nd (t ) = nd )P S d  >x
(1−ϵ(t )) i=1 λi (t )
i=1 ni <(1−ϵ(t ))λi (t )

d
  d

. P (N1 (t ) = n1 , . . . , Nd (t ) = nd ) [(1 − ϵ(t ))λi (t )]ci m−β F 1 (x)
i=1 ni <(1−ϵ(t ))λi (t ) i =1

 d  d
= [(1 − ϵ(t ))λi (t )]ci m−β F 1 (x) P (Ni (t ) < (1 − ϵ(t ))λi (t ))
i=1 i =1
 
d

= o F 1 (x) λi (t )ci .
i=1

γ
Proof of Lemma 5.2.6. For any sufficiently small 0 < δ < 1, γ > µ1 and m = m K , 1+δ defined in Theorem 3.1, then,
 

according to (3.2), as x ≥ γ i=1 λi (t )ci ,

d

d
 
P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn+ > x)
i=1 (1+ϵ(t ))λi (t )<ni <(1+δ)λi (t )

d
   
≤ P (N1 (t ) = n1 , . . . , Nd (t ) = nd )P S  >x
(1+δ) di=1 λi (t )

i=1 (1+ϵ(t ))λi (t )<ni <(1+δ)λi (t )

d
  d

. P (N1 (t ) = n1 , . . . , Nd (t ) = nd ) [(1 + δ)λi (t )]ci m−β F 1 (x)
i=1 (1+ϵ(t ))λi (t )<ni <(1+δ)λi (t ) i =1
1644 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647

d
 d

= (1 + δ) λi (t )ci m−β F 1 (x) P ((1 + ϵ(t ))λi (t ) < Ni (t ) < (1 + δ)λi (t ))
i =1 i =1
 
d

= o F 1 (x) λi (t )ci .
i =1

Proof of Lemma 5.2.7. For any sufficiently small 0 < δ < 1, γ > 0, as x ≥ γ λi (t )ci , according to Lemma 5.2.1,
d
i=1

d
 
P (N1 (t ) = n1 , N2 (t ) = n2 , . . . , Nd (t ) = nd )P (Sn+ > x)
i=1 ni >(1+δ)λi (t )

d
  d
 x
. P (N1 (t ) = n1 , . . . , Nd (t ) = nd ) ni ci F 1
i=1 ni >(1+δ)λi (t ) i =1
v
 v
d
ni µ i 

d
  e
i=1
+ P (N1 (t ) = n1 , . . . , Nd (t ) = nd ) 
 

i=1 ni >(1+δ)λi (t )
 x 

d
 x
. E (Ni (t )I(Ni (t )>(1+δ)λi (t )) )ci F 1
i=1
v
 v
d
ni µ i 

d
  e
i=1
+ P (N1 (t ) = n1 , . . . , Nd (t ) = nd )   =: I1 + I2 .
 
i=1 ni >(1+δ)λi (t )
 x 

If v = β + ϵ such that 1 < β + ϵ ≤ 2, we can deduce that

d
(n1 µ1 + n2 µ2 + · · · + nd µd )β+ϵ ≤ (β + ϵ) (ni µi )β+ϵ .

(5.10)
i =1

Hence,
 
 e β+ϵ
β+ϵ β+ϵ
 
I2 ≤ (β + ϵ) P (N1 (t ) = n1 )(n1 µ1 ) + ··· + P (Nd (t ) = nd )(nd µd ) .
x ni >(1+δ)λi (t ) nd >(1+δ)λd (t )

By means of Lemma 5.1.1, we can derive that F 1 (x) ≥ k1 x−(β+ε/2) . Then,

I2
lim sup sup
t →∞ x≥γ d d
λi (t )ci
λi (t )ci F 1 (x)

i=1

i=1
d
eβ+ϵ x−(β+ϵ) (β + ϵ) (µi )β+ϵ E ((Ni (t ))β+ϵ I(Ni (t )>(1+δ)λi (t )) )

i =1
≤ lim sup sup
d
t →∞ x≥γ d λ (t )c
λi (t )ci k1 x−(β+ε/2)

i=1 i i

i =1
= 0.
Moreover,
  according to the equivalent definition (2.2) of an ERV class, we can deduce that, for a sufficiently large x,
F1 x
β+ϵ
≤ 2(β + ϵ)β F 1 (x). Hence,

I1
lim sup sup
d
= 0. (5.11)
t →∞ x≥γ d λ (t )c
λi (t )ci F 1 (x)

i=1 i i

i=1

This completes the Proof of Lemma 5.2.7, and therefore concludes the Proof of Theorem 3.2.
 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647 1645

