Você está na página 1de 78

Expected

Risk-Adjusted Return
for
Insurance Based Models

Tatiana Solcà

Diploma thesis in mathematics at


ETH Zürich under the supervision of
Prof. Dr. P. Embrechts and Dr. U. Schmock

Spring 2000
Ai miei genitori:
per avermi dato la possibilita’
di raggiungere questo traguardo.
Acknowledgements

It is my pleasure to thank Prof. Paul Embrechts for giving me the opportunity


to work in this very interesting field.

I am very grateful to Dr. Uwe Schmock for his helpful suggestions and sup-
port. He always managed to find time to discuss problems related to my
work.

Next I would like to thank Francesco for his encouragement, constant support
and patience.

Let me finally thank all my friends and schoolmates for their advice and all-
round help during this work. In particular, I thank Nicola for his pleasant and
constructive “pineletters” and “talks” from Paris, Giacomo and Andrea for
answering my tedious questions concerning LATEX and Carlo for the pleasant
time spent working in the computer room.

Hereby, I express my sincere gratitude to all these people.

v
Contents

Acknowledgements v

1 Introduction 1

2 Preliminaries 3
2.1 Risk and Risk Measures . . . . . . . . . . . . . . . . . . . . . 3
2.1.1 Notation and properties . . . . . . . . . . . . . . . . . 3
2.1.2 The tail conditional expectation . . . . . . . . . . . . . 4
2.2 Convergence of random variables . . . . . . . . . . . . . . . . 5
2.3 Stationary process and ergodic theorem . . . . . . . . . . . . . 6
2.4 Properties of weak convergence . . . . . . . . . . . . . . . . . 8
2.5 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5.1 The idea and the notation . . . . . . . . . . . . . . . . 9
2.5.2 The different variants of the model . . . . . . . . . . . 11

3 Expected shortfall 13
3.1 The asymptotical limit of upper bounds . . . . . . . . . . . . 15
3.2 The n-dimensional model . . . . . . . . . . . . . . . . . . . . . 24
3.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4 Standard deviation 33
4.1 The simplest cases: n = 1 and n = 2 . . . . . . . . . . . . . . 34
4.2 The general case . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 The second variant of the model . . . . . . . . . . . . . . . . . 43
4.4 The third variant of the model . . . . . . . . . . . . . . . . . . 49

5 Capital allocation 53
5.1 Covariance principle . . . . . . . . . . . . . . . . . . . . . . . 53
5.2 Expected-shortfall principle . . . . . . . . . . . . . . . . . . . 60

A Calculating expected shortfall 63

B The Lagrange multipliers rule 67


vii
viii CONTENTS
Chapter 1

Introduction

An insurance company, nowadays, takes an interest in the applications of


actuarial techniques for measuring risk and for assessing profitable areas of
business. Its management is faced continually with the task of reconciling
the conflicting interests of policyholders and shareholders. The former are
interested in strong financial strength, while the latter are more concerned
with a return on equity that is commensurate with the risk inherent in their
investment. In order to satisfy these needs, the management selects profitable
business and limits the company’s risk.
Obviously, an insurance company’s liabilities cannot be entirely foreseen.
If companies are to maintain a high degree of financial security, they must
efficiently manage asset and liability portfolios, as well as understand and
keep control of the underlying risks.
One difficulty is that an insurance company faces many different types of
risk, and they are not at all easy to model. For a thorough understanding of
them it is essential to have quantitative models, but even though there are
many different measures of risk, up to today, none of them can be considered
the “best”.
In this paper, we focus on the kinds of risks which can be represented
by random variables. In particular, we analyze a model denoting the risk
portfolio of an insurance company. We suppose that the management steers
the company by choosing the numbers of independent risks so as to improve
the overall expected risk-adjusted return, defined as the expected return
divided by the assigned risk capital.
In other words, with respect to the defined model, we try to determine if
an optimal portfolio exists. For these purposes, we organized this paper in
four parts.
In Chapter 2, we first develop some approaches to measuring risk by con-
sidering the statement of axioms on a risk measure done by Artzner et al.,

1
2 CHAPTER 1. INTRODUCTION

which leads to the concept of coherent risk measure. Moreover, we briefly


introduce some basic concepts of probability theory (like convergence of ran-
dom variables and ergodic theorem), which will be seen to be useful for our
purpose.
Then, in Section 2.5, we give a description of the model considered in this
whole work and we indicate which methods will be used to measure risk. In
more detail, we suppose that the whole profit R of an insurance company
consisting of n units can be denoted by the sum of the stochastic gains Ri
of every unit i ∈ {1, . . . , n}. Ri is defined as the revenues minus the costs
and the losses in the form illustrated in (2.3). Then we will analyze this
model using the expected shortfall risk measure, the coherent risk measure
suggested by Artzner et al. (1998) and, later, by means of the standard
deviation risk measure, which is very popular in practice.
In Chapter 3 we discuss the model using the expected shortfall for quan-
tifying risk. We will estimate the performance of the company examining
the expected risk-adjusted return r, i.e., r = E[R]/ρ(R), where ρ denotes
the risk measure. In fact, the company’s aim will be to invest its resources
optimally and maximize r. We therefore will try to determine an optimal
portfolio by choosing the values of N1 , . . . , Nn , which denote the number of
contracts of the respective business units 1, . . . , n, such that a maximum for
r is attained. In particular, we focus on a proposition which shows the exis-
tence of the limit of the upper bounds for the expected risk-adjusted return
r.
In Chapter 4 we repeat the same approaches using the standard deviation
risk measure. In this case we concentrate on the optimization problem defined
in (4.2), i.e., we will try to determine the optimal number of contracts of every
unit i ∈ {1, . . . , n} in order to maximize E[R]/C subject to the constraint
ρ(R) ≤ C, where C denotes the capital the company wants to invest. We
will examine three different variants of the model defined in Section 2.5 and
we will show that this optimization problem has a solution for every variant.
In Chapter 5 we consider two different capital allocation principles, name-
ly the covariance principle and the expected shortfall principle. We will
calculate the covariance principle for ̺(R) = −E[R] + κ σ(R) with κ > 0,
and for two variants of the model R considered in Chapter 4. Moreover, for
simplicity we consider a company consisting only of two units, but the same
results can be computed in a similar way for the general case, too, and we
calculate E[Ri | R ≤ c] for the multivariate normal case.
In the Appendix we briefly recall some useful technical rules to calculate
the expected shortfall.
Chapter 2

Preliminaries

2.1 Risk and Risk Measures


It is not easy to define risk, and we will avoid attempting to give an exact
definition. Nevertheless in a recent paper, Artzner et al. (1998) have come
up with an appropriate description of what risk actually is. In this paper,
we consider risk related to the variability of the future value of a position
due to uncertain events. Therefore, we treat those kinds of risks which can
be represented by random variables, and which indicate the possible future
values of positions.

2.1.1 Notation and properties


Let Ω be the set of possible states of nature, and assume it is finite. By
a random variable X we denote the final net worth of a position for each
element of Ω. Let G be the set of all risk, i.e., the set of all real valued
functions on Ω. Remark that G can be identified with Rn , where n = card(Ω).

Definition 2.1. A measure of risk is a mapping ρ from G into R.

Then the real number ρ(X) can be interpreted, when positive, as the
minimum extra cash to add to the risky position X, or when negative, as the
cash amount that can be subtracted from the position.
We now consider some properties for a risk measure ρ defined on G listed
in the form of axioms.

Axiom (Translation invariance).


For all X ∈ G and all real numbers α: ρ(X + αr) = ρ(X) − α,
where r is the rate of return on a reference riskless investment.

3
4 CHAPTER 2. PRELIMINARIES

Axiom (Subadditivity).
For all X1 , X2 ∈ G: ρ(X1 + X2 ) ≤ ρ(X1 ) + ρ(X2 ).
Axiom (Positive homogeneity).
For all λ ≥ 0 and all X ∈ G: ρ(λX) = λρ(X).
Axiom (Monotonicity).
For all X and Y ∈ G with X ≤ Y : ρ(Y ) ≤ ρ(X).
Axiom (Relevance).
For all X ∈ G with X ≤ 0 and X = 0: ρ(X) ≥ 0.

Remarks.
• Translation invariance means that adding (resp. subtracting) the sure
initial amount α to (from) the initial position, and investing it in the
reference instrument (with rate of return r) the risk measure only de-
creases (resp. increases) by α.
• Subadditivity reflects the diversification of portfolios and ensures that
the risk measure behaves reasonably when adding two positions; we
can say: “a merger does not create extra risk”.

These axioms on measures of risk are related to the axioms on acceptance


sets, but we won’t treat this topic (for more details see Artzner et al. (1998)).
We are interested in the following definition then we argue that any risk
measure which is to be used to effectively regulate or manage risks satisfies
these axioms.
Definition 2.2. A risk measure satisfying the four axioms of translation
invariance, subadditivity, positive homogeneity and monotonicity is called
coherent.
In their paper Artzner et al. suggest a specific coherent measure called
tail conditional expectation and in the following chapter we will study the
model of a portfolio using this risk measure.

2.1.2 The “tail conditional expectation” measure of


risk
In practice there are various methods of measuring risk, and these axioms
are not restrictive enough to specify a unique risk measure. The choice of
precisely which measure to use should be made on the basis of additional
considerations.
2.2. CONVERGENCE OF RANDOM VARIABLES 5

In this work we consider tail conditional expectation (expected shortfall),


which, under some assumptions, is the least expensive among those which
are coherent and accepted by regulators 1 since they are more conservative
than the value-at-risk measurement. Managers and regulators are primarily
interested in setting “minimal requirements” or a maximal limit on the po-
tential losses. With a shortfall approach, one can answer the question “how
bad is bad?” by measuring the negative of the average future net worth X
of a position, given that X is below the quantile c ≤ 0, i.e.,

ρ(X) = E[−X | X ≤ c ] , (2.1)

provided that P (X ≤ c) > 0.


In the following sections we will focus on definitions and theorems, which
will be useful later for our purposes.

2.2 Convergence of random variables


Definition 2.3. Let X1 , X2 , . . . and X be real-valued random variables on
some probability space (Ω, F, P ). We say:

i) {Xn }n∈N converges to X almost surely (a.s.) if

P ({ω | lim Xn (ω) = X(ω)}) = 1,


n→∞

r
ii) {Xn }n∈N converges to X in r-th mean (−→) for r > 0 if
n→∞
E | Xnr | < ∞ for all n and E( | Xn − X | r ) −−→ 0,

P
iii) {Xn }n∈N converges to X in probability (−→) if for every ε > 0
n→∞
P ( | Xn − X | > ε) −−→ 0 .

D
iv) {Xn }n∈N converges to X in distribution (−→) if
n→∞
P (Xn ≤ x) −−→ P (X ≤ x)

for all points x at which FX (x) = P (X ≤ x) is continuous.


1
A regulator is a supervisor who takes into account the unfavorable states when allowing
a risky position.
6 CHAPTER 2. PRELIMINARIES

Remark. The following implications hold in general:


a.s. P D
i) Xn −→ X =⇒ Xn −→ X =⇒ Xn −→ X ,
r P D
ii) Xn −→ X =⇒ Xn −→ X =⇒ Xn −→ X.

We now recall the theorem of bounded convergence.


Theorem 2.4 (Lebesgue bounded convergence theorem). Consider a
a.s.
sequence {Xn }n∈N of variables with Xn −→ X. If there is a random variable
Y such that E | Y | < ∞ and | Xn | ≤ Y for all n, then
n→∞
E[Xn ] −−→ E[X] .

Proof. See Grimmet and Stirzaker (1992), Chapter 5.6. 


Remark. It is appropriate to specify that by the convergence in r-th mean
the values r = 1 and r = 2 are of most use. Therefore, in these cases, we
write, respectively
1
i) Xn −→ X in mean, instead of Xn −→ X,
2
ii) Xn −→ X in mean square, instead of Xn −→ X.

2.3 Stationary process and ergodic theorem


Definition 2.5. A real-valued process X1 , X2 , . . . is called stationary if for
every x1 , . . . , xn ∈ R and integer k > 0

P [X1 ≤ x1 , . . . , Xn ≤ xn ] = P [X1+k ≤ x1 , . . . , Xn+k ≤ xn ] .

Consider a probability space (Ω, F, P ) and a transformation T : Ω → Ω.


Definition 2.6. A transformation T : Ω → Ω will be called measurable if
for all A ∈ F: T −1 (A) = {ω ∈ Ω | T (ω) ∈ A} ∈ F.
Definition 2.7. A measurable transformation T : Ω → Ω will be called
measure-preserving if for all A ∈ F: P (T −1 (A)) = P (A).
Let T be a measure-preserving transformation on (Ω, F, P ).
Definition 2.8. A set A ∈ F is called invariant if T −1 (A) = A.
We denote by J the set of all invariant A ∈ F. Note that J is a σ-field.
2.3. STATIONARY PROCESS AND ERGODIC THEOREM 7

Definition 2.9. T is called ergodic if P (A) ∈ {0, 1} for every A ∈ J .

Theorem 2.10 (Ergodic theorem). Let T be a measure-preserving trans-


formation on (Ω, F, P ). Then for any random variable X such that E|X| <
∞:
n−1
1
lim X(T k (ω)) = E[X | J ] a.s. and in mean.
n→∞ n
k=0

Proof. See Breiman (1968), pp. 113–115. 

Corollary 2.11. Let T be a measure-preserving and ergodic transformation


on (Ω, F, P ). Then for any random variable X such that E|X| < ∞:

n−1
1
lim X(T k (ω)) = E[X] a.s. and in mean.
n→∞ n
k=0

Proof. See Breiman (1968), p. 115. 

These general definitions and results concerning invariance and ergodicity


can be applied to the original stationary process X1 , X2 , . . . by considering a
shift transformation T , i.e., if x = (x0 , x1 , . . . ) is a real sequence of values of
the stationary process then T x = (x1 , x2 , . . . ). For more details see Grimmet
and Stirzaker (1992). The corresponding form of the ergodic theorem for
stationary processes is:

Theorem 2.12 (Ergodic theorem for stationary processes). Let X1 ,


X2 , . . . be a stationary process such that E|X1 | < ∞, then
n
1
lim Xk = E[X|J ] a.s. and in mean.
n→∞ n
k=1

Proof. See Grimmet and Stirzaker (1992), Chapter 9.5. 

Furthermore, we define an ergodic stationary process and we write down


the corresponding version of the ergodic theorem.

Definition 2.13. Let T be the shift-operator and A a set of real sequences.


