Risk Theory

Maria de Lourdes Centeno

ISEG - Master in Actuarial Science
(These notes are exclusively for class use)
Contents

1 Introduction 1

2 The number of claims 2

2.1 The (a,b,0) class of distributions or Katz family of distributions 3
2.1.1 The Poisson distribution . . . . . . . . . . . . . . . . . 5
2.1.2 The Negative Binomial . . . . . . . . . . . . . . . . . . 6
2.1.3 The binomial distribution . . . . . . . . . . . . . . . . 8
2.2 The (a; b; 1) class of distributions . . . . . . . . . . . . . . . . 8
2.3 Compound frequency models . . . . . . . . . . . . . . . . . . . 10
2.4 Mixed frequency distributions . . . . . . . . . . . . . . . . . . 14
2.5 E¤ect of exposure on frequency . . . . . . . . . . . . . . . . . 16
2.6 Model selection . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Impact of coverage modi…cations 20

3.1 Deductibles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Policy limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Coinsurance, deductibles and limits . . . . . . . . . . . . . . . 24
3.4 Impact of deductibles on the claim frequency . . . . . . . . . . 25

4 Aggregate Loss Models 26

4.1 Collective risk model versus individual risk model . . . . . . . 26
4.2 Assumptions and characteristics of the compound model . . . 26
4.3 Special cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.3.1 Compound Poisson Model . . . . . . . . . . . . . . . . 27
4.3.2 Compound Mixed Poisson Model . . . . . . . . . . . . 28
4.4 Analytic results . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.5 The aggregate claim distribution . . . . . . . . . . . . . . . . 30
4.5.1 Recursive method (Panjer’s recursion formula) . . . . . 30
4.5.2 Constructing Arithmetic Distributions . . . . . . . . . 34
4.6 Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.6.1 The NP approximation . . . . . . . . . . . . . . . . . . 39

i
CONTENTS ii

4.6.2 Translated Gamma Approximation . . . . . . . . . . . 40

5 Premium principles 42
5.1 Some premium calculation principles . . . . . . . . . . . . . . 43
5.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6 Risk measures 47
6.1 Coherent risk measures . . . . . . . . . . . . . . . . . . . . . . 47
6.2 Value-at-Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

7 Reinsurance 50
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.1.1 What is reinsurance? . . . . . . . . . . . . . . . . . . . 50
7.1.2 Functions of Reinsurance . . . . . . . . . . . . . . . . . 51
7.2 Forms of reinsurance . . . . . . . . . . . . . . . . . . . . . . . 53
7.3 Quota-share treaty . . . . . . . . . . . . . . . . . . . . . . . . 54
7.4 Surplus treaty . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
7.4.1 Surplus treaty: insurer’s vs. reinsurer’s experience . . . 56
7.5 Excess of loss covers . . . . . . . . . . . . . . . . . . . . . . . 57
7.6 Aggregate excess, or stop loss covers . . . . . . . . . . . . . . 60

8 Ruin Theory 61
8.1 Continuous time model . . . . . . . . . . . . . . . . . . . . . . 61
8.1.1 The adjustment coe¢ cient . . . . . . . . . . . . . . . . 62
8.1.2 A Functional equation to the ultimate ruin probability 63
8.1.3 The maximum aggregate loss . . . . . . . . . . . . . . 66
8.2 Discrete time model . . . . . . . . . . . . . . . . . . . . . . . . 70
Chapter 1

Introduction

The main aim of Risk Theory is the study and development of mathematical
models that describe well the technical aspects of the insurance business. Ac-
tuarial mathematics started in the seventeen century with the …rst mortality
table of Sir Edmund Halley.
The classic theory was associated to life insurance and considered the
portfolios as a sum of policies. A new era started in the beginning of the
20th century with the works of Lundberg. The theory developed by him and
some other Swedish authors became known as Collective Risk Theory. The
collective risk theory model has its applications to Life and Non-Life Insur-
ance, although the most important applications are for Non-Life Insurance,
also called General Insurance. By historical reasons and by the big di¤er-
ences between the two branches the study of life and of non-life is usually
made separately.
Unless otherwise stated we are considering the Collective Risk Model.
We will not distinguish here between claims and losses. We leave that to
the Claim Reserving course.

1
Chapter 2

The number of claims

We consider that a risk is a policy or a set of policies, maybe the entire

portfolio.
The claim number processes are by its nature counting processes. When
the time is …xed we have a counting distribution, i.e. a discrete random
variable taking values on the non-negative integers.
We will denote by N the number of claims that happen in a …xed period
of time, for a given risk. Let its probability function be

pk = PrfN = kg; k = 0; 1; 2; :::

As we know in this case we can de…ne the probability generating function

X
1
N
PN (z) = E[z ] = pk z k :
k=0

The moment generating function, when it exists is

X
1 X
1
rk
MN (r) = E[erN ] = pk erk = E[X k ] :
k=0 k=0
k!

The cumulant generating function gN (r) is the logarithm of the moment

generating function
X
1
sk
gN (s) = k
k=1
k!
where n are the cumulants of N: 1 = E[N ]; 2 = V [N ] and 3 = 3 (N ):

2
CHAPTER 2. THE NUMBER OF CLAIMS 3

Table 2.1: (a; b; 0) class of distributions

Distribution Probability function a b

k
e
Poisson 0
k!
r+k 1 r k
1
Negative Binomial 1+ 1+ 1+ (r 1) 1+
k

m q
Binomial q k (1 q)m k
1 q (m + 1) 1 q q
k

The moments of N can be calculated by several ways. The derivatives at

zero of the moment generating function give the raw moments. The derivat-
ives of the probability generating function at 1 give the factorial moments.
The derivatives at zero of the cumulant generating function give the cumu-
lants, which are function of the moments.

2.1 The (a,b,0) class of distributions or Katz

family of distributions
We will denote by N the number of claims that happen in a …xed period of
time, for a given risk. Let its probability function be

pk = PrfN = kg; k = 0; 1; 2; :::

De…nition 2.1 pk = PrfN = kg; k = 0; 1; 2 : : : is a member of the (a; b; 0)

class of distributions if there exist constants a and b such that
b
pk = a+ pk 1 ; k = 1; 2; : : : : (2.1)
k
We can easily check that the Poisson, the binomial and the negative
binomial are members of this family, with a and b chosen according to Table
2.1.