5.3. Proof of Theorem 3.3

(i) For any fixed y > 0, 0 < x ≤ 1 and any 0 < θ ≤ 1, when M is sufficiently large, according to Theorem 3.1, the following
inequality holds:
d

P (τd (M ) ≤ yM x ) ≥ P (S[yM x ] > M ) ≥ (1 − θ ) [yi Mix ]ci F 1 (M ).
i=1

According to Lemma 5.1.1, for any β < β ′ < ∞, the following expression holds:
 
d
1 1 
−β ′
lim inf log P (τd (M ) ≤ yM ) ≥ lim inf
x
log (1 − θ ) yi Mix
[ ]ci k1 M
M →∞ log M M →∞ log M i=1

= x − β ′.
If β ′ → β , then inequality (3.6) can be obtained.
(ii) For any fixed y > 0 , 0 < x ≤ 1 and any fixed γ > µ1 , there exist a positive constant K and m = m(K , γ ) such that
µ1 e1/K
µ1 e1/K < γ and 1
Km γ
− 1 < −1. For a sufficiently large M, the following inequality holds:
d

P (τd (M ) ≤ yM x ) ≤ P (S[yM x ] > M ) ≤ (1 + θ ) [yi Mix ]ci m−β F 1 (M ).
i=1

According to Lemma 5.1.1, for any 0 < α ′ < α , the following inequality holds:
 
d
1 1 
−α ′
lim sup log P (τd (M ) ≤ yM ) ≤ lim supx
log (1 + θ ) yi Mix
[ ]ci m −β
k2 M
M →∞ log M M →∞ log M i =1

≤ x − α′ .
If α ′ → α , inequality (3.7) can be derived.

5.4. Proof of Theorem 3.4

According to Theorem 3.2, the inequalities (3.8) and (3.9) can be derived with the method used for the Proof of
Theorem 3.3.

5.5. Proof of Theorem 4.1

According to (3.1), for every k (0 < k < 1), the total OpVaR is bounded by

n (k) = inf{x ∈ R : P (Sn ≤ x) ≥ k}

VaR+ +

= inf{x ∈ R : P (Sn+ > x) ≤ 1 − k}

 
 d
. inf x ∈ R : ni ci F 1 (x) ≤ 1 − k
i=1
 

 

 
1−k
 
= inf x ∈ R : F1 (x) ≥ 1 − d
  
ni ci 

 
 
i=1
 

1−k
 
= F1← 1 − d .
 
  
ni ci
i =1

On the other hand, according to (3.2), for every k (0 < k < 1), the total OpVaR is bounded by

n (k) = inf{x ∈ R : P (Sn ≤ x) ≥ k}

VaR+ +
 
d

& inf x∈R: ni ci m −β
F 1 ( x) ≤ 1 − k
i=1
1646 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647
 

 

 
1−k
 
= inf x ∈ R : F1 (x) ≥ 1 − mβ d
  
ni ci 

 
 
i=1
 

1−k
 
= F1← 1 − mβ d .
 
  
ni ci
i=1

5.6. Proof of Theorem 4.2

According to Theorem 3.2, the total OpVaR can be bounded with the method used in the Proof of Theorem 4.1.

6. Concluding remarks

VaR can be calculated under the assumption of certain kinds of weak dependences in a number of ways. Almost all of the
traditional approaches to VaR computations (the delta method, historical simulation, Monte Carlo simulation and stress-
testing) were comprehensively compared in Refs. [48,13]. In addition, the application of stable processes within sample and
forecast evaluations was recommended for VaR estimation and the satisfactory evaluation of possible losses.
An interesting outcome of this research would be the introduction of the method of risk analysis into the discussion of
operational risk management which was thoroughly evaluated by Wu and Olson [37,38]. The four stages of risk analysis
are risk identification, risk analysis, risk mitigation, and risk monitoring (i.e. the risk management process). Within the
context of Basel III, there are 56 kinds of operational losses, and the classification of loss data is strongly associated with risk
identification. In the second stage of risk analysis, various risk evaluation models from econometrics, finance, and statistics
can be further developed. In the third stage of risk analysis, the advanced measurement tools of operational risk (including
value-at-risk, expected shortfall, etc.) can be applied to reduce the capital charge, which is required to cover the yearly
operational risk exposure of a bank within a confidence interval of 99.9%. The parameters of a more realistic scenario (such
as the operational losses in the 2008 financial crisis) can be used to strengthen the stability of risk models. In the final stage
(risk monitoring), prior information provided by experts and validated predictive scorecards can be used to supervise the
banking system.