A stationary process is said to be ergodic if P {(X0 , X1 , . . . ) ∈ A} = 0 or 1,
whenever A is shift-invariant, i.e., A is invariant with respect to the shift-
operator T .
8 CHAPTER 2. PRELIMINARIES

Remark. Let {Xn }n∈N0 be a real-valued stationary process. Then the follow-
ing conditions are equivalent:

(a) {Xn }n∈N0 is ergodic,

(b) P {(X0 , X1 , . . . ) ∈ A} = 0 or 1, for every invariant set A,



(c) limn→∞ n1 nj=1 φ(Xj , Xj+1 , . . . ) = E[φ(X0 , X1 , . . . )], for every measur-
able function φ of real sequences, provided the expectation exists,

(d) limn→∞ n1 nj=1 φ(Xj , . . . , Xj+k ) = E[φ(X0 , . . . , Xk )], for every k ∈ N0
and every measurable function φ of k + 1 variables, provided the ex-
pectation exists.

For more details see Karlin and Taylor (1975), Chapter 9.5.
So, since a stationary process X1 , X2 , . . . is ergodic if every shift-invariant
event has probability zero or one, if J has this zero-one property, of course
the average converges to E[X1 ]. Hence, we restrict ourselves to:

Theorem 2.14 (Ergodic theorem for ergodic stationary processes).


If X1 , X2 , . . . is a stationary ergodic process such that E|X1 | < ∞, then
n
1
lim Xk = E[X1 ] a.s. and in mean.
n→∞ n
k=1

Proof. See Grimmet and Stirzaker (1992), Chapter 9.5. 

2.4 Properties of weak convergence


Let S be a metric space and S the Borel σ-field in S, i.e., S is the smallest
σ-field containing all the open sets. Let Cb (S) be the set of all bounded
continuous real functions f on S, and P be a probability measure on S, i.e.,
a nonnegative, countably additive set function with P (S) = 1.

Definition 2.15. Let Pn for n ∈ N and P be probability measures on (S, S).


w
We say that {Pn }n∈N converges weakly to P (Pn −→ P ) if for all f ∈ Cb (S):
 
f dPn −→ f dP .
S S

Definition 2.16. A set A in S, whose boundary ∂A satisfies P (∂A) = 0, is


called a P-continuity set.
2.5. THE MODEL 9

Theorem 2.17 (Portmanteau Theorem). Let Pn for n ∈ N and P be


probability measures on (S, S). These five conditions are equal:
w
i) Pn −→ P ,
 
ii) lim supn→∞ S f dPn = S f dP , for all bounded uniformly continuous
functions f : S → R,

iii) lim supn→∞ Pn (F ) ≤ P (F ), for all closed F ⊂ S,

iv) lim inf n→∞ Pn (G) ≥ P (G), for all open G ⊂ S,

v) limn→∞ Pn (A) = P (A), for all P -continuity sets A.

Proof. See Billingsley (1968), pp. 11–14. 

2.5 The model


2.5.1 The idea and the notation
The aim of this work is to study a model for the portfolio of an insurance
company by means of two different risk measures which take the whole profit
of the company into consideration. We investigate some possible risk port-
folios, which are directly dependent on the number of the contracts to be
stipulated and on the expected losses, so as to determine, if possible, the
optimal number of contracts in order to maximize the profit. Hence, we are
going to assume that an insurance company consists of n organizational units
or business segments. The whole profit of the company (positive value means
gain, negative value means loss) will be denoted by
n

R= Ri , (2.2)
i=1

with Ri describing the stochastic gain of the unit i during a fixed time period
(usually one year). Moreover, we suppose that for every unit i ∈ {1, . . . , n}
Ni

Ri = νi Ni − Xi,j − Yi Ni (2.3)
j=1

where:

• νi is the premium income for one contract of unit i,


10 CHAPTER 2. PRELIMINARIES

• Ni is the number of contracts of unit i, N1 + · · · + Nn is the whole


number of contracts of the insurance company,
• {Xi,j }j∈N is a sequence of random variables for i ∈ {1, . . . , n}: Xi,j
represents
Ni the loss associated with the risk in the j-th contract of unit
i and j=1 Xi,j is the total (annual) claim amount of unit i,
• Yi is a random variable which represents, for one contract of unit i,
a safety loading needed to obviate both the possible approximation
error in the calculation of the optimal premium and unforeseeable,
catastrophic events.
Therefore, we will examine how to determine an optimal portfolio, which
guarantees a maximal profit. We will discuss this problem for two risk mea-
sures: the expected shortfall, the coherent risk measure suggested by Artzner
et al. (1998), and the standard deviation, which is very popular in practice.
We will analyze the portfolio by means of a risk adjusted performance mea-
surement, this means that we compute the return in the way commonly called
RORAC (return on risk adjusted capital). In particular, considering any risk
measure ρ and provided that ρ(R) = 0, we define the expected risk-adjusted
return for a risk R as
E[R]
r(R, ρ) = . (2.4)
ρ(R)
In practice, the company tries hard to improve its results and this means that
its aim is to maximize r. Therefore, we will try to determine the optimal
value for the numbers of contracts Ni of every unit i ∈ {1, . . . , n}, such that
this maximum is attained.
We now introduce our running examples of risk measure. We start con-
sidering the expected shortfall defined by
ρ(R) = E[−R | R ≤ c] with c ≤ 0 . (2.5)
The expected shortfall is an alternative risk measure to the quantile which
overcomes some of the theorical deficiencies of the latter. In particular, this
risk measure gives some information about the size of potential losses, given
that a loss bigger than c has occurred.
Then, the same consideration will be repeated using the standard devia-
tion risk measure defined by
ρ(R) = −E[R] + κ σ(R), (2.6)
where κ > 0 is some positiveconstant and σ(R) denotes the standard devi-
ation operator, i.e., σ(R) = Var(R).
2.5. THE MODEL 11

2.5.2 The different variants of the model


In Chapter 4 we will examine the expected risk-adjusted return for a risk
R considering the standard deviation risk measure. In particular, we will
study three different variants of the model representing the whole profit of a
company consisting of n units, previously defined as
n 
 Ni
 
R= νi Ni − Xi,j − Yi Ni . (2.7)
i=1 j=1

Moreover, we make the following assumptions, which are valid through all of
Chapter 4.
• {Xi,j }j∈N , for all i ∈ {1, . . . , n}, are sequences of independent iden-
tically distributed (i.i.d.) random variables, with Xi,j having a finite
mean denoted by µi , i.e,
µi = E[Xi,j ] for all i ∈ {1, . . . , n} and j ∈ N,

• Y1 , . . . , Yn are random variables having finite mean denoted by µ̃i , i.e,


µ̃i = E[Yi ] for all i ∈ {1, . . . , n},

• all the sequences {Xi,j }j∈N , with i ∈ {1, . . . , n}, and the random vari-
ables Y1 , . . . , Yn are independent.
Remark. In Chapter 3, we will examine the above-defined model with the
aid of the expected shortfall as risk measure, but for the moment we do
not need such strong assumptions. In fact it is not necessary to require
that {Xi,j }j∈N are sequences of i.i.d. random variables, but rather we will
prove some statements for which it is enough to assume the existence of real
constants µ1 , . . . , µn such that, for all i ∈ {1, . . . , n},
N
1  N →∞
Xi,j −−−→ µi in mean.
N j=1

Therefore, for a model of the form of (2.7) with the above-mentioned


assumptions, there are different variants depending on the choice of the dis-
tributions of the random variables and on the nature of the parameters Ni
for i ∈ {1, . . . , n}.
Then, in Chapter 4 we will examine three different cases which result
from (2.7) if we choose Ni for i ∈ {1, . . . , n} in different ways. In particular,
we consider the Ni ’s first as positive integers, then as Poisson-distributed
random variables and, finally, as the sum of both.
12 CHAPTER 2. PRELIMINARIES

More precisely, these possibilities can be represented as follows:


  i

1. R = ni=1 νi Ni − N j=1 Xi,j − Yi Ni ,


with Ni a positive integer for all i ∈ {1, . . . , n},
  i

2. R = ni=1 νi Ni − N j=1 Xi,j − Yi Ni ,


with Ni ∼ POIS(λi ), λi > 0 for all i ∈ {1, . . . , n},
  i

3. R = ni=1 νi Ni − N j=1 Xi,j − Y i N i ,


with Ni = Ni + Ni , where Ni are positive integers and Nipois are
fix pois fix

Poisson-distributed random variables, i.e., Nipois ∼ POIS(λi ), λi > 0,


for all i ∈ {1, . . . , n}.
Chapter 3

Results for the expected


shortfall risk measure

In this chapter, we will examine the portfolio of a company, analyze the profit
represented by the model previously described, and consider the expected
shortfall as an aid to quantifying risk.
Without loss of generality and, in order to simplify the following cal-
culations, we start by examining a one-dimensional model. This means we
assume that an insurance company consists of only one business unit. Hence,
we denote the whole profit by

N

R(N ) = νN − Xj − Y N, (3.1)
j=1

where X1 , X2 , . . . is a real-valued process which represents the claim sizes;


Y is a random variable having finite mean µ̃ and N is some positive integer.
Recall that it is assumed that Y and the sequence {Xj }j∈N are independent,
and that we still do not require any particular properties for the process
{Xj }j∈N . Moreover, we assume ν − µ − µ̃ > 0, i.e., the company chooses a
premium income rate greater than the expected losses.
In order to estimate the performance of the company, we consider the
expected risk-adjusted return for the risk R(N ) which, in this case, can be
represented by

E R(N )
rN = , c ≤ 0, (3.2)
E − R(N ) | R(N ) ≤ c

provided that P R(N ) < c > 0.

13
14 CHAPTER 3. EXPECTED SHORTFALL

More generally, in Section 3.2 we will take the same approach considering
an n-dimensional model
n 
 Ni
 
R(N1 , . . . , Nn ) = νi Ni − Xi,j − Yi Ni (3.3)
i=1 j=1

with all the assumptions listed in Section 2.5.1.

Now, we begin with the following lemma which is valid in general for any
risk represented by a random variable R ∈ L1 (Ω, F, P ).
Lemma 3.1. Let R be an integrable random variable on a probability space
(Ω, F, P ), this means R ∈ L1 (Ω, F, P ), and let c0 denote the infimum of the
support of the distribution of R, i.e., c0 = inf {c ∈ R | P (R ≤ c) > 0}.
Then the map
(c0 , ∞) ∋ c −→ E[−R | R ≤ c]
is non-increasing.
Proof. First, we consider E[−R | R ≤ c]. This conditional expectation is
defined by
E − R 1{R≤c}
E[−R | R ≤ c] = .
P [R ≤ c]
Then we consider any constants c1 , c2 ∈ (c0 , ∞) such that c1 < c2 .
Given that, since {R ≤ c1 } and {c1 < R ≤ c2 } are disjoint, it holds that

1{R≤c2 } = 1{R≤c1 } + 1{c1 <R≤c2 } ,


we can write:
E[−R | R ≤ c2 ]
E[−R | R ≤ c1 ] P [R ≤ c1 ] + E[−R | c1 < R ≤ c2 ] P [c1 < R ≤ c2 ]
= .
P [R ≤ c2 ]
Then, in order to prove the monotony, it suffices to show that

E[−R | R ≤ c2 ] ≤ E[−R | R ≤ c1 ] .

We split the proof into two cases.


i) If P (c1 < R ≤ c2 ) = 0, then it holds that 1{R≤c2 } = 1{R≤c1 } P -a.s.
and P (R ≤ c1 ) = P (R ≤ c2 ). Therefore, it follows that
E[−R 1{R≤c2 } ] E[−R 1{R≤c1 } ]
E[−R | R ≤ c2 ] = = = E[−R | R ≤ c1 ] .
P [R ≤ c2 ] P [R ≤ c1 ]
3.1. THE ASYMPTOTICAL LIMIT OF UPPER BOUNDS 15

ii) If P (c1 < R ≤ c2 ) > 0, then we can write


E[−R | R ≤ c2 ]
E[−R | R ≤ c1 ] P [R ≤ c1 ] + E[−R | c1 < R ≤ c2 ] P [c1 < R ≤ c2 ]
=
P [R ≤ c2 ]

E[−R | R ≤ c1 ] P [R ≤ c2 ] − P [c1 < R ≤ c2 ]


=
P [R ≤ c2 ]
E[−R | c1 < R ≤ c2 ] P [c1 < R ≤ c2 ]
+
P [R ≤ c2 ]
= E[−R | R ≤ c1 ]

P [c1 < R ≤ c2 ]
+ E[−R | c1 < R ≤ c2 ] − E[−R | R ≤ c1 ]
    P [R ≤ c2 ]
−c2 ≤···<−c1 ≥−c1  
  0<···≤1
<0
< E[−R | R ≤ c1 ] .

Thus, E[−R | R ≤ c2 ] ≤ E[−R | R ≤ c1 ], as desired. 

3.1 The asymptotical limit of upper bounds


In this section we focus on the portfolio of a company represented by the
model defined in (3.1). In particular, we will study the expected risk-adjusted
return rN for the risk R(N ). The main question is whether it is possible
to find a maximal value for rN thereby determining an optimal value for
N such that this maximum is attained. We will show that the limit of
the upper bounds for the expected risk-adjusted return exists under quite
general conditions. We also derive an explicit formula for the limit. Recall
that, considering the expected shortfall risk measure, rN is defined by (3.2).
Now, for any c ≤ 0 we can write
E[R(N )]
E R(N )
rN = = R(N ) NR(N ) .
E − R(N ) | R(N ) ≤ c E − N | N ≤ Nc
Therefore, if we define
N
1 
RN = ν − Xj − Y, (3.4)
N j=1

since E[R(N )] = N (ν − µ − µ̃), it holds that


ν − µ − µ̃
rN = c
. (3.5)
E − RN | RN ≤ N
16 CHAPTER 3. EXPECTED SHORTFALL

In particular, the aim of this chapter is to prove the following proposition.

Proposition 3.2. Let X1 , X2 , . . . be a real-valued


N process and Y be a inte-
1
grable random variable. Let RN = ν − N j=1 Xj − Y and assume that a
real constant µ exists such that

N
1  N →∞
Xj −−−→ µ in mean .
N j=1

Moreover, suppose that P (ν − µ − Y = 0) = 0 . Then, it holds that

N →∞
E[RN | RN ≤ 0] −−−→ E[ν − µ − Y | ν − µ − Y ≤ 0] .

For the proof of this proposition we need the next lemma.

Lemma 3.3. Let RN be defined as above and assume that a real constant
µ exists such that

N
1  N →∞
Xj −−−→ µ in mean.
N j=1

Let Z be an integrable random variable and f be a bounded uniformly Lip-


schitz continuous function. Then, it holds that

N →∞
E Z f (RN ) −−−→ E Z f (ν − µ − Y ) .

Proof. Let ε > 0. Since Z ∈ L1 , because of the theorem of Lebesgue, a real


constant K > 0 exists such that


E | Z | 1{ | Z | >K} ≤ ε .