In fact they are the only members of this family, as it was proved by
(Sundt & Jewell 1981).
CHAPTER 2. THE NUMBER OF CLAIMS 4

Figure 2.1: Class (a; b; 0) of distributions

CHAPTER 2. THE NUMBER OF CLAIMS 5

Theorem 2.1 The only distributions taking values on the nonnegative in-
tegers satisfying (2.1)are the Poisson, binomial and negative binomial (which
includes the geometric).
Proof. In Figure 2.1 the lines b = (m + 1)a for m = 1; 2; : : : and
a < 0 correspond to the binomial distribution. The positive ordinate axis
corresponds to the Poisson and the region with b > a and 0 < a < 1
correspond to the negative binomial. Consider the region such that b < a,
ie. a + b < 0, which implies that (a + b)p0 < 0 which is to say (2.1) can not
hold for k = 1.
Suppose now that a < 0 and b a. The line b = a with a < 0 is the
degenerate variable at the origin because p1 = (a + b)p0 = 0. For all other
a < 0 and b > a either there exists a k such that pk+1 is nule or pk will
become negative for some k big enough. But pk+1 will become null for some
k when b is of the form b = (m + 1)a for m = 1; 2; : : :.
Finally in the region a 1 and b > a we have that a + b=k a a=k
(k 1)=k which implies that p2 > p1 =2, p3 > p1 =3; : : : ; pn > p1 =n; : : :. Sum-
ming pk over , we would have
X
1
1 1
pk > p1 1 + + + ::: ;
k=1
2 3

which diverges.

2.1.1 The Poisson distribution

k
e
pk = ; k = 0; 1; 2; :::
k!
(z 1)
PN (z) = e

(er 1)
MN (r) = e

gN (s) = ln MN (s) = (es 1)

k
E[N (N 1) : : : (N k + 1)] = ; k = 1; 2; : : : ;
E[N ] = ;
2
E[N 2 ]= + ;
2 3
E[N 3 ]= + 3 + :
CHAPTER 2. THE NUMBER OF CLAIMS 6

E[N ] = N = ;

2
Var[N ] = N = ;
p
N = 1= :

Theorem 2.2 Let N1 ; :::; Nn be independent Poisson random variables with

parameters 1 ; 2 ; :::;P n : Then N = N1 + ::: + Nn has a Poisson distribution
with parameter = ni=1 i :

Theorem 2.3 Suppose that the N is a Poisson with mean : Suppose that
each event can be classi…ed into one of m types with probabilities r1 ; r2 ; :::; rm ;
(where r1 + r2 + ::: + rm = 1) independently of all the other events. Then the
number of events N1 ; ..., Nm classi…ed in each type are independent Poisson
random variables with means r1 ; r2 ; :::; rm :

Proof. For n = 2 :

PrfN1 = n1 ; N2 = n2 g = PrfN1 = n1 ; N2 = n2 jN = ng PrfN = ng =

(n1 +n2 )
n n1 n2 e 1
= r r = (r1 )n1 (r2 )n2 e (r1 +r2 )
=
n1 1 2 n! n1 !n2 !
e r1 ( r1 )n1 e r2 ( r2 )n2
=
n1 ! n2 !
Hence
X
1
PrfN1 = n1 g = PrfN1 = n1 ; N2 = n2 g =
n2 =0

( r1 )n1 X e
r1 1 r2
e ( r2 )n2 e r1
( r1 )n1
= =
n1 ! n =0
n2 ! n1 !
2

and
r2
( r2 )n2e
PrfN2 = n2 g =
n2 !
As the joint probability is the product of the marginals the r.v. are inde-
pendent.

2.1.2 The Negative Binomial

We have already seen that the Negative Binomial can be regarded as a mix-
ture of Poissons where the structure distribution is a gamma, i.e. given
CHAPTER 2. THE NUMBER OF CLAIMS 7

= ; N is a Poisson random variable with mean ; and is the observation

of a random variable gamma distributed with parameters ( = r; = ):

pk = PrfN = kg = E[PrfN = kj g] =
Z 1 k r 1
e e =
= r d =
0 k! (r)
k r
r+k 1 1
= ; k = 0; 1; :::
k 1+ 1+

where
x x(x 1):::(x k 1) (x + 1)
= =
k k! (k + 1) (x k + 1)
with x > k 1 in the last expression.
The probability generating function is

PN (z) = E[z N ] = E[E[z N j ]] = E[e (z 1)

] = (1 (z 1)) r

and the moment generating function

MN (t) = (1 (et 1)) r ;

so the cumulant generating function

gN (s) = ln MN (s) = r ln(1 (et 1)):

From here we get

E[N ] = r

2
V ar[N ] = r + r

3 2 3
E[(N N) ] = (r + 3r + 2r )
The Poisson distribution may be regarded as the limit of the negative
binomial when r ! 1; ! 0; and the product r is constant (= ):
CHAPTER 2. THE NUMBER OF CLAIMS 8

r
lim 1 (z 1) = exp lim r ln 1 (z 1) =
r!1 r r!1 r
( )
ln 1 r
(z 1)
= exp lim 1
= (L´Hôpital´s rule)
r!1 r
( 1
)
r2
(z 1) 1 r
(z 1)
= exp lim 2
=
r!1 r
r (z 1)
= exp lim = (L´Hôpital´s rule)
r!1 (r (z 1))
= exp f (z 1)g

As a special case of the negative binomial when r = 1; we have the

geometric distribution. In this case
k
1
pk = ; k = 0; 1; :::
1+ 1+

PrfN > m + njN mg = PrfN > ng: It is used, as the exponential in the
continuos cas, to divide the distributions in heavy tail and light tail. The
negative binomial has a heavy tail when r < 1 and a light tail when r > 1:

2.1.3 The binomial distribution

The binomial is quite di¤erent from the Poisson or negative binomial. The
variance is smaller than the mean. It describes the situation were m inde-
pendent risks are each subject to the probability q of making a claim. It
is used for life insurance. The number of claims for each individual is a
Bernoulli with parameter q; and the number of claims of the m independent
and identical individuals is a Binomial (m; q):

PN (z) = (1 + q(z 1))m :

2.2 The (a; b; 1) class of distributions

At times the distributions of class (a; b; 0) do not describe well the charac-
teristics of some data sets encountered in practice. This may be the case
when there are too many zeros or too few zeros. In this section we study
adjustments to modi…cation at zero to the (a; b; 0) class of distributions.
CHAPTER 2. THE NUMBER OF CLAIMS 9

De…nition 2.2 Let pk = PrfN = kg; k = 0; 1; 2; : : : be the p.d.f. of a

discrete r.v. taking values at the nonnegative integers. This distribution
belongs to the (a; b; 1) class of distributions if there are constants a and b
such that
b
pk = a + pk 1 ; k = 2; 3 : : : : (2.2)
k

Zero-modi…ed distributions If we modify theprobability at zero of any

member of the (a; b; 0) class, i.e.

1 pM
0
pM
k = pk ; k = 1; 2; : : : ; (2.3)
1 p0

with pM 0 ; 0 pM
0 < 1; to be the modi…ed probability, we obtain a mem-
ber of the (a; b; 1) class which we designate by zero-modi…ed distribution:
The members of the (a; b; 1) class are the zero-modi…ed Poisson , the zero-
modi…ed binomial and the zero-modi…ed negative binomial. Note that these
distributions can be regarded as a mixture of a member of the class (a; b; 0)
with a degenerate distribution at the origin.
When pM 0 = 0 the modi…ed distribution is called truncated at zero.
The (a; b; 1) class does not contain only the zero-modi…ed members of the
class (a; b; 0). The space parameter of the negative binomial can be extended
to case where r > 1; r 6= 0. As the negative binomial probability function
is of the form
r k
r+k 1 1
pk = ; k = 0; 1; 2 : : : ;
k 1+ 1+

verifying the recursion with a = 1+ and b = (r 1) 1+ ; for k = 1; 2; : : :the

same happening with the same a and b; for the modi…ed negative binomial
for k = 2; 3; : : : ; i.e.

b
pM
k = a+ pM
k 1 ; k = 2; 3; : : : .
k

Noticing that a + b=k > 0; 8k 2 if and only if a + b=2 > 0; which is

equivalent to r > 1: Hence we only have to guaranty that pM
1 > 0 to have
pM
k > 0; k = 1; 2; : : :. But
r
1 pM
0 1
pM
1 = rr (2.4)
1 1 1+ 1+
1+
CHAPTER 2. THE NUMBER OF CLAIMS 10

is positive for > 0 and

P 0 < pM 0 < 1: Additionally we can show that
pk k=1;2;::: is such that k=0;1;:: pM
M
k = 1:
The ETNB is also de…ned for r > 1; 6= 0. When r ! 0; the limiting
case of the ETNB is the logarithmic distribution.

b
pTk = a+ pTk 1 =
k
!
1+ k 1
= pTk 1 = pTk 1 :
1+ k 1+ k
Hence
k 1
k 1k 2 1 T
pTk = ::: p =
1+ k k 1 2 1
k 1
1
= pT1 :
1+ k
Adding from 1 to 1 we get that
1
pT1 = 1 P1 k
=
1
1+ k=0 1+ k
1
= =
1+ ln 1 1+
1
= = ;
1+ ln 1 (1 + ) ln(1 + )
1+

1
from where
k
1
pTk = ; k = 1; 2; :::
1+ k ln(1 + )

2.3 Compound frequency models

We can create a larger counting class by compounding two or more distribu-
tion. Let N be a counting distribution with probability generating function
PN (z) and let fMi g be a sequence of i.i.d. counting random variables, inde-
pendent from N; with probability generating function PM (z): The probability
generating function of
S = M0 + M1 + M2 + : : : + MN ;
1
P1 xk
ln(1 x) = k=1 k ; 0<x<1
CHAPTER 2. THE NUMBER OF CLAIMS 11

with M0 0; is
PS (z) = PN (PM (z)):
A possible interpretation consist on considering N the number of accidents
and Mi the number of claims from accident i. S would represent the total
number of claims.

Example 2.1 Let N and fMi g have Poisson distribution, with parameters
1 and 2 respectively. The probability generating function of S is

1( e 2 (z 1) 1)
PS (z) = e

called Poisson-Poisson or Neyman Type A. J

Let pn = PrfN = ng, n = 0; 1; 2; : : :, fn = PrfMi = ng; n = 0; 1; 2; : : :and

gn = PrfS = ng; n = 0; 1; 2; : : :. Then

X
1
gk = PrfS = kg = PrfM0 + M1 + : : : + Mn = kjN = ng PrfN = ng
n=0
X
1
= PrfM0 + M1 + : : : + Mn = kg PrfN = ng
n=0
X1
= pn f k n ; i = 0; 1; 2; : : : ;
n=0

n
where f is the n-fold convolution of f:

Theorem 2.4 (Panjer recursion formula) For the model just described and
if N is a member of the (a; b; 0) family,

1 Xk
j
gk = a+b fj gk j ; k = 1; 2; : : : (2.5)
1 af0 j=1 k
g0 = PN (f0 ): (2.6)

Proof. From (2.1),

npn = a(n 1)pn 1 + (a + b)pn 1: (2.7)

As
X
1
n
PS (z) = PN (PM (z)) = p n PM (z);
n=0
CHAPTER 2. THE NUMBER OF CLAIMS 12

then
X
1
n 1
PS0 (z) = npn PM 0
(z)PM (z);
n=0
n 1 0
and multiplying both members of (2.7) by PM (z)PM (z) and summing we
get
PS0 (z) = aPM (z)PS0 (z) + (a + b)PM0
(z)PS (z): (2.8)
But
X
1
PS (z) = z i gi
i=0

and
X
1
PS0 (z) = iz i 1 gi ;
i=1

hence (2.8) is equivalent to

X
1 X
1 X
1 X
1 X
1
i 1 j k 1 j 1
iz gi = a z fj kz gk + (a + b) jz fj z k gk
i=1 j=0 k=1 j=1 k=0

which is equivalent to
X
1 X
1 X
1 1 X
X 1
i 1 k+j 1
iz gi = a kz fj gk + (a + b) jz k+j 1 fj gk : (2.9)
i=1 k=1 j=0 k=0 j=1

As the coe¢ cien z i 1

must be the same in both sides of (2.9), we get

X
i X
i
igi = a (i j)fj gi j + (a + b) jfj gi j
j=0 j=0

X
i X
i
= aif0 gi + ai f j gi j +b jfj gi j ;
j=1 j=1

from where we obtain (2.5).