Acknowledgments

We would like to thank the anonymous referee for his/her comments on the previous version of this manuscript. This
research was supported in part by the National Natural Science Foundation of China under (Grant No. 11201001), the Science
Research Grant of Shaanxi province under (Grant No. 2011JM1019) and the Xi’an 2013 Annual Social Science Planning Fund
under (Grant No. 13IN14).

References

[1] A. Chapelle, Y. Crama, G. Hübner, J.-P. Peters, Practical methods for measuring and managing operational risk in the financial sector: a clinical study,
Journal of Banking and Finance 32 (2008) 1049–1061.
[2] A. Frachot, P. Georges, T. Roncalli, Loss distribution approach for operational risk, Groupe de Recherche Opérationnelle, Cré dit Lyonnais, Working
Paper, 2001. URL: http://ssrn.com/abstract=1032523.
[3] A. Frachot, P. Georges, T. Roncalli, Loss distribution approach in practice, Groupe de Recherche Opérationnelle, Cré dit Lyonnais, Working Paper, 2003.
URL: http://gro.creditlyonnais.fr/content/wp/lda-practice.pdf.
[4] A. Frachot, T. Roncalli, E. Salomon, The correlation problem in operational risk, Groupe de Recherche Opérationnelle, Crédit Lyonnais, Working Paper,
2004.
[5] A. Resheter, Dependence of Operational Losses and the Capital at Risk, Swiss Banking Institute, University of Zurich, 2008, URL:
http://www.gloriamundi.org.
[6] Basel Committee on Banking Supervision, Basell II: The New Basel Capital Accord, Second Consultantive Paper. Report to the Bank for International
Settlements, 2001. URL: http://www.bis.org.
[7] Basel Committee on Banking Supervision, Basell II: The New Basel Capital Accord, Third Consultantive Paper, Report to the Bank for International
Settlements, 2003. URL: http://www.bis.org.
[8] C. Alexander, Operational Risk. Regulation, Analysis and Management, Financial Times Prentice Hall, London, 2003.
[9] Eric Cope, G. Antonini, Obeserved correlations and dependencies among operational losses in the ORX consortium database, The Journal of Operational
Risk 3 (2008) 47–74.
[10] F. Biagini, S. Ulmer, Asymptotics for operational Risk quantified with expected shortfall, ASTIN Bulletin 39 (2009) 735–752.
[11] H.H. Panjer, Operational Risk: Modeling Analytics (I), first ed., John Wiley & Sons, 2006.
[12] J. Havlický, Quantification of operational risk loss within the loss distributional approach, in: Euro Working Group on Financial Modelling, Stockholm,
2008.
[13] J.D.C. Semper, J.M.T. Beltrán, Sector concentration risk: a model for estimating capital requirements, Mathematical and Computer Modelling 54 (2011)
1765–1772.
 Z. Lu / Mathematical and Computer Modelling 58 (2013) 1634–1647 1647