Moreover, f is Lipschitz continuous, i.e., a real constant α exists such that

| f (RN ) − f (ν − µ − Y ) | ≤ α | RN − ν − µ − Y | .
3.1. THE ASYMPTOTICAL LIMIT OF UPPER BOUNDS 17

Then,

E | Z f (RN ) − Z f (ν − µ − Y ) |

≤ E | Z | | f (RN ) − f (ν − µ − Y ) |

≤ E | Z | 1{ | Z | >K} | f (RN ) − f (ν − µ − Y ) |
 
≤ 2 sup|f |

+ E | Z | 1{ | Z | ≤K} | f (RN ) − f (ν − µ − Y ) |
 
≤K

≤ E | Z | 1{ | Z | >K} 2 sup|f | + K E | f (RN ) − f (ν − µ − Y ) |
   
≤ε ≤α |RN −(ν−µ−Y )|

N →∞
≤ 2 ε sup |f | + K α E | RN − ν − µ − Y | −−−→ 2 ε sup |f |
 
1 N
=| N j=1 Xj −µ|

Therefore, since ε is arbitrary, we can conclude that, as desired,


N →∞
E | Z f (RN ) − Z f (ν − µ − Y ) | −−−→ 0 .


Proof (Proposition 3.2). We remember that E[RN | RN ≤ 0] is defined by



E RN 1{RN ≤0}
E[RN | RN ≤ 0] = . (3.6)
P [RN ≤ 0]

We first consider E[RN 1{RN ≤0} ]. Clearly, it holds that


 N  
1 
E RN 1{RN ≤0} = E ν − Xj − Y 1{RN ≤0} (3.7)
N j=1
  N  
1
= νP [RN ≤ 0] − E Xj 1{RN ≤0} − E Y 1{RN ≤0} .
N j=1

Since
N →∞
RN −−−→ ν − µ − Y in mean ,
and because the convergence in mean implicates the convergence in distrib-
ution, it holds that
N →∞
P (RN ≤ 0) −−−→ P (ν − µ − Y ≤ 0) .
18 CHAPTER 3. EXPECTED SHORTFALL

Therefore, since it is assumed that P (ν − µ − Y = 0) = 0, we can prove, on


the one hand, that
 N  
1 N →∞
E Xj 1{RN ≤0} −−−→ µ P (ν − µ − Y ≤ 0) (3.8)
N j=1

and, on the other, that


N →∞
E Y 1{RN ≤0} −−−→ E[Y | ν − µ − Y ≤ 0] P (ν − µ − Y ≤ 0) . (3.9)

From   N 
1  N →∞
E  Xj − µ −−−→ 0
N j=1

it follows that
  N     N   
 1   1 
E  Xj 1{RN ≤0} − µ 1{RN ≤0}  = E  Xj − µ 1{RN ≤0}
N j=1 N j=1  
≤1
  N 
1  N →∞
≤E  Xi − µ −−−→ 0 .
N j=1

Thus, it still remains to be calculated

lim E[µ 1{RN ≤0} ] and lim E[Y 1{RN ≤0} ] .


N →∞ N →∞

Let Z be any integrable random variable. We will determine

lim E[Z 1{RN ≤0} ]


N →∞

and then we will employ the result for Z = µ and Z = Y . As shown in


Figure 3.1, let fn and gn bounded continuous functions defined by

1 ,
 if x ∈ (−∞, − n1 ] ,
fn (x) = −nx , if x ∈ (− n1 , 0) ,


0, if x ∈ [0, ∞) ,
and 
1 ,
 if x ∈ (−∞, 0] ,
gn (x) = 1 − nx , if x ∈ (0, n1 ) ,

if x ∈ [ n1 , ∞) .

0,
3.1. THE ASYMPTOTICAL LIMIT OF UPPER BOUNDS 19

fn (x) gn (x)

1 1

x 1 x
− n1 n

Figure 3.1: Continuous approximation of the indicator function 1(−∞,0] from be-
low by fn and from above by gn .

Instead of 1{RN ≤0} it is useful to write 1(−∞,0] ◦ RN , then it follows that

1(−∞,− 1 ] ◦ RN ≤ fn (RN ) ≤ 1(−∞,0] ◦ RN ≤ gn (RN ) ≤ 1(−∞, 1 ] ◦ RN


n n

and clearly for Z ≥ 0 it holds that



E Z fn (RN ) ≤ E Z 1(−∞,0] ◦ RN ≤ E Z gn (RN ) . (3.10)

For a general integrable random variable Z, consider the decomposition Z =


Z + − Z − with Z + = max{Z, 0} and Z − = max{−Z, 0}.
Then, considering the left and right side respectively of the inequality
(3.10) separately, since fn and gn are Lipschitz continuous functions, i.e.,

| fn (RN ) − fn (ν − µ − Y ) | ≤ n | RN − ν − µ − Y |

and analogously

| gn (RN ) − gn (ν − µ − Y ) | ≤ n | RN − ν − µ − Y | ,

it follows from Lemma 3.3 that


N →∞
E Z fn (RN ) −−−→ E Z fn (ν − µ − Y )
and N →∞
E Z gn (RN ) −−−→ E Z gn (ν − µ − Y ) .
n→∞
Moreover, since fn (ν − µ − Y ) −−→ 1(−∞,0) ◦ (ν − µ − Y ) pointwise, from the
Lebesgue theorem 2.4 it follows that
n→∞
E Z fn (ν − µ − Y ) −−→ E Z 1(−∞,0) ◦ (ν − µ − Y )
20 CHAPTER 3. EXPECTED SHORTFALL

and analogously
n→∞
E Z gn (ν − µ − Y ) −−→ E Z 1(−∞,0] ◦ (ν − µ − Y ) .

Now, by the assumption P [ν − µ − Y = 0] = 0,



E Z 1(−∞,0) ◦ (ν − µ − Y ) = E Z 1(−∞,0] ◦ (ν − µ − Y )

and consequently we obtain the statements (3.8) and (3.9) replacing Z = µ


and Z = Y , respectively, i.e.,
N →∞

E µ1{RN ≤0} −−−→ E µ1{ν−µ−Y ≤0} = µ P ν − µ − Y ≤ 0 ,

and
N →∞
E Y 1{RN ≤0} −−−→ E Y 1{ν−µ−Y ≤0}
= E[Y | ν − µ − Y ≤ 0] P (ν − µ − Y ≤ 0) .

Thus, to conclude,
N →∞

E[RN 1{RN ≤0} ] −−−→ ν − µ − E[Y | ν − µ − Y ≤ 0] P (ν − µ − Y ≤ 0)

and given that this last quantity equals

E[ν − µ − Y | ν − µ − Y ≤ 0] P (ν − µ − Y ≤ 0) ,

the statement in the proposition follows directly from the definition of the
conditional expectation. 

Consequently, from Lemma 3.1 and from Proposition 3.2 we obtain the
limit of the upper bounds for the expected risk-adjusted return rN . In fact,
we also have
N →∞
rN −−−→ r∞ ,
N →∞
as RN −−−→ ν − µ − Y in mean and in law and hence
c
N →∞

P RN ≤ −−−→ P ν − µ − Y ≤ 0
N
and
N →∞
E RN 1RN ≤c/N −−−→ E (ν − µ − Y ) 1ν−µ−Y ≤0 .
3.1. THE ASYMPTOTICAL LIMIT OF UPPER BOUNDS 21

Remark. We obtain N the same result if we assume a model of the form of


R(N ) = νN − j=1 Xj − Y N such that X1 , X2 , . . . is a real stationary
ergodic process having finite mean µ = E[X1 ], and P (ν − µ − Y = 0) = 0.
In fact, as a result of the Ergodic Theorem 2.14, it holds that
N
1  N →∞
Xj −−−→ µ a.s. and in mean.
N j=1

Therefore, the requirements of the Proposition 3.2 are satisfied and it follows
that, if N → ∞, the expected risk-adjusted return for R(N ) converges to
r∞ .

We would like to illustrate Proposition 3.2 by concentrating on normally


distributed random variables and by giving a numerical example.
Example 1. Assume that an insurance company consists of one business
unit, i.e., n = 1. Then we consider
N

R(N ) = νN − Xj − Y N.
j=1

Suppose that {Xj }j∈N is an i.i.d. sequence of random variables having normal
distribution and that Y is a normally distributed random variable indepen-
dent of the sequence {Xj }j∈N . In particular, we write

Xj ∼ N µ, σ 2 for all j ∈ N ,

Y ∼ N µ̃, σ̃ 2 .
Because of the particular properties of the normal distribution it holds that
N  σ2 
1 
Xj ∼ N µ,
N j=1 N
and therefore
N
1   σ2 
RN = ν − Xj − Y ∼ N ν − µ − µ̃, + σ̃ 2 .
N j=1 N

For more details see Johnson et al. (1994), Chapter 13, Section 3.
Based on RN , we can calculate the expected risk-adjusted return rN .
Given c ≤ 0, define
c/N − E[RN ] c/N − ν + µ + µ̃
cN =  =  .
Var(RN ) σ 2 /N + σ̃ 2
22 CHAPTER 3. EXPECTED SHORTFALL

Using (A.3) from the Appendix, we obtain



E R(N )
rN =
E − R(N ) | R(N ) ≤ c
E[RN ]
=
E[−RN | RN ≤ c/N ]
ν − µ − µ̃
= 
′ .
−ν + µ + µ̃ + σ 2 /N + σ̃ 2 log Φ (cN )

Note that the last expression is mathematically meaningful even if N is not


an integer. If c = 0, the above expression simplifies to

−cN
rN =
′ .
cN + log Φ (cN )

For every c ≤ 0 we get

−ν + µ + µ̃
c∞ := lim cN = ,
N →∞ σ̃
hence
ν − µ − µ̃
r∞ = lim rN =

N →∞ −ν + µ + µ̃ + σ̃ log Φ (c∞ )
−c∞
=
′ .
c∞ + log Φ (c∞ )

Let us now choose the parameter as follows

ν = 5, µ = 1, µ̃ = 2, σ = 2, σ̃ = 1

and we consider different values for c. In this particular case we choose

c = 0, c = −1, c = −2, c = −5, respectively.

In Figure 3.2, a graphical representation of this situation is shown.


In Figure 3.3, we can observe that the returns {rN }N ∈N converge to r∞
for any c ≤ 0, as asserted in Proposition 3.2 . In fact,

2 ∼
r∞ =
′ = 5.359 .
−2 + log Φ (−2)
3.1. THE ASYMPTOTICAL LIMIT OF UPPER BOUNDS 23

6
5
4
3
r
2
1
0

2 4 6 8 10
N

Figure 3.2: This figure shows the risk-adjusted returns rN of the form mentioned
above for the parameter values under the assumptions that parameter c = 0 (dot-
ted line), c = −1 (short-dashed line), c = −2 (dashed line), c = −5 (solid line),
respectively, and the limit r∞ (long-dashed line).
6
5
4
3
r
2
1
0

0 10 20 30 40 50
N

Figure 3.3: Plot of the same returns rN shown in Figure 3.2 for greater values
of N .
24 CHAPTER 3. EXPECTED SHORTFALL

3.2 The n-dimensional model


The previous considerations are also valid for an n-dimensional model. In
fact, in general we can denote the whole profit of a company consisting of n
units by

 N1 
1 
R(N1 , . . . , Nn ) = N1 ν1 − X1,j − Y1 + · · ·
N1 j=1
 Nn 
1 
+ Nn νn − Xn,j − Yn ,
Nn j=1

where Ni are some positive integers for all i ∈ {1, . . . , n} and Y1 , . . . , Yn are
integrable random variables. Moreover, assume that real constants µ1 , . . . , µn
exist such that, for all i ∈ {1, . . . , n},

Ni
1  Ni →∞
Xi,j −− −→ µi in mean.
Ni j=1

Then we define
R(N1 , . . . , Nn )
R̃(N1 , . . . , Nn ) =
N1 + · · · + N n

and we obtain that the expected risk-adjusted return for all c ≤ 0 can be
written as

E R(N1 , . . . , Nn )
r(N1 , . . . , Nn ) =
E − R(N1 , . . . , Nn ) | R(N1 , . . . , Nn ) ≤ c

E R(N1 ,...,Nn )

=  N1 +···+Nn 
E −R(N 1 ,...,Nn )  R(N1 ,...,Nn )
N1 +···+Nn  N1 +···+Nn
≤ c
N1 +···+Nn

E R̃(N1 , . . . , Nn )
=   .
 c
E − R̃(N1 , . . . , Nn )R̃(N1 , . . . , Nn ) ≤ N1 +···+Nn

Moreover, we assume that N1 , . . . , Nn → ∞, such that

Ni
−→ ti exists, for all i = 1, . . . , n .
N1 + · · · + N n
3.2. THE N -DIMENSIONAL MODEL 25

Then, clearly, it follows that


n

n
(t1 , . . . , tn ) ∈ [0, 1] and ti = 1 (3.11)
i=1
and that
n

N1 ,...,Nn →∞
R̃(N1 , . . . , Nn ) −−−−−−−→ R̃∞ (t) := ti (νi − µi − Yi ) in mean,
i=1

for all t = (t1 , . . . , tn ) satisfying (3.11). Therefore assuming that


P R̃∞ (t) = 0 = 0,
it holds that
N ,...,Nn →∞
E[R̃(N1 , . . . , Nn ) | R̃(N1 , . . . , Nn ) ≤ 0] −−1−−−−−→ E R̃∞ (t) | R̃∞ (t) ≤ 0 .
Note that this is the same statement we arrived at (3.2) for the simplest
model. For this reason, we omit the proof in this case, because it suffices to
adapt the arguments considered there.
Analogously, as before, we obtain that the limit of the upper bounds
exists and will be denoted by

N ,...,Nn →∞

r N1 , . . . , Nn −−1−−−−−→ r∞ t1 , . . . , tn , (3.12)
where
n

i=1 ti (νi− µi − µ̃i )
r∞ t1 , . . . , tn = .
E − R̃∞ (t1 , . . . tn ) | R̃∞ (t1 , . . . tn ) ≤ 0
We can observe that the limit of the upper bounds obtained for the

n-
dimensional model is directly
dependent on (t1 , . . . tn ). If P R̃∞ (t) = 0 = 0
for every t = (t1 , . . . , tn satisfying (3.11), then (3.12) holds for
all these t and
it makes sense to determine the maximum of r∞ t1 , . . . , tn for (t1 , . . . , tn )
satisfying (3.11). In fact, if r∞ is upper semi-continuous, then this maximum
is attained by t∗1 , . . . , t∗n , which can be calculated through
n
∗ ∗ i=1 ti (νi − µi − µ̃i )
(t1 , . . . , tn ) = argmax .
(t1 ,...,tn ) E − R̃∞ (t1 , . . . tn ) | R̃∞ (t1 , . . . tn ) ≤ 0
ti ∈[0,1]

Then, for all (t1 , . . . , tn ) satisfying (3.11), it holds that



r∞ t1 , . . . , tn ≤ r∞ t∗1 , . . . , t∗n < ∞ .