Para terminar a demonstração note-se que
X
1 X
1
g0 = pn PrfM0 + M1 + : : : + Mn = 0g = pn f0n = PN (f0 ):
n=0 n=0
CHAPTER 2. THE NUMBER OF CLAIMS 13

Example 2.2 When the primary distribution is Poisson with parameter ;

S is a compound Poisson distribution and we get

X
k
gk = jfj gk j ; k = 1; 2; : : : (2.10)
k j=1
g0 = exp( (1 f0 ): (2.11)

When the secondary distribution is speci…ed, let’s say with distribution X,

we say that S has a Poisson-X distribution.J

Theorem 2.4 was proved by (Panjer 1981), initially with f0 = 0, though

(2.10) had been obtained previously by (Adelson 1966).

Theorem 2.5 Suppose that Si has a compound Poisson distribution with

Poisson parameter i and secondary distribution fqi;n : n = 0; 1; 2; :::g for
i = 1; 2; 3; :::; k: Suppose also that S1; S2 ; :::; Sk are independent random vari-
ables. Then S = S1 + ::: + Sk also has a compound Poisson distribution with
parameter = 1 + ::: + k and secondary distribution fqn : n = 0; 1; 2; :::g
where qn = [ 1 q1;n + ::: + n qk;n ]= :

Proof. Let Qi (z) bePthe generating function associated to fqi;n : n =

0; 1; 2; :::g; i.e. Qi (z) = 1
n=0 z n
q i;n : Then

Y
k Y
k
Pk
i [Qi (z) 1] i [Qi (z) 1]
PS (z) = PSi (z) = e =e i=1

i=1 i=1
Pk Pk iQ
i Qi (z) i (z) Q(z)
= e i=1 =e i=1 =e ;

where Q(z) is the probability generating function associated to fqn : n =

0; 1; 2; :::g .
When N is a member of the (a; b; 1) family the theorem may be adapted;
and we obtain:

Theorem 2.6 For the model here described and when N is a member of the
(a; b; 1) family,
Pk
(p1 (a + b)p0 )fk + j=1 (a + bj=k) fj gk j
gk = ; k = 1; 2; : : : (2.12)
1 af0
g0 = PN (f0 ): (2.13)
CHAPTER 2. THE NUMBER OF CLAIMS 14

The members of the (a; b; 0) family may me regarded as distributions that

have as probability generating function, a function of the form

PN (z; ) = B( (z 1)); (2.14)

where is a parameter and B(:) is a function independent of . In the Poisson

case, = and B(x) = ex ; for the binomial = p and B(x) = (1 + x)K and
for the negative binomial = and B(x) = (1 x) r .

Theorem 2.7 If PN [z; ] satis…es (2.14) for given and B(z) independent
of ; then PS (z) = PN [PM (z); )] can be written as
T
PS (z) = PN [PM (z); (1 f0 )]; (2.15)
T
where PM (z) is the p.g.f. of the secondary distribution truncated at the origin.

Proof.
PS (z)=PN [PM (z); ]
=PN [E[z Mi ; ]]
=PN E[z 0 ; ] PrfMi = 0g + E[z Mi jMi > 0; ] PrfMi > 0g
T
=PN f0 + (1 f0 )PM (z);
T
=B f0 + (1 f0 )PM (z) 1
T
=B (1 f0 ) P M (z) 1
T
=PN [PM (z); (1 f0 )]:

This result shows that changing the probability at the origin in the second-
ary distribution does not create a new compound distribution (only changes
the parameter).
An important case of the compound distribution is the Poisson-extended
truncated negative binomial. When r = 1 we get the Poisson-truncated
geometric; when r = 0 we get the Poisson-logarithmic and when r = 0:5
we get the Poisson - Inverse Gaussian.

2.4 Mixed frequency distributions

As we have already seen the negative binomial can be regarded as a mixture
of the Poisson with the Gamma. In general
CHAPTER 2. THE NUMBER OF CLAIMS 15

Z
pk = pk ( )u( )d

if is a continuous random variable or

X
pk = pk ( )u( )

if is a discrete random variable or more generally

Z
pk = pk ( )dU ( ):

The probability generating function is

Z
PN (z) = PN j = (z)dU ( )

Example 2.3 The zero-modi…ed distribution can be created by using a two

point mixture, of a generate distribution that places all probability at zero and
a distribution with the original probability function.