[14] K. Böcker, C. Kluppelbeg, Multivariate models for operational risk, Quantitative Finance 10 (2010) 855–869.
[15] L. Dalla Valle, P. Giudici, A Bayesian approach to estimate the marginal loss distributions in operational risk management, Computational Statistics &
Data Analysis 52 (2008) 3107–3127.
[16] M. Degen, P. Embrechts, D.D. Lambrigger, The quantitative modeling of operational risk: between g-and-h and EVT, Astin Bulletin 37 (2) (2007)
265–291.
[17] M. Haubenstock, L. Hardin, The loss distribution approach, in: C. Alexander (Ed.), Operational Risk. Regulation, Analysis and Management, Financial
Times Prentice Hall, London, 2003.
[18] Nan Li, D. Wu, Using text mining and sentiment analysis for online forums hotspot detection and forecast, Decision Support Systems 48 (2) (2010)
354–368.
[19] R. Nash, The three pillars of operational risk, in: Operational Risk, Regulation, Analysis and Management, Financial Times Prentice Hall, London, 2003.
[20] J.M. Netter, A.B. Poulsen, Operational risk in financial service providers and the proposed basel capital accord: an overview, Working Paper, Terry
College of Business, Department of Banking and Finance, University of Georgia, 2003.
[21] P. Embrechts, G. Puccetti, Aggregating risk capital, with an application to operational risk, Geneva Risk Insurance Review 31 (2006) 71–90.
[22] P. Embrechts, G. Puccetti, Bounds for functions of dependent risks, Finance and Stochastic 10 (2008) 341–352.
[23] P. Embrechts, H. Furrer, R. Kaufmann, Quantifying regulatory capital for operational risk, Working Paper, RiskLab, ETH Zürich, 2003.
[24] P. Giudici, A. Bilotta, Modelling operational losses: a Bayesian approach, Quality and Reliability Engineering International 20 (5) (2004) 407–417.
[25] R. Kuhn, P. Neu, Functional correlation approah to operational risk in banking orginations, Physica A 322 (2003) 650–666.
[26] S. Bachelier, Operational Risk: Some Practical Questions, RiskLab-madrid, Universidad Autonoma de Madrid, 2007, URL:
http://www.crest.fr/ckfinder/userfiles/files/pageperso/relie.
[27] U. Anders, The path to operational risk economic capital, in: C. Alexander (Ed.), Operational Risk. Regulation, Analysis and Management, Financial
Times Prentice Hall, London, 2003.
[28] V. Chavez-Demoulin, P. Embrechts, J. Neŝlehová, Quantitative models for operational risk: extremes, dependence and aggregation, Journal of Banking
& Finance 30 (2006) 2635–2658.
[29] Z. Lu, Modeling the yearly value-at-risk for operational risk in Chinese commercial banks, Mathematics and Computers in Simulation 82 (4) (2011)
604–616.
[30] D. Wu, Z. Yang, L. Liang, Using DEA-neural network approach to evaluate branch efficiency of a large Canadian bank, Expert Systems with Applications
31 (1) (2006) 108–115.
[31] D. Wu, Z. Yang, S. Vela, L. Liang, Simultaneous analysis of production and investment performance of Canadian life and health insurance companies
using data envelopment analysis, Computers and Operations Research 34 (1) (2007) 180–198.
[32] D. Wu, BiLevel programming data envelopment analysis with constrained resource, European Journal of Operational Research 207 (2) (2010) 856–864.
[33] D. Wu, D.L. Olson, Enterprise risk management: a DEA VaR approach in vendor selection, International Journal of Production Research 48 (16) (2010)
4919–4932.
[34] D.L. Olson, D. Wu, Earthquakes and risk management in China, Human and Ecological Risk Assessment 16 (3) (2010) 478–493.
[35] D.L. Olson, D. Wu, Simulation of fuzzy multi-attribute models for grey relationships, European Journal of Operational Research 175 (1) (2006) 111–120.
[36] D. Wu, Chi-Guhn Lee, Stochastic DEA with ordinal data applied to a muti-attribute pricing problem, European Journal of Operational Research 207 (3)
(2010) 1679–1688.
[37] D. Wu, D.L. Olson, Enterprise risk management: coping with model risk in a large bank, Journal of the Operational Research Society 61 (2010) 179–190.
[38] D. Wu, D.L. Olson, Enterprise risk management: small business scorecard analysis, Production Planning & Control 120 (4) (2009) 362–369.
[39] D. Wu, D.L. Olson, Supply chain risk, simulation, and vendor selection, International Journal of Production Economics 114 (2) (2008) 646–655.
[40] D. Wu, D.L. Olson, Introduction to the special issue on ‘‘Optimizing risk management: methods and tools’’, Human and Ecological Risk Assessment 15
(2) (2009) 220–226.
[41] D. Wu, John R. Birge, Serial chain merger evaluation model and application to mortgage banking, Decision Sciences 43 (1) (2012)
http://dx.doi.org/10.1111/j.1540-5915.2011.00340.x.
[42] D. Wu, K. Xie, G. Chen, et al., Modeling technological innovation risks of an entrepreneurial team using system dynamics: an agent-based perspective,
Technology Forecasting and Social Change 77 (6) (2010) 857–869.
[43] D. Wu, K. Xie, G. Chen, et al., A risk analysis model in concurrent engineering product development, Risk Analysis 30 (9) (2010) 1440–1453.
[44] C. Su, Q. Tang, T. Jiang, A contribution to large deviations for heavy-tailed random sums, Science in China (Series A) 44 (4) (2001) 438–444.
[45] D.B.H. Cline, T. Hsing, Large deviation probabilities for sums and maxima of random variables with heavy or subexponential tails, Texas A&M
University, 1991, Preprint.
[46] Y. Liu, Large deviations, risk theory and their applications in finance and insurance, Ph.D. Dissertation, Wuhan University Publishing House, 2004
(in Chinese).
[47] N.H. Bingham, C.M. Godie, J.M. Teugels, Regular Variation, Cambridge University Press, 1987.
[48] I. Khindanova, S. Rachev, E. Schwartz, Stable modeling of value at risk, Mathematical and Computer Modelling 34 (2001) 1223–1259.