Note that, if Y1 , . . . , Yn are

independent with continuous distribution func-
tions, then P R̃∞ (t) = 0 = 0 certainly holds for all (t1 , . . . , tn ) satisfying
(3.11).
26 CHAPTER 3. EXPECTED SHORTFALL

Example 2. Assume that an insurance company consists of two business


units, i.e., n = 2. Then
N1
 N2

R(N1 , N2 ) = ν1 N1 − X1,j − Y1 N1 + ν2 N2 − X2,j − Y2 N2 .
j=1 j=1

Assume that Xi,j ∼ N µi , σi2 ) and Yi ∼ N µ̃i , σ̃i2 ) for all i ∈ {1, 2} and
j ∈ N. Furthermore, assume that all the random variables are independent.
For simplicity, we denote µ̂i = νi − µi − µ̃i for i ∈ {1, 2},
µ̂ = N1 µ̂1 + N2 µ̂2
and
σ̂ 2 = N1 σ12 + N12 σ̃12 + N2 σ22 + N22 σ̃22 .
Then we have

R(N1 , N2 ) ∼ N µ̂, σ̂ 2 .
Using (A.3) as in the previous example, we obtain for c ≤ 0
E[R(N1 , N2 )]
r(N1 , N2 ) =
E[−R(N1 , N2 ) | R(N1 , N2 ) ≤ c]
µ̂
=
′  .
−µ̂ + σ̂ log Φ (c − µ̂)/σ̂

Assume that N1 , N2 → ∞ such that


N1 N ,N2 →∞
−−1−− −→ t exists.
N 1 + N2
Then
µ̂ N ,N2 →∞
−−1−− −→ t µ̂1 + (1 − t) µ̂2
N1 + N2
and 
σ̂ N ,N2 →∞
−−1−− −→ t2 σ̃12 + (1 − t)2 σ̃22 .
N1 + N 2
Hence
N ,N →∞
r(N1 , N2 ) −−1−−
2
−→ r∞ (t) ,
with
t µ̂1 + (1 − t) µ̂2
r∞ (t) = 

−t µ̂1 − (1 − t) µ̂2 + t2 σ̃12 + (1 − t)2 σ̃22 log Φ c(t)


−c(t)
=

,
c(t) + log Φ c(t)
3.2. THE N -DIMENSIONAL MODEL 27

where
t µ̂1 + (1 − t) µ̂2
c(t) = −  .
t2 σ̃12 + (1 − t)2 σ̃22

We now want to find the t∗ ∈ [0, 1] which maximizes the limiting expected
risk-adjusted return [0, 1] ∋ t → r∞ (t). We start with the following lemma.
Lemma 3.4. The function
c
(−∞, 0] ∋ c → −

c + log Φ (c)
is monotonely decreasing.
Proof. Using the substitution z = x2 /2, we get for every c ∈ (−∞, 0)
 c  c ∞
x e−z  ϕ(c)
Φ(c) = ϕ(x) dx < ϕ(x) dx = √  =− ,
−∞ c c

c 2π c2 /2
−∞

hence
1

log Φ (c) < −1 .
c
Therefore, it suffices to show that g : (−∞, 0) → R with
1

g(c) = log Φ (c), c ∈ (−∞, 0) ,
c
is monotonely decreasing.
Since
 ϕ(c) ′ −c2 ϕ(c) Φ(c) − ϕ(c) Φ(c) − c ϕ(c)2
g ′ (c) = = ,
c Φ(c) c2 Φ(c)2
we obtain that for every c ∈ (−∞, 0)
g ′ (c) ≤ 0 ⇐⇒ (c2 + 1)Φ(c) + c ϕ(c) ≥ 0
 1 ϕ(c)
⇐⇒ 1 + 2 Φ(c) ≥ − .
c c
Since for c < 0
  c 
1 1
1 + 2 Φ(c) = 1 + 2 ϕ(x) dx
c c
−∞
c 
1 ϕ(x) 
c ϕ(c)
≥ 1 + 2 ϕ(x) dx = −  =− ,
−∞ x x −∞ c
the lemma is proved. 
28 CHAPTER 3. EXPECTED SHORTFALL

01 2 3 4 5 6
01 2 3 4 5 6

r(c=-1)
r(c=0)

10 10
8 8
10 6 10
6 8 8
N2 4 6 N2 4 6
4 2 4
2 2 N1
0 2 N1 0 0
0

01 2 3 4 5 6
01 2 3 4 5 6

r(c=-5)
r(c=-2)

10 10
8 8
10 6 10
6 8 8
N2 4 6 N2 4 6
2 4 2 4
2 N1 2 N1
0 0 0 0

Figure 3.4: Plot of the risk-adjusted return r(N1 , N2 ): at the top left for the
parameter c = 0, at the top right for c = −1, at the bottom left for c = −2, at the
bottom right for c = −5.

Due to the lemma, it remains for us to find the t∗ ∈ [0, 1] which minimizes
[0, 1] ∋ t → c(t). Solving the equation c′ (t∗ ) = 0 leads to

µ̂1 σ̃22
t∗ = ,
µ̂1 σ̃22 + µ̂2 σ̃12
and 
µ̂21 µ̂22
c(t∗ ) = − + .
σ̃12 σ̃22

This is the minimum in [0, 1], because


µ̂1 µ̂2
c(1) = − > c(t∗ ) and c(0) = − > c(t∗ ) .
σ̃1 σ̃2
3.2. THE N -DIMENSIONAL MODEL 29

0
0

5
10
N1 8
6
4
10
2 N2

Figure 3.5: Plot of the risk-adjusted return r(N1 , N2 ) for the parameters c = −5,
c = −2, c = −1, c = 0, respectively.

In Figures 3.4–3.7 we give a graphical representation assuming that

ν1 = 5, µ1 = 1, µ̃1 = 2, σ1 = 2, σ̃1 = 1,
ν2 = 3, µ2 = 1, µ̃2 = 1, σ2 = 1, σ̃2 = 1,

and choosing

c = 0, c = −1, c = −2, c = −5, respectively.


30 CHAPTER 3. EXPECTED SHORTFALL

0
0 10 0.5
20 30 t
N 40 501

Figure 3.6: Plot of the risk-adjusted return r(N1 , N2 ) for N1 = t N and N2 =


(1 − t) N and for the parameters c = −5, c = −2, c = −1, c = 0, respectively.

0 0.2 0.4 0.6 0.8 1


t

Figure 3.7: Plot of the limit r∞ (t).


3.3. CONCLUSION 31

In this case, it holds that


−c(t)
r∞ (t) =

,
c(t) + log Φ c(t)
where
t+1
c(t) = √ .
t2 − 2 t + 1
Moreover, the t∗ ∈ [0, 1] which maximises r∞ (t) is
2
t∗ =
3
and therefore
r∞ (t∗ ) ∼
= 6.429 .
We can observe that, for the same values of N1 and N2 , increasing c,
r(N1 , N2 ) becomes greater. This is a direct consequence of Lemma 3.1 which
shows that the map c → E[−R | R ≤ c] decreases monotonously for any
c ∈ (c0 , ∞), where c0 = inf {c ∈ R | P (R ≤ c) > 0}. Moreover, assuming that
N1 , N2 → ∞ such that N1 /(N1 + N2 ) → t, we can observe in the following
figures that in any case – for every c ≤ 0 – the return r(N1 , N2 ) converges to
r∞ (t). In particular, in Figure 6, we can observe that r∞ (t) ≤ r∞ (t∗ ) ∼
= 6.429
for all t ∈ [0, 1].

3.3 Conclusion
If we examine the portfolio of an insurance company represented by the model
(3.1) and we try to optimize it by determining the number of contracts, we
can conclude that, for the expected risk-adjusted return rN defined by (3.2),
a limit exists.
Considering the n-dimensional model we obtain a limit of the upper
bounds which depends directly on the partition of the contracts between
the different n units of the company. An optimal partition can be calculated
in the case of normal distributions. In general, it has to be done numerically.
Nevertheless, we cannot find the optimal number of contracts which should
assure a maximal risk-adjusted return to the company.
Usually, the company has a fixed capital C at its disposal to invest. So,
considering that the insurance company wants to invest this capital C, it is
possible to calculate an optimal solution to the following problem:

E R(N1 , . . . Nn )
maximize
C

subject to ρ R(N1 , . . . , Nn ) ≤ C and Ni ≥ 0 .


32 CHAPTER 3. EXPECTED SHORTFALL

In this case, using the expected shortfall as risk measure, we do not obtain
a general solution, but it exists and can be found numerically if the distri-
butions of Yi for i = 1, . . . , n are known. If the random variables Yi for
i = 1, . . . , n are independent and normally distributed, this numerical cal-
culation is not very complicated, because of the particular properties of the
normal distribution.
However, we can observe that the obtained solution is not always a good
one, because frequently the optimal values for Ni for i = 1, . . . , n are not
integers. For this reason we have to consider once more which optimal integer
values we shall choose.
In the following chapters we will show that we obtain better solutions to
such a problem if we take the standard deviation risk measure into consid-
eration.
Chapter 4

Results for the standard


deviation risk measure

In this chapter, we want to study the model of the portfolio of an insurance


company defined in Section 2.5 once again. We will now examine the risk-
adjusted performance of the company by studying the return
E[R]
r=
ρ(R)
and using the standard deviation risk measure of R defined by

ρ(R) = −E[R] + κσ(R),

where κ > 0 is some positive constant and σ(R) denotes the standard devi-
ation of the risk R.
In consequence, we assume the existence of second moments for the fol-
lowing, because the definition of σ(R) clearly requires
 this.
Considering the portfolio of the form R = ni=1 Ri , let Ri be real-valued
random variables on (Ω, F, P ) having finite expectation and finite second
moment for all i ∈ {1, . . . , n}
Ni

Ri = νi Ni − Xi,j − Yi Ni , (4.1)
j=1

where Ni are some positive integers for all i ∈ {1, . . . , n}. For the sequences
{Xi,j }j∈N and for the random variables Yi with i ∈ {1, . . . , n} the assump-
tions mentioned in Section 2.5.2 are still valid. Moreover, we assume that
(R1 , . . . , Rn ) are non-trivial and this means that ρ(R) takes values other than
zero, where R is the portfolio of (R1 , . . . , Rn ).

33
34 CHAPTER 4. STANDARD DEVIATION

Let C > 0 be the capital that the company wants to invest. Now we look
at the following problem:

E[R]
maximize (4.2)
C
subject to ρ(R) ≤ C
Ni ≥ 0 integers for i ∈ {1, . . . , n}
but not all of them equal to zero.

We try to determine the optimal number of contracts in order to obtain a


maximal return r, which means a better performance for the company.
First, we will examine the simplest cases with n = 1 and n = 2, and
then we will take the same approach for a general n-dimensional model. In
Section 4.3 we will repeat the same procedure, again analyzing the same
model, but this time with the assumption that Ni are Poisson-distributed
random variables with parameters λi > 0 for all i = 1, . . . , n. This last
assumption is convenient for us, because in this case the optimal solution of
the optimization problem consists of real positive values which are no longer
required to be integers.

4.1 The simplest cases: n = 1 and n = 2


We start considering n = 1. Let the portfolio be represented by

N

R(N ) = νN − Xj − Y N (4.3)
j=1

and take the assumptions mentioned in Section 2.5. We then have that N
is a positive integer, {Xj }j∈N are uncorrelated random variables with finite
mean µ, Y has finite mean µ̃ and the sequence {Xj }j∈N and Y are uncor-
related. Moreover, we assume that both Xj for all j ∈ N and Y have finite
variances denoted by σ 2 = Var(Xj ), independent of j ∈ N, and σ̃ 2 = Var(Y )
respectively.
Then we can compute

E R(N ) = N (ν − µ − µ̃)


ρ R(N ) = −N (ν − µ − µ̃) + κ N σ 2 + N 2 σ̃ 2 .
4.1. THE SIMPLEST CASES: N = 1 AND N = 2 35

We get that the optimization problem (4.2) can be written as


N (ν − µ − µ̃)
maximize
C √
subject to − N (ν − µ − µ̃) + κ N σ 2 + N 2 σ̃ 2 ≤ C , (4.4)
N > 0 integer.
For the sake of practicality we define
ν − µ − µ̃
a= .
C
Note that a is a positive constant because ν − µ − µ̃ is assumed to be
positive and the capital C to invest is positive too. Moreover, from now on,
we assume
κ
σ̃ > a. (4.5)
C

This means that even for the case σ 2 = 0, we get ρ R(N ) > 0 for every
N ∈ N, hence, accepting contracts involves real risk. Furthermore, for any
σ2 ≥ 0 ,

ρ R(N )
lim > 0,
N →∞ N
i.e., the risk per contract stays positive even in the limit N → ∞.
Returning to the optimization problem; if we square the constraint (4.4)
we reduce the previous formulation to
maximize aN
 κ2   κ2 
subject to N 2 σ̃ 2
− a2
+ N σ 2
− 2a − 1 ≤ 0, (4.6)
C2 C2
N > 0 integer.
Given that κσ̃/C is assumed to be greater than a, we obtain that the con-
straint (4.6) is fulfilled if N ∈ [z1 , z2 ] ∩ N, where with z1 and z2 we denote
the zeros of the constraint itself. So, since a is positive and the function to
maximize becomes greater with increasing N , the value of N which assures
all requirements is the greatest positive integer which satisfies the condition
(4.6). We have that the greatest zero of (4.6) has the following form:
 2   2  
κ2 2 κ2 2
− Cκ 2 σ 2 − 2a + C 2 σ − 2a + 4 C 2 σ̃ − a2

z2 = 
2

2 Cκ 2 σ̃ 2 − a2

2aC 2 − κ2 σ 2 + κ κ2 σ 4 − 4aσ 2 C 2 + 4σ̃ 2 C 2
=
.
2 κ2 σ̃ 2 − a2 C 2
36 CHAPTER 4. STANDARD DEVIATION

We observe that z2 is clearly positive but not necessarily an integer. So,


in consequence of these arguments we can conclude that the solution of the
optimization problem (4.4) is

2 2 2
√ 
2aC − κ σ + κ κ 2 σ 4 − 4aσ 2 C 2 + 4σ̃ 2 C 2
N∗ =
,
2 κ2 σ̃ 2 − a2 C 2

where with ⌊x⌋ we denote the greatest integer smaller than x, i.e.,

⌊x⌋ = max{k ∈ Z | k ≤ x} .