Example 2.4 Determine the p.f. of a mixed binomial with a beta mixing
distribution (called binomial-beta).
Z 1 m k (a + b) a
pk = q (1 q)m k
q 1
(1 q)b 1
dq =
0 k (a) (b)
Z 1
(a + b) (m + 1) (k + a) (m + b k) (a + m + b)
= q k+a 1
(1 q)m+b k 1
dq =
(a) (b) (k + 1) (m k + 1) (a + m + b) 0 (k + a) (m + b k)
a+k 1 b+m k 1
(a + b) (m + 1) (k + a) (m + b k) k m k
= = a+b+m 1
=
(a) (b) (k + 1) (m k + 1) (a + m + b)
m
a b
k m k
= a b
; k = 0; 1; 2::
m

because by de…nition

n n+k 1
= ( 1)k :
k k

Homework 2.1 Show that the composition of a Poisson with the ETBNB
with r = 0:5 can be obtained as a mixture of the Poisson with the inverse
Gaussian.
CHAPTER 2. THE NUMBER OF CLAIMS 16

2.5 E¤ect of exposure on frequency

The claim frequency, N; is the number of claims of a portfolio (or sub-
portfolio) of policies in a given time period. We should expect N to in-
crease with the exposure (number of lives, number policies, square meters
of insured buildings,...). Suppose that the portfolio consists of n entities,
each of them producing claims Nj in the period under consideration. Then
N = N1 + N2 + : : : + Nn . If we suppose that Nj ate i.i.d., then

PN (z) = [PN1 (z)]n : (2.16)

If instead of n there were n entities, then N ; would have probability

generating function

PN (z) = [PN1 (z)]n = [PN (z)]n =n

: (2.17)
p
When a counting random variable N is such that for all n, n P (z) is
still a probability generating function we say that N is in…nitely divisible. In
that case N will have the same form as N but with di¤erent parameter.
Both the Poisson and the negative binomial are in…nitely divisible.

Example 2.5 Consider a health plan for a group of 300 teachers of a school,
and suppose that the number of claims of the group is considered to follow
a negative binomial with parameters r= 10 and = 3. The distribution of
the number of claims for another similar group of 450 teachers could still be
considered negative binomial with the same and r= 15 = 10 450=300: J

2.6 Model selection

The 2 goodness-of-…t test is an adequate test to study the …tness of di¤erent
discrete distributions to data.
Let’s consider a random sample from a population N; and suppose that
we want to test the hypothesis that the probability function of N is a given
one with r unknown parameters 1 ; :::; r . We classify the sample in m + 1
classes, numbered from 0 to m. Let nk be the number of observations in class
k and let p k ( 1 ; :::; r ) be the probability that N takes values in class k; with
k = 0; 1; : : : ; m: The statistic
2
X
m nk npk ( ^1 ; : : : ; ^r )
2
( ^1 ; : : : ; ^r ) = ; (2.18)
k=0 npk ( ^1 ; : : : ; ^r )
CHAPTER 2. THE NUMBER OF CLAIMS 17

where ( ^1 ; : : : ; ^r ) are the maximum likelihood estimators, obtained with the

classi…ed data, is asymptotically distributed according to a 2 with m r
degrees of freedom when the hypothesis is true.
The maximum likelihood estimates are obtained maximizing in 1 ; :::; r
the function
Y
m
[pk ( 1 ; :::; r )]nk : (2.19)
k=0

The application of the test requires that the expected number of elements
in each class is high enough. In the exercises we will consider the value of 5.
Note that this is slightly di¤erent than maximizing the likelihood function
with the original sample.

Table 2.2: Observed frequency of an automobile portfolio

k nk
0 96 978
1 9 240
2 704
3 43
4 9
>4 0
Total 106974
CHAPTER 2. THE NUMBER OF CLAIMS 18

Table 2.3: Observed and …tted distribution - Poissn model

k nk nPk
0 96 978 96696.46
1 9 240 9767.21
2 704 493.29
3 52 17.04
2
106 974 =191.04
^ =0.101009

Table 2.4: Observed and …tted frequency - negative binomial model.

k nk nPk
0 96 978 96977,79
1 9 240 9240,88
2 704 702,89
3 52 52,44
2
106 974 = 0:005479
rb=1.67651
^ =0.060262782

Table 2.5: Observed frequency

k nk
0 103 704
1 14 075
2 1 766
3 255
4 45
5 6
6 2
7 ou mais 0
Total 119 853
CHAPTER 2. THE NUMBER OF CLAIMS 19

Table 2.6: Fitted models 2.5

Number of claims Observed frequency Fitted distribution

Poisson Negative B. P.I.G P.- E.T.N.B.
0 103 704 102 673.07 103 722.68 103 709.92 103 704.58
1 14 075 15 885.18 13 992.95 14 054.97 14 072.58
2 1 766 1 228.85 1 855.67 1 784.79 1 769.24
3 255 244.67 254.43 255.22
4 45 40.40 41.98
other 8 65.90 37.02 8.49 9.41
Parameters b =0.154716 rb =1.03518 b =0.144668 b =0.144720
b =0.149851347 b =0.310447361 rb =-0.55169
b =0.3480422
Chi-square 1 340.776 12.146 0.779 0.435
Degrees of freedom 2 2 3 2
p value 0.000000 0.002304 0.854595 0.804631
-Log Likelihood 54 888.98 54 587.55 54 604.73 54 604.57
Chapter 3

Impact of coverage
modi…cations

3.1 Deductibles
De…nition 3.1 In an insurance policy, the deductible is the amount of the
loss that must be reached before an insurer will cover any expenses.

De…nition 3.2 If an ordinary deductible is in force, the insurer will pay the
value of the loss in excess of the deductible.

In presence of an ordinary deductible the insurer is responsible for the

loss
0 X d
YL =
X d X > d:
When a loss greater than d happens the insurer will pay Y P = Y L jY L > 0:
The per-payment variable is the per-loss variable given that there was a
payment.
The density function of Y P is
fX (x + d)
fY P (x) = ; x > 0;
SX (d)
its survival function is
SX (x + d)
SY P (x) = ; x > 0;
SX (d)
the distribution function is
FX (x + d) FX (d)
FY P (x) = ; x > 0;
SX (d)

20
CHAPTER 3. IMPACT OF COVERAGE MODIFICATIONS 21

and the hazard rate function is

fX (x + d)
hY P (x) = = hX (x + d); x > 0:
SX (x + d)

The random variable Y L is of a mixed type. It has a mass point in zero

of value FX (d) and it is continuous afterwards.

FX (d) x=0
fY L (x) =
fX (x + d) x > 0:

FY L (x) = FX (x + d); x 0;
and
SY L (x) = SX (x + d); x 0:

De…nition 3.3 A franchise deductible is a threshold which needs to be ex-

ceeded in order for the insurer to be liable for the entirety of the claim.