Now let n = 2. The portfolio has the form


N1
 N2

R(N1 , N2 ) = ν1 N1 − X1,j − Y1 N1 + ν2 N2 − X2,j − Y2 N2 , (4.7)
j=1 j=1

where N1 , N2 are positive integers, {X1,j }j∈N and {X2,j }j∈N are sequences
of uncorrelated random variables which are uncorrelated to Y1 and Y2 . We
suppose that the sequences {X1,j }j∈N and {X2,j }j∈N are uncorrelated and Y1
and Y2 are uncorrelated, too. Furthermore, we assume that the first two
moments of {X1,j }j∈N and {X2,j }j∈N do not depend on j ∈ N. Moreover, we
define

µi = E[Xi,j ] and σi2 = Var(Xi,j ), for all j ∈ N and for all i = 1, 2,


µ̃i = E[Yi ] and σ̃i2 = Var(Yi ), for i = 1, 2,

where both means and variances are finite.


Then, we calculate

E R(N1 , N2 ) = N1 (ν1 − µ1 − µ̃1 ) + N2 (ν2 − µ2 − µ̃2 ) ,

ρ R(N1 , N2 ) = −N1 (ν1 − µ1 − µ̃1 ) − N2 (ν2 − µ2 − µ̃2 )



+ κ N1 σ12 + N12 σ̃12 + N2 σ22 + N22 σ̃22 .

In order to simplify the following calculations we define two constants as


above
ν1 − µ1 − µ̃1 ν2 − µ2 − µ̃2
a1 = and a2 = .
C C
Once again, we assume that these constants are positive, because we consider
contracts with positive expectation.
4.1. THE SIMPLEST CASES: N = 1 AND N = 2 37

The initial optimization problem (4.2) therefore turns out to have the
following form

maximize a1 N1 + a2 N2

κ
subject to − a1 N1 − a2 N2 + N1 σ12 + N12 σ̃12 + N2 σ22 + N22 σ̃22 ≤ 1 ,
C
(4.8)
N1 , N2 ≥ 0 integers, not both of them equal to zero.

It is useful, analogously, as before, to square the constraint (4.8), so we obtain

maximize a1 N 1 + a2 N 2
 κ2   κ2   2 
2 2 2 κ 2 2
subject to N1 σ − 2a1 + N 2 σ − 2a2 + N σ̃ − a
C2 1 C2 2 1
C2 1 1
 κ2 
+ N22 2
σ̃22 − a22 − 2a1 a2 N1 N2 − 1 ≤ 0 , (4.9)
C
N1 , N2 ≥ 0 integers, not both of them equal to zero.

Note that, both here and in the n-dimensional case, we make the same as-
sumption as in (4.5) and, in general, this implies
κ
σ̃i > ai for all i ∈ {1, . . . , n}.
C
In particular in this case, where n = 2, it is assumed that
κ κ
σ̃1 > a1 and σ̃2 > a2 . (4.10)
C C
Now we consider only the constraint (4.9). It can be written as
 κ2   κ2 
N1 σ2
2 1
− 2a1 + N2 σ2
2 2
− 2a2 + q ≤ 1 , (4.11)
C C
where with q we denote a real quadratic form

q = XT AX, (4.12)
with !
N1
X=
N2

and A representing the symmetric matrix of q relative to N1 and N2


κ2 2 2
!
C 2 σ̃1 − a1 −a 1 a2
A= κ2 2 2
.
−a1 a2 C 2 σ̃2 − a 2
38 CHAPTER 4. STANDARD DEVIATION

This notation is useful, since any real quadratic form q, with q = XT AX as in


(4.12) can be reduced to a diagonalized representation, and by an orthogonal
change of variables the expression (4.11) can be rewritten in a form related to
an ellipse. It therefore follows that the optimization problem can be solved
directly.
So we denote the eigenvalues of A by α1 and α2 and the associated or-
thonormal eigenvectors by P1 and P2 , then we can represent q in the form
!T !
x x
q= B = α1 x2 + α2 y 2 , (4.13)
y y
where !
T α1 0
B = P AP =
0 α2
and   !
p11 p12
P = P1 , P2 = ,
p21 p22

i.e., P is the orthogonal matrix with the orthonormal eigenvectors of A in


its columns. For more details see Gilbert and Gilbert (1995) Chapter 8.5.
From now on, we denote Y = (x, y)T . Then, by a change of variables from
N1 , N2 to x, y according to the rule that X = PY, the constraint considered
until now can be re-written as follows:

β1 x + β2 y + α1 x2 + α2 y 2 ≤ 1. (4.14)

Recall that if A is a real and symmetric matrix, its eigenvalues are real, so
α1 , α2 , β1 and β2 are real constants dependent on a1 , a2 , σ̃1 , σ̃2 , κ and C.
In particular, α1 and α2 have the following forms:
"  #
1 κ2 2
κ2 2
αi = σ̃ + σ̃22 − a21 − a22 ± (σ̃ 2 − σ̃22 ) − a21 + a22 + 4a21 a22 ,
2 C2 1 C2 1

where we choose α1 to correspond to the plus sign and α2 to the minus sign.
Both constants are positive due to (4.10).
Then the values of the constants β1 and β2 can be calculated too, but for
reasons of space, we will not list these formulae at full length. However it
holds that
 κ2   κ2 
2 2
βi = p1i σ − 2a1 + p 2i σ − 2a2
C2 1 C2 2
for i = 1, 2.
4.1. THE SIMPLEST CASES: N = 1 AND N = 2 39

The normalized eigenvector Pi associated to αi can be represented as


follows
κ2 2 2
!
vi C 2 σ̃1 − a1 − αi + a1 a2
Pi = , with vi = 2 .
vi  −a1 a2 − Cκ 2 σ̃22 − a22 − αi

Consequently, the form of the constraint mentioned in (4.14) can be trans-


formed once again, so as to get the following form related to an ellipse

β1 x + β2 y + α1 x2 + α2 y 2 ≤ 1
    βi
⇐⇒ α1 (x − γ1 )2 − γ12 + α2 (x − γ2 )2 − γ22 ≤ 1, with γi = ,
2αi

2
2
⇐⇒ α1 x − γ1 + α2 x − γ2 − α1 γ12 − α2 γ22 ≤ 1

2
2
α1 x − γ1 α2 y − γ2
⇐⇒ + ≤ 1. (4.15)
1 + α1 γ12 + α2 γ22 1 + α1 γ12 + α2 γ22

To make it easier, we define



1 + α1 γ12 + α2 γ22
εi = for i = 1, 2 .
αi

Clearly, ε1 and ε2 represent the two half axes of the ellipse given in (4.15).
By the same change of variable, the function to maximize

a1 N 1 + a2 N 2
becomes
ã1 x + ã2 y (4.16)
with
ãi = a1 p1i + a2 p2i for i = 1, 2 .

To recapitulate, we now want to solve the following problem:

maximize f (x, y)
subject to the condition g(x, y) ≤ 0 , (4.17)

where
f (x, y) = ã1 x + ã2 y ,

2
2
x − γ1 y − γ2
g(x, y) = + − 1.
ε21 ε22
40 CHAPTER 4. STANDARD DEVIATION

f (x, y) = const

ε2
ε1
(γ1 , γ2 )

g(x, y) = 0

Figure 4.1: Illustration of the optimization problem 4.17

Let us ignore the restrictions coming from N1 , N2 ∈ N0 with N1 , N2 = 0, 0


for the moment. An optimization problem of this form can be illustrated by
Figure 4.1, where the line f (x, y) = const is orthogonal to (ã1 , ã2 ).
We see that the required maximal value for f (x, y) is reached on the
boundary of the ellipse described by g(x, y) = 0, at the point on the ellipse
whose tangent is parallel to the straight line. The optimal solution is (x∗ , y ∗ )
and it can be calculated with the aid of Lagrange multipliers. In fact, by
introducing suitable multipliers, the constrained extremum problem can be
treated as one of an ordinary extremum. More precisely, as follows from the
Lagrange multipliers rule, it holds that:
if f has a relative extremum in (x∗ , y ∗ ) subject to the constraint g(x∗ , y ∗ ) = 0
and if gy (x∗ , y ∗ ) = 0, then a real number δ called the Lagrange multiplier
exists such that (x∗ , y ∗ , δ) is the critical point of the function H defined by
H(x, y, δ) = f (x, y) + δg(x, y). For more details see Appendix B.
Therefore, the pertinent problem to resolve is

Hx (x, y, δ) = fx (x, y) + δgx (x, y) = 0

Hy (x, y, δ) = fy (x, y) + δgy (x, y) = 0 (4.18)


Hδ (x, y) = g(x, y) = 0.
4.1. THE SIMPLEST CASES: N = 1 AND N = 2 41

We get a system of three equations and in our case we find a solution with the
aid of the computer. Then, through the inverse change of variables, we can
finally determine the values for N1 and N2 , respectively, that solve the initial
problem (4.8). We obtain that the number of contracts which guarantee a
maximal risk-adjusted return are

∗ a1 σ̃22 κ s + C 2 a1 a2 σ22 − a22 σ12 − 2a1 σ̃22 + κ2 σ12 σ̃22


N1 = 

2 2 2 2 2 2 2 2
2 C a1 σ̃2 + a2 σ̃1 − κ σ̃1 σ̃2

and

a2 σ̃12 κ s + C 2 a1 a2 σ12 − a21 σ22 − 2a2 σ̃12 + κ2 σ22 σ̃12


N2∗ = 
 ,
2 2 2 2 2 2 2 2
2 C a1 σ̃2 + a2 σ̃1 − κ σ̃1 σ̃2

where


2  
C 2 4 σ̃12 σ̃22 − a1 σ12 σ̃22 − a2 σ22 σ̃12 − a1 σ22 − a2 σ12 + κ2 σ14 σ̃22 + σ24 σ̃12
s= .
a21 σ̃22 + a22 σ̃12

Note that N1∗ and N2∗ defined above are not necessarily integers, so they have
to be rounded in practice.
Remark. We can proceed analogously even if we consider a model of the same
form as (4.7) and we allow correlation between the two random variables
Y1 , Y2 . In fact, it holds that

E R(N1 , N2 ) = N1 ν1 − µ1 − µ̃1 + N2 ν2 − µ2 − µ̃2


and

Var R(N1 , N2 ) = Var(R1 + R2 )


= Var(R1 ) + Var(R2 ) + 2 Cov(R1 , R2 )
= N1 σ̃12 + N12 σ̃12 + N2 σ̃22 + N22 σ̃22 + 2 N1 N2 ρY σ̃1 σ̃2 ,
because
Cov(R1 , R2 ) = N1 N2 Cov(Y1 , Y2 )
= N1 N2 ρY σ̃1 σ̃2 ,

where ρY = corr(Y1 , Y2 ). Therefore, if we want to solve the optimization


problem of the form of (4.2) for such a model R, we must find an optimal
solution in the same way as before. As a matter of fact, in this case the
42 CHAPTER 4. STANDARD DEVIATION

optimization problem turns out to be very similar to (4.9). In particular, the


constraint has the following form
 κ2   κ2   2 
2 2 2 κ 2 2
N1 σ − 2a 1 + N 2 σ − 2a 2 + N σ̃ − a
C2 1 C2 2 1
C2 1 1
 κ2 

+ N22 σ̃ 2
− a2
− 2 N 1 N 2 a 1 a 2 − ρ Y σ̃1 σ̃2 −1≤0
C2 2 2

and can be written as


 κ2   κ2 
2 2
N1 σ − 2a1 + N2 σ − 2a2 + q ≤ 0,
C2 1 C2 2
where with q we denote a real quadratic form
!T !
N1 N1
q= A
N2 N2
with !
κ2 2
− a21
σ̃
C2 1
−a1 a2 + ρY σ̃1 σ̃2
A= κ2 2 .
−a1 a2 + ρY σ̃1 σ̃2 C2 2
σ̃ − a22

Thus, adapting the previously considered argumentations (such as change of


variables, form related to an ellipse, Lagrange multipliers) we can compute
the optimal solution.

4.2 The general case


If we consider a general model related to a firm consisting of n units, i.e.,
n

R= Ri , with Ri of the form of (4.1),
i=1

we can proceed analogously, as before. In fact, similar arguments can be


adapted to solve an n-dimensional optimization problem, too.
Starting from the initial problem of the form of (4.2), we square the
constraint ρ(R) ≤ C and then we represent it with help of a quadratic form
defined from an n × n−matrix A. Diagonalizing A and by an orthogonal
change of variables, we can rewrite the constraint in a form related to a conic
section. Therefore, we get a problem with new variables, i.e., x1 , . . . , xn
instead of N1 , . . . , Nn and it has the following form

maximize f (x1 , . . . , xn )
subject to the condition g(x1 , . . . , xn ) ≤ 0,
4.3. THE SECOND VARIANT OF THE MODEL 43

where, as before, the maximal value for f (x1 , . . . , xn ) is reached on the bound-
ary described by g(x1 , . . . , xn ) = 0.
This means that the optimal solution (x∗1 , . . . , x∗n ) can be found employ-
ing the Lagrange multipliers rule (see Appendix B for more details). As a
consequence, we have to solve a system of n + 1 equations with respect to
the unknowns x1 , . . . , xn , δ
$
Hxi (x1 , . . . , xn , δ) = fxi (x1 , . . . , xn ) + δgxi (x1 , . . . , xn ) = 0 , i = 1, . . . , n
Hδ (x1 , . . . , xn ) = g(x1 , . . . , xn ) = 0 .
(4.19)
Then, by the inverse orthogonal change of variables, we can determine the so-
lution of the initial problem with respect to the original variables N1 , . . . , Nn .
Due to the length of the expression related to the two-dimensional case, we
avoid listing the values of N1∗ , . . . , Nn∗ , i.e., the optimal solution of the initial
problem.

4.3 The second variant of the model: the Ni’s


are random variables
In this section, we still consider the same model representing the whole port-
folio of a company – but with one difference. In fact, we assume that the
number of contracts for every unit i, denoted by Ni with i ∈ {1, . . . , n}, are
Poisson-distributed random variables with parameter λi > 0.
To resume, we have
n 
 Ni
 
R= νi Ni − Xi,j − Yi Ni , (4.20)
i=1 j=1
with

• Ni ∼ POIS(λi ) for all i ∈ {1, . . . , n},

• {Xi,j }j∈N are independent1 sequences of i.i.d. random variables with


finite means and finite variances for all i ∈ {1, . . . , n}

µi = E[Xi,j ] and σi2 = Var(Xi,j ),


1
Note that these independence assumptions are introduced for simplicity. In fact, even
if we consider a model which does not fulfil these supplementary requirements, we can
proceed in a similar way.
44 CHAPTER 4. STANDARD DEVIATION

• Y1 , . . . , Yn are independent1 random variables with both means and


variances finite

µ̃i = E[Yi ] and σ̃i2 = Var(Yi ),

• the sequences {Xi,j }j∈N , i ∈ {1, . . . , n}, are independent of the random
variables Y1 , . . . , Yn .

Moreover, the number of contracts N1 , . . . , Nn are assumed to be independent


of the sequences {Xi,j }j∈N for all i ∈ {1, . . . , n} and of the random variables
Y1 , . . . , Yn .
Recall that the aforementioned assumption

κ νi − µi − µ̃i
σ̃i >
C C

is still valid for all i ∈ {1, . . . , n}.


We are now interested in the solution of the optimization problem

E[R]
maximize (4.21)
C
subject to ρ(R) ≤ C
λi > 0 for all i ∈ {1, . . . , n}.

We observe that this problem is analogous to (4.2), but that this time we have
weaker conditions, in fact the values of the solution are no longer required
to be integers.
Before considering this problem, we must do some calculations needed in
the following. First, in the next Proposition we recall two particular equalities
concerning expectation and variance.

Proposition 4.1. Let


X be an R-valued random variable defined on a prob-
ability space Ω, F, P with a finite second moment and let G be a sub-σ-field
of F. Then

E[X] = E E[X | G]

Var(X) = E Var X | G + Var E[X | G ] .

Proof. See Fristedt and Gray (1997), Chapter 23. 


4.3. THE SECOND VARIANT OF THE MODEL 45

Then, since Ni is independent of the sequence {Xi,j }j∈N and of Yi and


because of the Wald Identity we obtain
n
  Ni
 
E[R] = E Ni νi − Xi,j − Ni Yi
i=1 j=1
n
 Ni
  !
= E[Ni νi ] − E Xi,j − E[Ni Yi ]
i=1 j=1
n  
= νi E[Ni ] − E[Ni ]E[Xi,j ] − E[Ni ]E[Yi ]
i=1
n


= λi νi − µi − µ̃i
i=1

and since all the sequences {Xi,j }j∈N with i ∈ {1, . . . , n} are independent of
Y1 , . . . , Yn it holds that
n
  Ni
 
Var(R) = Var Ni νi − Xi,j − Ni Yi
i=1 j=1
n
 Ni
   
= Var(Ni νi ) + Var Xi,j + Var(Ni Yi ) − 2 Cov Ni νi , Ni Yi
i=1 j=1
Ni
 ! Ni
 !!
− 2 Cov Ni νi , Xi,j + 2 Cov Xi,j , Ni Yi .
j=1 j=1

First, we calculate the single variances and covariances separately:


Var(Ni νi ) = νi2 Var(Ni ) = λi νi2 ,
Ni
 

Var Xi,j = E[Ni σi2 ] + Var(Ni µi ) = λi σi2 + µ2i ,


j=1

Var(Ni Yi ) = E[Ni2 σ̃i2 ] + Var(Ni µ̃i ) = (λ2i + λi )σ̃i2 + λi µ̃2i ,


Ni
 ! Ni !
Cov Ni νi , Xi,j = νi Cov Ni , Xi,j
j=1 j=1
 Ni
  Ni
 !
= νi E Ni Xi,j − E[Ni ]E Xi,j
j=1 j=1
 
2
= νi E Ni µi − E[Ni ]E[Ni ]µi
= λi νi µi ,
46 CHAPTER 4. STANDARD DEVIATION

Ni
 !  Ni
  Ni
 

Cov Xi,j , Ni Yi = E N i Yi Xi,j − E Xi,j E Ni Yi
j=1 j=1 j=1
2
= E Ni2 µi µ̃i − E Ni µi µ̃i
= λi µi µ̃i ,
   
Cov Ni νi , Ni Yi = νi Cov Ni , Ni Yi
 
2 2
= νi E[Ni Yi ] − E[Ni ] E[Yi ]
= λi νi µ̃i .

Then, if we resume, we obtain


n 


Var(R) = λi νi2 + λi σi2 + µ2i + (λ2i + λi )σ̃i2 + λi µ̃2i + 2λi µi µ̃i


i=1

− 2λi νi µi − 2λi νi µ̃i
n 

= λ2i σ̃i2 + λi νi2 + µ2i + µ̃2i + 2µi µ̃i − 2νi µi − 2νi µ̃i
i=1


+ σi2 + σ̃i2
n 


= λ2i σ̃i2 + λi (νi − µi − µ̃i )2 + σi2 + σ̃i2 .
i=1

Remark. Note that, if we allow dependence between the sequences {Xi,j }j∈N ,
for i ∈ {1, . . . , n}, and also between the random variables Y1 , . . . , Yn , it holds
that
n
 Ni
 

Var(R) = Var Ni νi − Xi,j − Ni Yi
i=1 j=1

 n
n−1   Ni
 Nk
 
+2 Cov Ni νi − Xi,j − Ni Yi , Nk νk − Xk,j − Nk Yk ,
i=1 k>i j=1 j=1

but we will not deal with this case in detail.


In order to determine the solution of (4.21) we take the same approach
as in the previous section. This means that we consider the simplest cases
with n = 1 and n = 2. Because of the length of the expressions we will
not reiterate the general case, but all the following argumentations can be
4.3. THE SECOND VARIANT OF THE MODEL 47

adapted in order to solve an n-dimensional optimization problem with respect


to the model (4.20).
Let n = 1 and
N
R(N ) = νN − Xj − Y N
j=1

with the above-mentioned assumptions. Recall that in this section N is a


Poisson-distributed random variable with parameter λ > 0.
From the previous calculations we obtain

E R(N ) = λν − µ − µ̃

Var R(N ) = λ2 σ̃ 2 + λ (ν − µ − µ̃)2 + σ 2 + σ̃ 2 .

We now have the standard deviation risk measure for R(N ) which is



ρ R(N ) = −λ(ν − µ − µ̃) + κ λ2 σ̃ 2 + λ (ν − µ − µ̃)2 + σ 2 + σ̃ 2 .

Consequently, in this case the initial problem (4.21) has the following form

λ(ν − µ − µ̃)
maximize
C 

subject to − λ(ν − µ − µ̃) + κ λ2 σ̃ 2 + λ (ν − µ − µ̃)2 + σ 2 + σ̃ 2 ≤ C ,


λ > 0.

Analogously, as before, it is useful to introduce some positive constants in


order to get shorter expressions. We define
ν − µ − µ̃
a= ,
C

2
b = ν − µ − µ̃ + σ 2 + σ̃ 2 .

We remember that κ/C is assumed to be greater than a, because of the


argumentations already mentioned.
Moreover, if we square the constraint we obtain a new formulation of the
problem to solve, which is very similar to (4.6).

maximize aλ
 κ2   κ2 
subject to λ2 σ̃ 2
− a2
+ λ b − 2a − 1 ≤ 0, (4.22)
C2 C2
λ > 0.
48 CHAPTER 4. STANDARD DEVIATION

We observe that, with respect to the problem (4.6), the only difference
is the value of the constant related to the variable λ. In this case we will
not rewrite all the passages to calculate the solution, because those in the
previous section are easily adaptable. Therefore, we can determine that the
optimal solution of (4.22) is

∗ 2aC 2 − κ2 b + κ κ2 b2 − 4abC 2 + 4σ̃ 2 C 2
λ =
.
2 κ2 σ̃ 2 − a2 C 2

Now, let n = 2 and


N1
 N2

R(N1 , N2 ) = ν1 N1 − X1,j − Y1 N1 + ν2 N2 − X2,j − Y2 N2 ,
j=1 j=1

with the assumptions cited at the beginning of this section. From the previ-
ous calculation we obtain

E R(N1 , N2 ) = λ1 ν1 − µ1 − µ̃1 + λ2 ν2 − µ2 − µ̃2



Var R(N1 , N2 ) = λ21 σ̃12 + λ1 (ν1 − µ1 − µ̃1 )2 + σ12 + σ̃12


+ λ22 σ̃22 + λ2 (ν2 − µ2 − µ̃2 )2 + σ22 + σ̃22 ,

and introducing the constants

ν1 − µ1 − µ̃1
a1 = ,
C
ν2 − µ2 − µ̃2
a2 = ,
C
b1 = (ν1 − µ1 − µ̃1 )2 + σ12 + σ̃12 ,
b2 = (ν2 − µ2 − µ̃2 )2 + σ22 + σ̃22 ,

it follows that the problem to be solved in this case is

maximize a 1 λ 1 + a2 λ 2

κ
subject to − a1 λ1 − a2 λ2 + λ1 b1 + λ21 σ̃12 + λ2 b2 + λ22 σ̃22 ≤ 1, (4.23)
C
λ 1 , λ2 > 0 .

Remember that we assume


κ κ
σ̃1 > a1 and σ̃2 > a2 .
C C
4.4. THE THIRD VARIANT OF THE MODEL 49

Once again, squaring the main constraint, we get a new formulation very
similar to (4.9) with the exception of the constants related to λ1 and to λ2
respectively, i.e.,
maximize a1 λ1 + a2 λ2
 κ2   κ2   2 
2 κ 2 2
subject to λ1 b1 − 2a1 + λ2 b2 − 2a2 + λ1 σ̃ − a1
C2 C2 C2 1
 κ2 
+ λ22 σ̃ 2
− a2
2 − 2a1 a2 λ1 λ2 − 1 ≤ 0 (4.24)
C2 2
λ 1 , λ2 > 0 .
We can calculate the solution for this problem in the same way as in the
previous section. Without rewriting all the passages concerning the respec-
tively real quadratic form q, its matrix and the orthogonal change of variables
– which bring us to a form of the constraint corresponding to an ellipse – we
obtain that the optimal value of the function to maximize is reached by

∗ a1 σ̃2 κ s + κ2 σ̃2 b1 + C 2 a1 a2 b2 − a22 b1 − 2a1 σ̃2


λ1 = 

2 2 2 2
2 C a1 σ̃2 + a2 σ̃1 − κ σ̃1 σ̃2

and
√ 2 2
2

a 2 σ̃1 κ s + κ σ̃1 b 2 + C a 1 a2 b 1 − a1 b 2 − 2a2 σ̃1


λ∗2 = 
 ,
2 2 2 2
2 C a1 σ̃2 + a2 σ̃1 − κ σ̃1 σ̃2

where we use s as short notation for





2 
κ2 σ̃2 b21 + σ̃1 b22 + C 2 4 σ̃1 σ̃2 − a2 σ̃1 b2 − a1 σ̃2 b1 − a1 b2 − a2 b1
s= .
a21 σ̃2 + a22 σ̃1
All these argumentations are also valid for an n-dimensional problem. Con-
sequently, the initial optimization problem (4.21) has optimal solutions for
every model of the form of (4.20), but because of the length of the expression
we will not list general solutions.

4.4 The third variant of the model: Ni is the


sum of a positive integer and a random
variable for i ∈ {1, . . . , n}
In this section, we consider a different variation of the model examined in
this work and defined in Section 2.5. This new variation is based on the
50 CHAPTER 4. STANDARD DEVIATION

definition of Ni , i.e. the variables which represent the number of contracts of


every unit i of the company, for i ∈ {1, . . . , n}. We define
Ni = Nifix + Nipois for i ∈ {1, . . . , n}.
This means that the number of contracts of any unit i is defined as the sum
of a fixed integer Nifix and a Poisson-distributed random variable Nipois with
parameter λi ≥ 0.
We observe that this variant of the model includes both versions we con-
sidered previously. In fact, if we set Nifix = 0 for all i ∈ {1, . . . , n}, we obtain
the same variant as those examined in Section 4.3 or, if we set λi = 0 for all
i ∈ {1, . . . , n}, we have the initial model considered in Sections 4.1 and 4.2.
We represent the whole portfolio through
fix pois
n  Ni +Ni
 fix pois



R= νi Ni + Ni − Xi,j − Yi Nifix + Nipois ,
i=1 j=1

with the assumptions mentioned. In particular, we remember that {Xi,j }j∈N


with i ∈ {1, . . . , n} are independent sequences of i.i.d. random variables,
independent of the random variables Y1 , . . . , Yn and that Y1 , . . . , Yn are in-
dependent, too. Moreover, N1 , . . . , Nn are assumed to be independent of all
the sequences {Xi,j }j∈N with i ∈ {1, . . . , n} and of Y1 , . . . , Yn . Analogously,
we define the finite means and finite variances of Xi,j and Yi , respectively, by
µi = E[Xi,j ] and σi2 = Var(Xi,j ), for all j ∈ N and all i ∈ {1, . . . , n} ,
µ̃i = E[Yi ] and σ̃i2 = Var(Yi ), for all i ∈ {1, . . . , n}.
Our interest turns to the following optimization problem in order to de-
termine the portfolio which guarantees a maximal return.
E[R]
maximize , (4.25)
C
subject to ρ(R) ≤ C,
λi ≥ 0 for all i ∈ {1, . . . , n} ,
Nifix ≥ 0, integers for all i ∈ {1, . . . , n} .
We will consider only the simplest case with n = 1, because the form of Ni
already entails solving a two-dimensional optimization problem. A solution
for a more general problem can be computed with the same argumentations,
which can be easily adapted. For the sake of practicality, we rewrite the risk
R as follows
Ñ
R = ν Ñ − Xj − Y Ñ ,
j=1
4.4. THE THIRD VARIANT OF THE MODEL 51

where
Ñ = N + N p ,
with N denoting some positive integer and N p a Poisson-distributed random
variable with parameter λ ≥ 0. Using the formulae cited in the previous
section, we compute
E[R] = (N + λ)(ν − µ − µ̃) ,
+N p
N !

Var(R) = Var ν (N + N p ) + Var Xj + Var Y (N + N p )


j=1
+N p
N ! +N p
N !
p p
− 2 Cov ν(N + N ), Xj + 2 Cov Xj , Y (N + N )
j=1 j=1

− 2 Cov ν(N + N p ), Y (N + N p )

= λ (ν − µ − µ̃)2 + σ 2 + σ̃ 2 + N σ 2 + (N 2 + λ2 )σ̃ 2 .
To simplify the calculations we introduce the constants
a = (ν − µ − µ̃)/C ,
b = (ν − µ − µ̃)2 + σ 2 + σ̃ 2 .
Moreover, as in previous sections we assume κσ̃/C > a. Therefore, the
optimization problem can be written as
maximize a(N + λ) (4.26)
κ
subject to − a(N + λ) + λb + N σ 2 + (N 2 + λ2 ) σ̃ 2 ≤ 1 ,
C
λ ≥ 0, N ≥ 0 .
If we square the main constraint we obtain
maximize a(N + λ) (4.27)
 κ2   κ2   κ2 
subject to λ2 2 2 σ̃ 2 − a2 + λ 2 b − 2a + N 2 2
σ̃ 2 − a2
C C C
 κ2   κ2 
+N σ 2 − 2a − 2N λ 2 σ̃ 2 − 2a − 1 ≤ 0,
C2 C
λ ≥ 0, N ≥ 0.
Then, if we consider only the constraint, we can rewrite it in a new form
related to an ellipse. In fact, introducing a quadratic form
 κ2   2   κ2 
2 2 2 2 κ 2 2 2
q = λ 2 2 σ̃ − a + N σ̃ − a − 2N λ 2 σ̃ − 2a
C C2 C
52 CHAPTER 4. STANDARD DEVIATION

and computing its corresponding matrix A, we can determine an orthogonal


matrix P with the orthonormal eigenvectors of A in its columns. Then, by
a change of variables from N, λ to x, y according to
! !
N x
=P ,
λ y

we obtain that the constraint can be written as

β1 x + β2 y + α1 x2 + α2 y 2 ≤ 1.