In this case
0 X d
YL =
X X > d:
and Y P = Y L jY L > 0:

FX (d) x = 0
fY L (x) =
fX (x) x > d:

SX (d) 0 x d
SY L (x) =
SX (x) x > d:

FX (d) 0 x d
FY L (x) =
FX (x) x > d:
0 0<x<d
hY L (x) =
hX (x) x > d:
and for Y P
fX (x)
fY P (x) = SX (d)
x > d:
(
1 0 x d
SY P (x) = SX (x)
SX (d)
x > d:
CHAPTER 3. IMPACT OF COVERAGE MODIFICATIONS 22

(
0 0 x d
FY P (x) = FX (x) FX (d)
SX (d)
x > d:

0 0<x d
hY P (x) =
hX (x) x > d:

Homework 3.1 X is a Pareto ( ; ): Determine these quantities for any d:

Proposition 3.1 For any ordinary deductible, the expected cost per loss is
Z 1 Z 1
E[X] E[X ^ d] = (1 FX (x))dx = SX (x)dx
d d

and the expected cost per payment is

R1
E[X] E[X ^ d] d
SX (x)dx
= = eX (d)
1 FX (d) SX (d)

Proposition 3.2 For any franchise deductible, the expected cost per loss

E[X] E[X ^ d] + d[1 FX (d)]

and the expected cost per payment is

E[X] E[X ^ d]
+d
1 FX (d)

Homework 3.2 Determine the four expectations for the Pareto ( ; ):

De…nition 3.4 The loss elimination ratio is the ratio of the decrease in the
expected payment (it is per loss)with an ordinary deductible to the expected
payment with no deductible.

E[X ^ d]
:
E[X]

Homework 3.3 Determine the loss elimination ratio for the Pareto ( ; ):

When there are deductibles, the e¤ect of in‡ation is magni…ed.

CHAPTER 3. IMPACT OF COVERAGE MODIFICATIONS 23

Proposition 3.3 For an ordinary deductible of d; after uniform in‡ation of

1 + r; the expected cost per loss is

d
(1 + r) E[X] E X^
1+r
d
and the expected cost per payment, if SX 1+r
> 0 is
d
(1 + r) E[X] E X^ 1+r
d
:
SX 1+r

x
Proof. After in‡ation losses are Y = (1 + r)X; with SY (x) = SX 1+r
:

With an ordinary deductible

0 Y d
YL =
Y d Y >d
R1 R1 x
R1
E[YL ] = d
SY (x)dx = d
SX 1+r
dx = d SX (y) (1 + r)dy
1+r
d
= (1 + r) E[X] E X^ 1+r

Homework 3.4 Calculate the expected cost per loss and per payment after
in‡ation for a Pareto(3; 2000), d = 500 and r = 0:10:

3.2 Policy limits

The opposite of a deductible is a policy limit. Given a loss X the insurance
company pays only up to a limit u; i.e. pays Y = min(X; u):

FX (x) x < u
FY (x) =
1 x u:
With in‡ation we have the following result:

Proposition 3.4 For a policy limit of u; after uniform in‡ation of 1 + r;

the expected cost is
u
(1 + r)E X ^ :
1+r
CHAPTER 3. IMPACT OF COVERAGE MODIFICATIONS 24

3.3 Coinsurance, deductibles and limits

8 d
< 0 X < 1+r
L d u
Y = [(1 + r)X d] 1+r
X < 1+r
: u
(u d) X 1+r

Note that
u d
Y L = (1 + r) X^ X^
1+r 1+r

Hence
u d
E[Y L ] = (1 + r) E X ^ E X^ =
1+r 1+r
Z u
1+r
= (1 + r) SX (x)dx =
d
1+r

E[Y L ]
E[Y P ] = d
:
SX 1+r
Further moments are more di¢ cult to obtain:

Z u
L k
1+r u
E[ Y ] = k
[(1 + r)x d]k fX (x)dx + k
(u d)k SX
d 1+r
1+r
Z u k
k k
1+r d k u
= (1 + r) x fX (x)dx + (u d)k SX
d 1+r 1+r
1+r
8" # u
9
< k 1+r Z u k 1 =
d 1+r d
= ( (1 + r))k x SX (x) + k x SX (x)dx
: 1+r d
d 1+r ;
1+r
1+r

u
+ ( (u d))k SX
1+r
(" # Z u
)
k k 1
u d u 1+r d
= ( (1 + r))k SX + k x SX (x)dx
1+r 1+r 1+r d 1+r
1+r

u
+ ( (u d))k SX
1+r
Z u k 1
k 1+r d
= k ( (1 + r)) x SX (x)dx
d 1+r
1+r
CHAPTER 3. IMPACT OF COVERAGE MODIFICATIONS 25

In the Loss Models book (Klugman, Panjer & Willmot (2008)) it is shown
that
( " # " #
2 2
2 u d
E[ Y L ] = ( (1 + r))2 E X^ E X^
1+r 1+r
d u d d
2 E X^ +2 E X^
1+r 1+r 1+r 1+r

3.4 Impact of deductibles on the claim fre-

quency
In presence of deductibles, the aggregate claim amount can be calculated as
L P
X
N X
N
S= YiL = YiP
i=0 i=0

What is the relation between N L and N P ?

L
X
N
P
N = Ii
i=0

where Ii is a Bernoulli r.v. taking the value 1 whenever a loss results in a

payment. Let’s assume that this probability is p:
If the probability generating function of N L belongs to the (a; b; 0) or
(a; b; 1) class of distributions

PN L (z) = B( (z 1):

In that case
PN L PN L h i
NL
PN P (z) = E z i=0 Ii =E E z i=0 Ii jN L
=E E z I
=

= PN L (zp + (1 p)) = PN L (1 + p(z 1)) = B( (p(z 1)) = B( p(z 1)):

Chapter 4

Aggregate Loss Models

4.1 Collective risk model versus individual

risk model
Given a risk (policy, group of policies, a portfolio,...), in the Collective Risk
Model we add the payments as they are made.