We observe that we get the same result as in the previous sections. The
only differences are the values of the constants, but, in any case, the way to
proceed is similar. Without writing all the calculations needed to compute
these values, we can now solve the optimization problem as we did in Section
4.2, and we obtain that the optimal solution is
4 2
2
2 2
2 2

κ σ̃ 2σ + b − aC κ 6σ̃ − a(b + σ ) + 4a3 C 4


N∗ =

2κ2 σ̃ 2 a2 C 2 − κ2 σ̃ 2
2 2

3κ σ̃ − 2a2 C 2 s
+ 2 2 2 2

2κ σ̃ a C − κ2 σ̃ 2

and

∗ 4aC 2 − κ2 σ 2 + b − 2 s
λ = ,
2κ2 σ̃ 2
where we use s as short notation for

16a4 C 6 − 8 a3 (σ 2 + b) + 2a2 σ̃ 2 κ2 C 4 − (σ 2 + b)2 + σ 4 σ̃ 2 κ6


s=
2 2 4a2 C 2 − 5κ2 σ̃ 2

a (σ + b)2 + 4a(2b + 3σ 2 )σ̃ 2 − 4σ̃ 4 κ4 C 2


+ .
4a2 C 2 − 5κ2 σ̃ 2
Chapter 5

Capital allocation

An insurance company has a total target return and spreads down the total
return to different portfolios when establishing a business plan. The company
is also interested in comparing the individual returns ri = E[Ri ]/Ci of the
various business units i ∈ {1, . . . , n}.
The vital question concerning this is: how can one choose Ci ? The idea
is to introduce a capital allocation principle and to split up the risk capital
amongst the various business units, but there is no general answer to the
question of how the risk capital should be allocated. There are different
classes of capital allocation methodologies, and a formal description can be
found in Albrecht (1997).
In the following sections, we consider two different capital allocation prin-
ciples, namely the covariance principle and the expected shortfall principle.

5.1 Covariance principle


We now want to calculate the well-known covariance principle for the special
case of RAC(R) = −E[R] + κσ(R) with κ > 0 and with R having the
following form
n n  Ni
 
R= Ri = νi Ni − Xi,j − Yi Ni
i=1 i=1 j=1

where
• N1 , . . . , Nn are positive integers,
• {Xi,1 , Xi,2 , . . . , Xi,Ni } are finite sequences of i.i.d. random variables with
finite means and finite variances for all i ∈ {1, . . . , n} with
µi = E[Xi,j ] and σi2 = Var(Xi,j ),

53
54 CHAPTER 5. CAPITAL ALLOCATION

• the linear correlation coefficients satisfy


ρik = corr(Xi,j , Xk,l )
for all i, k ∈ {1, . . . , n}, i = k and j ∈ {1, . . . , Ni }, l ∈ {1, . . . , Nk },
• Y1 , . . . , Yn are random variables with finite means and finite variances
µ̃i = E[Yi ] and σ̃i2 = Var(Yi ),

• the sequences {Xi,1 , . . . , Xi,Ni }, for every i ∈ {1, . . . , n}, are indepen-
dent of the random variables Y1 , . . . , Yn .
Remark. There are restrictions on the correlation coefficients ρik with i, k in
{1, . . . , n}, i = k. The covariance matrix of the (N1 + · · · + Nn )-dimensional
random vector
(X1,1 , . . . , X1,N1 , X2,1 , . . . , X2,N2 , . . . , Xn,1 , . . . , Xn,Nn )
has to be positive semidefinite. If this is the case, then there exist at least
multivariate normally distributed random variables with the prescribed de-
pendence structure.
In general it holds that
Cov(Ri , R)
Ci = C ,
Var(R)
where C denotes the capital the company wants to invest and Ci is the capital
the company will allocate to business unit i.
Let us compute the enumerator and the denominator separately. On one
side, we have that, for the enumerator, it holds that
  n 
Cov(Ri , R) = Cov Ri , Rk
k=1
  n 
= Var(Ri ) + Cov Ri , Rk
k=1
k =i

 Ni
  n

= Var νi Ni − Xi,j − Yi Ni + Cov Ri , Rk


j=1 k=1
k =i
n


= Ni σi2 + Ni2 σ̃i2 + Ni Nk ρik σi σk + ρ̃ik σ̃i σ̃k ,


k=1
k =i
5.1. COVARIANCE PRINCIPLE 55

since, for any k = i,


 Ni
 Nk
 

Cov Ri , Rk = Cov νi Ni − Xi,j − Yi Ni , νk Nk − Xk,l − Yk Nk


j=1 l=1
Ni
 Nk
 
= Cov Xi,j + Yi Ni , Xk,l + Yk Nk
j=1 l=1
Nk
Ni 


= Cov Xi,j , Xk,l + Ni Nk Cov Yi , Yk ,


j=1 l=1
   
=ρik σi σk =ρ̃ik σ̃i σ̃k

where ρ̃ik denotes the linear correlation coefficient, in particular for i, k in


{1, . . . , n} with i = k,

ρ̃ik = corr Yi , Yk .

On the other side, for the denominator, it holds that


n
 
Var(R) = Var Rj
j=1
n
 n−1 
 n

= Var Rj + 2 Cov Rj , Rk
j=1 j=1 k=j+1
n
 n−1 
 n

= Nj σj2 + Nj2 σ̃j2 +2 Nj Nk ρjk σj σk + ρ̃jk σ̃j σ̃k .


j=1 j=1 k=j+1

Therefore, it follows that




Ni σi2 + Ni2 σ̃i2 + nk =i Ni Nk ρik σi σk + ρ̃ik σ̃i σ̃k


Ci = C n 2 2 2

n−1 n
.
j=1 Nj σj + Nj σ̃j + 2 j=1 k=j+1 Nj Nk ρjk σj σk + ρ̃jk σ̃j σ̃k

Remark. In the case where {Xi,j }j∈N are independent sequences consisting
of i.i.d. random variables and Y1 , . . . , Yn are independent too, it holds that
all correlation coefficients are zero. Therefore, it follows that
Ni σi2 + Ni2 σ̃i2
Ci = C n 2 2 2

.
j=1 Nj σj + Nj σ̃j

We now consider the second variant of the initial model. In fact, we


assume that all Ni , with i ∈ {1, . . . , n}, which denote the number of contracts
for every unit i, are Poisson-distributed random variables with parameters
56 CHAPTER 5. CAPITAL ALLOCATION

λi > 0. To recapitulate, we will calculate the covariance principle for R


having the following form
n
 n 
 Ni
 
R= Ri = νi Ni − Xi,j − Yi Ni ,
i=1 i=1 j=1

where
• N1 , . . . , Nn are Poisson-distributed random variables, i.e.,
Ni ∼ POIS(λi ) with λi > 0,

• {Xi,j }j∈N are uncorrelated sequences of i.i.d. random variables with


finite means and finite variances for all i ∈ {1, . . . , n}
µi = E[Xi,j ] and σi2 = Var(Xi,j ),

• Y1 , . . . , Yn are random variables with finite means and finite variances


µ̃i = E[Yi ] and σ̃i2 = Var(Yi ),

• all the sequences {Xi,j }j∈N with i ∈ {1, . . . , n} are independent of the
random variables Y1 , . . . , Yn ,
• N1 , . . . , Nn are assumed to be independent of all the sequences {Xi,j }j∈N
with i ∈ {1, . . . , n} and of Y1 , . . . , Yn .
Remark. Consider sequences {Ui }i∈N and {Vi }i∈N of i.i.d. random variables.
For simplicity, assume that E[Ui ] = E[Vi ] = 0 and Var(Ui ) = Var(Vi ) = 1
for all i ∈ N. If the correlation ρ = Cov(Ui , Vj ) does not depend on i, j ∈ N,
then ρ has to be zero. To see this, fix n ∈ N. Then
n !
0 ≤ Var (Ui + Vi )
i=1
n
 n


= Var(Ui ) + Var(Vi ) + 2 Cov(Ui , Vj )


i=1 i,j=1
2
= 2n + 2n ρ
implies ρ ≥ −1/n. Similarly,
n
 !
0 ≤ Var (Ui − Vi ) = 2n − 2n2 ρ,
i=1

hence ρ ≤ 1/n. Since n ∈ N was arbitrary, ρ = 0. For this reason, the above
sequences {Xi,j }j∈N for i ∈ {1, . . . , n} are assumed to be uncorrelated.
5.1. COVARIANCE PRINCIPLE 57

In the above model it holds that

Cov(Ri , R)
Ci = C ,
Var(R)

with

  n 
Cov(Ri , R) = Var(Ri ) + Cov Ri , Rk
k=1
k =i
n


= λ2i σ̃i2 + λi (νi − µi − µ̃i ) + 2


σi2 + σ̃i2 + µ2i + Cov(Ri , Rk )
k=1
k =i

and
n
 n−1 
 n
Var(R) = Var(Ri ) + 2 Cov(Ri , Rk ) .
i=1 i=1 k=i+1

In detail, we first calculate the covariance between Ri and Rk for i = k


separately:

 Ni
 Nk
 
Cov(Ri , Rk ) = Cov νi Ni − Xi,j − Yi Ni , νk Nk − Xk,l − Yk Nk
j=1 l=1
 Nk
 
= Cov(νi Ni , νk Nk ) − Cov νi Ni , Xk,l − Cov(νi Ni , Yk Nk )
l=1
Ni
  Ni
 Nk
 
− Cov Xi,j , νk Nk + Cov Xi,j , Xk,l
j=1 j=1 l=1
Ni
 
+ Cov Xi,j , Yk Nk − Cov(Yi Ni , νk Nk )
j=1
 Nk
 
+ Cov Yi Ni , Xk,l + Cov(Yi Ni , Yk Nk ) ,
l=1

with

Cov(νi Ni , νk Nk ) = νi νk ρ̂ik λi λk ,
58 CHAPTER 5. CAPITAL ALLOCATION

 Nk
   Nk
 
Cov νi Ni , Xk,l = νi Cov Ni , Xk,l
l=1 l=1
 N k  Nk !
  
= νi E Ni Xk,l − E Ni E Xk,l
l=1 l=1
 
= νi E Ni Nk µk − λi λk µk

= νi µk Cov(Ni , Nk ) = νi µk ρ̂ik λi λk ,
 
Cov(νi Ni , Yk Nk ) = νi E Ni Yk Nk − E Ni E Yk Nk

= νi µ̃k Cov(Ni , Nk ) = νi µ̃k ρ̂ik λi λk ,

Ni
    Ni  Ni !

Cov Xi,j , νk Nk = νk E Nk Xi,j − E Nk E Xi,j
j=1 j=1 j=1

= νk µi ρ̂ik λi λk ,

Ni
 Nk
  Ni
 Nk
  Ni
 Nk
  
Cov Xi,j , Xk,l = E Xi,j Xk,l − E Xi,j E Xk,l
j=1 l=1 j=1 l=1 j=1 l=1

= E N i Nk µ i µk − µ i µk λ i λ k

= µi µk Cov Ni , Nk

= ρ̂ik λi λk µi µk ,

Ni
 
Cov Xi,j , Yk Nk = E Ni Nk µi µ̃k − λi µi λk µ̃k
j=1

= µi µ̃k Cov(Ni , Nk ) = µi µ̃k ρ̂ik λi λk ,
 
Cov(Yi Ni , νk Nk ) = νk E Yi Ni Nk − E Yi Ni E Nk

= νk µ̃i Cov(Ni , Nk ) = νk µ̃i ρ̂ik λi λk ,

 Nk
 
Cov Yi Ni , Xk,l = E Ni Nk µ̃i µk − λi µ̃i λk µk
l=1

= µ̃i µk Cov(Ni , Nk ) = µ̃i µk ρ̂ik λi λk ,
5.1. COVARIANCE PRINCIPLE 59

Cov(Yi Ni , Yk Nk ) = E Yi Ni Yk Nk − µ̃i µ̃k λi λk

= E Yi Yk E Ni Nk − µ̃i µ̃k λi λk


= Cov Yi , Yk Cov Ni , Nk + µ̃i µ̃k Cov Ni , Nk


+ λi λk Cov Yi , Yk
 
= ρ̃ik σ̃i σ̃k ρ̂ik λi λk + ρ̂ik λi λk µ̃i µ̃k
+ ρ̃ik σ̃i σ̃k λi λk ,

where with ρ̂ik and ρ̃ik we denote the linear correlation coefficients between
the dependent random variables, i.e.,

ρ̂ik = corr Ni , Nk and ρ̃ik = corr Yi , Yk ,

respectively. It follows that


 


Cov(Ri , Rk ) = ρ̂ik λi λk νi − µi − µ̃i νk − µk − µ̃k + ρ̃ik σ̃i σ̃k
+ λi λk ρ̃ik σ̃i σ̃k .

A similar calculation
√ with i = k for the nine terms given above gives similar
results with ρ̃ik λi λk replaced by λi and ρ̃ik σ̃i σ̃k replaced by σ̃i2 . However,
for the fifth term we get
Ni
 ! Ni
  Ni
  N1
 !2
2
Var Xi,j =E Xi,j Xi,l + E Xi,j − E Xi,j
j=1 j,l=1 j=1 j=1
j =l

= µ2i E[Ni (Ni − 1)] + E[Ni ] E[Xi,1


2
] − (E[Ni ] µi )2
= λ2i µ2i + λi E[Xi,1
2
] − λ2i µ2i
= λi σi2 + λi µ2i ,

hence

Var(Ri ) = λ2i σ̃i2 + λi (νi − µi − µ̃i )2 + σi2 + σ̃i2 + µ2i .

Therefore, we can conclude that


Cov(Ri , R) = λ2i σ̃i2 + λi (νi − µi − µ̃i )2 + σi2 + σ̃i2 + µ2i


n   

+ ρ̂ik λi λk νi − µi − µ̃i νk − µk − µ̃k + ρ̃ik σ̃i σ̃k


k=1 !
k =i
+ λi λk ρ̃ik σ̃i σ̃k ,
60 CHAPTER 5. CAPITAL ALLOCATION

and
n
 n−1 
 n
Var(R) = Var(Ri ) + 2 Cov(Ri , Rk )
i=1 i=1 k=i+1
n !

= λ2i σ̃i2 + λi (νi − µi − µ̃i ) + 2


σi2 + σ̃i2 + µ2i +
i=1
n−1 
 n
+2 λi λk ρ̃ik σ̃i σ̃k
i=1 k=i+1
 

!
+ ρ̂ik λi λk νi − µi − µ̃i νk − µk − µ̃k + ρ̃ik σ̃i σ̃k .

Remark. Note that where N1 , . . . , Nn are independent random variables, all


the {Xi,j }j∈N with i ∈ {1, . . . , n} are uncorrelated sequences and the ran-
dom variables Y1 , . . . Yn are also independent, it holds that all correlation
coefficients are zero. Thus, it follows that

λ2i σ̃i2 + λi (νi − µi − µ̃i )2 + σi2 + σ̃i2 + µ2i


Ci = C  
 .
n 2 2 2 + σ 2 + σ̃ 2 + µ2
j=1 λ σ̃
j j + λ j (νj − µ j − µ̃ j ) j j j

5.2 Expected-shortfall principle


An alternative to the covariance principle for the purpose of allocation of risk
capital are the conditional expectations.
nLet R1 , . . . , Rn be the stochastic gains of the business units and R =
i=1 Ri the whole profit and loss of an insurance company. Let c ≤ 0 be
the capital loss threshold, for example, the α-quantile rα of R. If a total loss
R ≤ c occurs, we consider the expected shares E[−Ri | R ≤ c] of the single
losses with respect to the total loss. Obviously, it holds that
n

E[−R | R ≤ c] = E[−Ri | R ≤ c] ,
i=1

where E[−R | R ≤ c] denotes the risk capital of the entire company, while
E[−Ri | R ≤ c] denotes the risk capital assigned to business unit i.
This risk capital allocation principle is additive, as is the covariance prin-
ciple, but in addition it can be applied for integrable random variable, so the
existence of second moments is no longer required. In order to compute these
conditional expectations we can use the procedures proposed in Appendix A.
5.2. EXPECTED-SHORTFALL PRINCIPLE 61

As an example, we consider R having the usual form for a company


consisting of two units, i.e.

 N1
   N2
 
R = R1 + R2 = ν1 N1 − X1,j − Y1 N1 + ν2 N2 − X2,j − Y2 N2 ,
j=1 j=1

with the following assumptions:

• N1 and N2 are positive integers,

• the sequences {X1,j }j∈N and {X2,j }j∈N are independent and consist of
independent, identically normally distributed random variables, i.e.,

Xi,j ∼ N µi , σi2 , for i = 1, 2 and j ∈ N,

• Y1 and Y2 are independent normally distributed random variables, i.e.,


Yi ∼ N µ̃i , σ̃i2 , for i = 1, 2,

• the sequences {X1,j }j∈N and {X2,j }j∈N are independent of Y1 and Y2 .

It is known that, given the independent random variables Xi,j that are
normally distributed for i = 1, 2 and all j ∈ N, the random variables R1 , R2
and R are also normally distributed, i.e.,


R1 ∼ N µ̂1 , σ̂12 , R2 ∼ N µ̂2 , σ̂22 and R ∼ N µR , σR2 ,

where we define

µ̂i = Ni νi − µi − µ̃i , σ̂i2 = Ni σi2 + Ni2 σ̃i2 , for i = 1, 2 ,

and
µR = µ̂1 + µ̂2 , σR2 = σ̂12 + σ̂22 .
We denote the density of the standard normal distribution by
1 1 2
ϕ(u) = √ e− 2 u , u ∈ R,

and the standard normal distribution function by
 t
Φ(t) = ϕ(u) du, t ∈ R.
−∞
62 CHAPTER 5. CAPITAL ALLOCATION

Note that t − µ 

R
P R≤t =Φ .
σR
We are now interested in the measures of risk

E Ri | R ≤ c

for a given value c ∈ R and i = 1, 2 . Using formula (A6) from Appendix A,


we obtain !
σ̂i2 ′ c − µR
E[Ri | R ≤ c] = µ̂i − (log Φ) .
σR σR
Appendix A

Calculating expected shortfall

In the following pages we will introduce some rules which will be seen to be
useful to calculate the expected shortfall.
Let X be an integrable random variable with distribution function FX .
Then, for all c ∈ R such that FX (c) < 1, it holds that
 ∞
1
E X |X > c = x FX (dx). (A.1)
1 − FX (c) c
LetX1 , . . . , Xn be exchangeable1 integrable random variables and let
n
X = i=1 Xi . Then, for all c ∈ R such that P (X > c) > 0 and for all
i, j ∈ {1, . . . , n}, it holds that

E Xi | X > c = E Xj | X > c .

Since the conditional expectation is linear, for all i ∈ {1, . . . , n}, it follows
that
1
E Xi | X > c = E X | X > c
n
and this last conditional expectation can be solved by means of (A.1).
Let X and Y be independent integrable random variables with distribu-
tion functions FX and FY , respectively, and let c ∈ R such that
P (X + Y > c) > 0. It then holds that

1

E X |X + Y > c =
x 1 − FY (c − x) FX (dx) ,
1 − FX ∗ FY (c) R
(A.2)
1

Let I be a
countable set. A sequence Xi , i ∈ I of random variables on a probability
space Ω, F,
P is exchangeable
if, for every permutation ρ of I, the distributions of
Xρ(i) , i ∈ I and Xi , i ∈ I are identical. Note that a finite or infinite i.i.d. sequence is
exchangeable.

63
64 APPENDIX A. CALCULATING EXPECTED SHORTFALL

where FX ∗ FY denotes the convolution of the distribution functions FX and


FY , that is, the distribution function of the sum X + Y , i.e.,

P (X + Y > c) = 1 − FX ∗ FY (c) .

This result is generalizable to independent random variables X1 , . . . , Xn


with distribution functions F1 , . . . , Fn , respectively. Let X = X1 and
Y = X2 + · · · + Xn . X and Y are likewise independent with distribution
functions FX = F1 and FY = F2 ∗ · · · ∗ Fn , respectively. The conditional
expectation E X1 | X1 + · · · + Xn > c can be then solved by means of (A.2).
Let
1 2
ϕ(t) = √ e−t /2 , t ∈ R,

and  x
Φ(x) = ϕ(t) dt , x ∈ R,
−∞

denote the density and the distribution function of the standard normal dis-
tribution, respectively.
Given X ∼ N (0, 1) and c ∈ R, we obtain with the substitution z = x2 /2
 c
1
E[X | X ≤ c] = xϕ(x) dx
Φ(c) −∞
 ∞
1
= −√ e−z dz
2π Φ(c) c2 /2
2
e−c /2
= −√
2π Φ(c)
ϕ(c)

=− = − log Φ (c).
Φ(c)

If Y ∼ N (µ, σ 2 ), then

Y −µ
X := ∼ N (0, 1)
σ
and with the above result we obtain

E[Y | Y ≤ c] = E[µ + σX | µ + σX ≤ c]
= µ + σE[X | X ≤ (c − µ)/σ] (A.3)

′  c − µ 
= µ − σ log Φ .
σ
65

In particular, for c = 0,
E[Y ] µ
=

E[−Y | Y ≤ 0] −µ + σ log Φ −µ σ
1
=

.
−1 + σµ log Φ −µ σ

Now consider independent random variables X, Y ∼ N (0, 1) and con-


stants c ∈ R, σ > 0. To compute E[X | σX + Y ≤ c], define
c
γ=√
1 + σ2
and note that
P (σX + Y ≤ c) = Φ(γ)

because (σX + Y )/ 1 + σ 2 ∼ N (0, 1). By conditioning on the σ-algebra
generated by X,
1
E[X | σX + Y ≤ c] = E E[X1{σX+Y ≤c} | X]
Φ(γ)
1
= E[XP (Y ≤ c − σX | X)] .
Φ(γ)
Since X and Y are independent,

P (Y ≤ c − σX | X) = Φ(c − σX) P -almost surely.

Since xϕ(x) = −ϕ′ (x), partial integration gives


 
E[XΦ(c − σX)] = xϕ(x)Φ(c − σx) dx = −σ ϕ(x)ϕ(c − σx) dx .
R R

Since

2
x2 + (c − σx)2 = (1 + σ 2 )x2 − 2cσx + c2 = 1 + σ 2 x − σγ + γ 2 ,

the substitution u = 1 + σ 2 x − σγ yields
 

ϕ(x)ϕ(c − σx) dx = ϕ(γ) ϕ 1 + σ 2 x − σγ dx


R R

ϕ(γ)
=√ ϕ(u) du
1 + σ2 R
ϕ(γ)
=√ .
1 + σ2
66 APPENDIX A. CALCULATING EXPECTED SHORTFALL

Therefore,

σ ϕ(γ)
E[X | σX + Y ≤ c] = − √
1 + σ 2 Φ(γ)
! (A5)
σ ′ c
= −√ (log Φ) √ .
1 + σ2 1 + σ2
2
More generally, for independent X ∼ N (µX , σX ) and Y ∼ N (µY , σY2 )
with σX > 0 and σY > 0, we obtain

E[X | X + Y ≤ c]
  
X − µX  σX X − µX Y − µY c − µX − µ Y
= µX + σX E + ≤ .
σX  σY σX σY σY

Since
X − µ X Y − µY
, ∼ N (0, 1) ,
σX σY
equation (A5) gives

E[X | X + Y ≤ c]
!
σX /σY c − µ − µY
= µX − σX  (log Φ) ′
 X
1 + (σX /σY )2 σY 1 + (σX /σY )2 (A6)
2
!
σX ′ c − µX − µ Y
= µX −  2 (log Φ)  2 .
σX + σY2 σX + σY2
Appendix B

The Lagrange multipliers rule

We consider the general case. Let m and n be natural numbers and U


be an open subset of Rn+m . Suppose x ∈ Rn , y ∈ Rm , f : U → R and
g : U → Rm . Let us examine the problem of finding the extrema of f (x, y) =
f (x1 , . . . , xn ; y1 , . . . , ym ) subject to the m constraints:

g1 (x1 , . . . , xn ; y1 , . . . , ym ) = 0 ,
.. ..
. .
gm (x1 , . . . , xn ; y1 , . . . , ym ) = 0 .

Theorem. Let m, n and U be defined as above.


Suppose that f ∈ C 1 (U, R) and g ∈ C 1 (U, Rm ). Moreover, assume that
(ξ, η) ∈ M = {(x, y) ∈ U | g(x, y) = 0} is such that the restriction f |M of f to
M has a local extremum at (ξ, η); this means that there is some neighborhood
V of (ξ, η) such that either f (x, y) ≤ f (ξ, η) for all (x, y) ∈ V ∩ M or
f (x, y) ≥ f (ξ, η) for all (x, y) ∈ V ∩ M . Suppose also that the n by n matrix
∂g(ξ, η)/∂y has nonzero determinant. Then, a vector λ = (λ1 , . . . , λm ) ∈ Rm
exists such that the function

H(x, y, z) = f (x, y) + "z, g(x, y)#


m
= f (x, y) + zj gj (x, y) , (x, y, z) ∈ U × Rm ,
j=1

has a critical point in (ξ, η, λ).

Proof. See Walter (1990), pp. 130–133. 

67
68 APPENDIX B. THE LAGRANGE MULTIPLIERS RULE

As consequence, in the point (ξ, η, λ) it holds that

Hxi = fxi + "λ, gxi # = 0 , i = 1, . . . , n ,


Hyk = fyk + "λ, gyk # = 0 , k = 1, . . . , m , (B.1)
Hλk = gk = 0 , k = 1, . . . , m .

The variables λ1 , . . . , λn are called Lagrange multipliers. The virtue of this


result is that if one seeks points (ξ, η) at which f |M has local extrema, then
one need only search among those (ξ, η) ∈ M for which the system (B.1) has
a solution λ.
Bibliography

Albrecht, P., (1997). Risk Based Capital Allocation and Risk Adjusted
Performance Management in Property/Liability Insurance: A Risk
Theoretical Framework. Joint Day Proocedings of the XXVIIIth In-
ternational AFIR Colloquium & the 7th International AFIR Collo-
quium. Cairns, Australia, 13 August 1997.
Arztner, P., Delbaen, F., Eber, J. M., Heath, D., (1998). Coherent mea-
sure of risk. Mathematical Finance, Volume 9/3, July 1999.
Billingsley, P. (1968). Convergence of Probability Measures. John Wiley
& Sohns, New York.
Billingsley, P. (1979). Probability and Measures. John Wiley & Sohns,
Inc. New York.
Breiman, L. (1968). Probability. Addison-Wesley Publishing Company.
Crouhy, M., Turnbull, S. M., Wakeman, L. M. (1999). Measuring risk-
adjusted performance. Journal of Risk. Volume 2/Number 1, Fall
1999.
Denault, M. (1999). Coherent allocation of risk capital. RiskLab Re-
search Paper. Available under http://www.risklab.ch/Papers.html .
Embrechts, P., Küppelberg, C. and Mikosch, T.(1999). Modelling Ex-
tremal Events for Insurance and Finance. Springer
Feller, W., (1966). An Introduction to Probability Theory and its Appli-
cations. Volume II. John Wiley & Sohns, Inc.
Fristedt, B., Gray, L. (1997). A Modern Approach to Probability Theory.
Birkhäuser, Boston.
Gilbert, J. and Gilbert, L. (1995). Linear Algebra and Matrix Theory.
Academic Press, Inc.
Grimmet, G. R. and Stirzaker, D. R. (1992). Probability and random
processes. Oxford Science Publications, Oxford.

69
70 BIBLIOGRAPHY

Johnson, N. L., Kotz, S., Balakrishnan, N. (1994). Continuous univari-


ate Distribution. Volume 1, 2nd Edition. John Wiley & Sohn, Inc..
Karlin, S., Taylor, H. M. (1975). A first course in stochastic processes.
2nd Edition, Academic Press, Inc. New York.
Schmock, U. and Straumann, D. Allocation of risk capital and perfor-
mance measurement. Working paper, private communication.
Stromberg, K. R. (1981). An introduction to classical real analysis.
Wadsworth International Group, Belmont, California.
Tasche, D. (1999). Risk contributions and performance measurement.
Research paper, Zentrum Mathematik (SCA), TU München.
Walter, W. (1990). Analysis II. Springer-Verlag.