X
N
S = X1 + X2 + ::: + XN = Xi ; (X0 0);
i=0

where fXi gi=1;2;::: are i.i.d. random variables and independent of N:

In the Individual Risk Model, the aggregate loss is the sum of the losses
of n contracts. In this case

S = X1 + X2 + ::: + Xn :

In this model the variables fXi gi=1;2;::: are supposed to be independent

but not identically distributed. The variables Xi usually have a mass point
at 0.

4.2 Assumptions and characteristics of the

compound model
We will use the Collective model. N is called frequency of claims, number of
claims. Xi is the payment, although sometimes referred as loss. S is called
the aggregate loss random variable (but it is the total payment) or aggregate
claims amount.

26
CHAPTER 4. AGGREGATE LOSS MODELS 27

X
1 X
1
Fs (x) = PrfS xg = pk PrfS xjN = kg = pk FXk (x); (4.1)
k=0 k=0

where FXk (x) represents the n fold convolution of the cdf of X: It can be
obtained as

1 x 0
FX0 (s) =
0 x<0
and Z +1
FXk (x) = FXk 1 (x y)dFX (y):
1

The most common case is the case where X is continuous and positive random
variable in which case
Z x
k
FX (x) = FXk 1 (x y)fX (y)dy; k = 1; 2; :::
0

In this case
X
1
fS (x) = pk fXk (x); x > 0; (4.2)
k=0

and PrfS = 0g = p0 :

MS (r) = PN [MX (r)]

= MN [ln MX (r)]:

E[S] = S = E[N ]E[X];

2
Var[S] = S = E[N ]Var[X] + Var[N ]E 2 [X];

3
3 [S] = 3 [N ]E [X] + 3Var[N ]E[X]Var[X] + E[N ] 3 (X):

4.3 Special cases

4.3.1 Compound Poisson Model
(MX (r) 1)
MS (r) = e : (4.3)
CHAPTER 4. AGGREGATE LOSS MODELS 28

E[S] = E[X];

Var[S] = E[X 2 ]; (4.4)

E[X 3 ]
S = ( E[X 2 ])3=2
:

4.3.2 Compound Mixed Poisson Model

MS (r) = M [(MX (r) 1)];

E[S] = E[ ]E[X];
Var[S] = E[ ]E[X 2 ] + Var[ ]E 2 [X]; (4.5)

E[ ]E[X 3 ]+3Var[ ]E[X]E[X 2 ]+ 3[ ]E 3 [X]

S = (Var[S])3=2
:

4.4 Analytic results

Example 4.1 Exponential severity When X is exponential with para-
meter ; FXk (x) is the Gamma distribution with parameters (k; ): From its
relation with the Poisson
X1
exp( x= )(x= )j X
k 1
exp( x= )(x= )j
FXk (x) = =1
j=k
j! j=0
j!

Hence
X
1 X
1 X
k 1
exp( x= )(x= )j
Fs (x) = pk FXk (x) =1 pk =
k=0 k=1 j=0
j!
X1
(x= )j X
1
= 1 exp( x= ) pk
j=0
j! k=j+1

If N has a Geometric distribution

j+1
X
1 X
1
1
k
1 1+
j+1
pk = = 1 = :
k=j+1 k=j+1
1+ 1+ 1+ 1+
1+
CHAPTER 4. AGGREGATE LOSS MODELS 29

Then
X1
(x= )j
j+1
Fs (x) = 1 exp( x= ) =
j=0
j! 1+
X1
(x= )j
j
= 1 exp( x= ) =
1+ j=0
j! 1+
x
= 1 exp( x= ) exp =
1+ 1+
x 1
= 1 exp :
1+ 1+
1
FS (0) = 1+
and

x 1
fS (x) = ; x>0
(1 + )2 1+
i.e. FS (x); has a mass point at zero and it is exponential hereafter.
Alternatively
MS (r) = PN (MX (r)) = [1 (MX (r) 1)] 1 =
1 1 1 r
= [1 ((1 r) 1)] = =
1 r r
1 1
= + (1 (1 + ) r)
1+ 1+
which corresponds to the moment generating function of a mixture of a degen-
erate random variable taking the value 0 and of an exponential with parameter
(1 + ) ; being 1+1 the weight at the zero value.

Theorem 4.1 Suppose that Sj has a compound Poisson distribution with

Poisson parameter j and severity distribution with cdf FXj (x); j = 1; 2; :::; n:
Suppose that S1 ; S2 ; :::; Sn are independent. Then S = S1 + S2 + :::+ Sn is
compound Poisson with Poisson parameter = 1 + ::: + n and severity
distribution with cdf
X n
j
FX (x) = FXj (x)
j=1

Corollary 4.1.1 Let x1 ; x2 ; : : : ; xn be di¤erent numbers and suppose that

N1 ; N2 ; : : : ; Nn are independent r.v. each of them Poisson distributed with
parameter i . Then
S = x1 N1 + x2 N2 + : : : + xn Nn
CHAPTER 4. AGGREGATE LOSS MODELS 30

is Compound Poisson with

X
n
= i
i=1

and
i
x = xi ; i = 1; 2; : : : ; m
fX (x) =
0 other. values

Homework 4.1 Consider a group life insurance, for simpli…cation with just
4 classes. The following table gives the number of lives nk in each class, the
bene…t bk and the death probability qk . Let S be the aggregate claims.

k q k bk nk
1 0.02 1 10
2 0.02 4 20
3 0.05 1 30
4 0.05 4 40

1. Considering that the 100 lives are independent, determine S, S and

S.

2. Compare these values with the values obtained by approximating N to

a Poisson (keeping the mean).

3. Calculate an approximation to the probability that S is greater than

twice its mean.

4.5 The aggregate claim distribution

4.5.1 Recursive method (Panjer’s recursion formula)
Theorem 4.2 If N is a member of the (a; b; 1) and X takes values on the
non-negative integers, the probability function of S, fS , satis…es
Pi
[p1 (a + b)p0 ]fX (i) + j=1 (a + bj=i) fX (j)fS (i j)
fS (i) = ; i = 1; 2; :(4.6)
::
1 afX (0)
fS (0) = PN (fX (0)): (4.7)
CHAPTER 4. AGGREGATE LOSS MODELS 31

Note that
p0 P if fX (0) = 0
fS (0) = (4.8)
PN (fX (0)) = 1 p f
k=0 k X
k
(0) if fX (0) > 0:
If the support of fX (:) is limited to r then equation (4.6) is
P
[p1 (a + b)p0 ]fX (i) + i^rj=1 (a + bj=i) fX (j)fS (i j)
fS (i) = ; i = 1; 2; : : :
1 afX (0)
(4.9)
Note also that some of the fX ’s may be zero.
Corollary 4.2.1 If N is a member of the (a; b; 0) family and X takes values
on the non-negative integers, the probability function of S, fS , satis…es

1 Xi
j
fS (i) = a+b fX (j)fS (i j); i = 1; 2; : : :(4.10)
1 afX (0) j=1 i
fS (0) = PN (fX (0)): (4.11)
Example 4.2 The number of claims is Poisson with mean 0.1. Each claim
takes the values 5 000 euros and 10 000 euros with probabilities 0.8 and 0.2
respectively. Determine the probability that the aggregate claims is grater
than 30 000 euros.
Choosing 5 000 euros for monetary unit we have that fX (1) = 0:8 and
fX (2) = 0:2. As = 0:1we get
fS (0)=e = 0:904837418
fS (1)= fX (1)fS (0) = 0:1 0:8 0:904837418 = 0:072386993
fS (2)= 2 [fX (1)fS (1) + 2fX (2)fS (0)] = 0:020992228
fS (3)= 3 [fX (1)fS (2) + 2fX (2)fS (1)] = 0:001524953
fS (4)= 4 [fX (1)fS (3) + 2fX (2)fS (2)] = 0:000240421:

i fS (i) FS (i)
0 0.904837418 0.904837418
1 0.072386993 0.977224411
2 0.020992228 0.998216640
3 0.001524953 0.999741592
4 0.000240421 0.999982014
5 0.000016046 0.999998060
6 0.000001817 0.999999877
The required probability is 0.000000123. J
CHAPTER 4. AGGREGATE LOSS MODELS 32

When the frequency distribution is a compound distribution involving

(a; b; 0) or (a; b; 1); distributions, the recursive formulas may be used more
than once to calculate the probabilities of the aggregate claims.

R code to compute fS (i)

Panjer.Poisson <- function (fx, lambda)
{ if (lambda * sum(fx) > 727) stop("Under‡ow; recursion fails")
cumul <- f <- exp(-lambda * sum(fx)) ## sum(fx) = 1-fx(0)
r <- length(fx)
x <- 0
repeat
{ x <- x+1
m <- min(x, r)
last <- lambda / x * sum(1:m * head(fx, m) * rev(tail(f, m)))
f <- c(f, last)
cumul <- cumul + last
if (cumul > 0.99999999) break }
return(f) }
Panjer.Poisson(c(0.8,0.2), 0.1)

R code with actuar (in actuar fx must start by the prob at 0)

## Recursive method
fx<- c(0,0.8,0.2)
Fs <- aggregateDist("recursive", model.freq = "poisson",
model.sev = fx, lambda = 0.1, x.scale = 1)
Fs(knots(Fs)) # cdf evaluated at its knots
di¤(Fs) # probability mass function

Example 4.3 (example 9.12-book) The number of claims has a Poisson-

ETNB distribution with Poisson parameter = 2 and ETNB parameters
= 3 and r = 0:2: The claim size distribution has probabilities 0.3, 0.5 and
0.2 at 0, 10 and 20 respectively. Calculate P r(S > 40).
In monetary units of 10, we have that fX (0) = 0:3; fX (1) = 0:5 and
fX (2) = 0:2:

PS (z) = P1 [P2 (PX (z))] = P1 (PS1 (z))

PS1 (z) = P2 (PX (z))
CHAPTER 4. AGGREGATE LOSS MODELS 33

where P1 is the pgf of the Poisson, P2 is the pgf of the ETNB and PX is the
pgf of the severity distribution.

[1 (fX (0) 1) r (1 + ) r
fS1 (0) = P2 (fX (0)) = =
1 (1 + ) r
[1 3(0:3 1) 0:2 (1 + 3) 0:2
= = 0:16369:
1 (1 + 3) 0:2
r
As a = 1+ = 0:75 and b = (r 1)a = 0:6; p0 = 0; p1 = (1+ )r+1 (1+ )
=
0:46947; we have that

Pmin(i;2)
[p1 (a + b)p0 ]fX (i) + j=1 (a + bj=i) fX (j)fS1 (i j)
fS1 (i) = =
1 afX (0)
Pmin(i;2)
0:46947 fX (i) + j=1 (0:75 0:6j=i) fX (j)fS1 (i j)
= ; i = 1; 2; : : :
0:775

fS1 (1) = 0:31873

fS1 (2) = 0:22002
fS1 (3) = 0:10686
fS1 (4) = 0:06692

Now we apply the recursion of the Poisson with S1 and we get

fS (0) = exp( (fS1 (0) 1) = 0:18775

2X
i
fS (i) = jfS1 (j)fS (i j); j = 1; 2; :::
i j=1

fS (1) = 0:11968
fS (2) = 0:12076
fS (3) = 0:10090
fS (4) = 0:08696
FS (4) = 0:61605
Pr(S > 4) = 0:38395:
CHAPTER 4. AGGREGATE LOSS MODELS 34

R code with actuar

## Recursive method
fx<- c(0.3,0.5,0.2)
Fs <- aggregateDist("recursive", model.freq = "negative binomial", model.sev
= fx, size = 0.2,prob=0.25,p0=0, x.scale = 1)
Fs(knots(Fs)) # cdf evaluated at its knots
di¤(Fs) # probability mass function
Gs <- aggregateDist("recursive", model.freq = "poisson", model.sev =
di¤(Fs), lambda=2, x.scale = 1)
Gs(knots(Gs)) # cdf evaluated at its knots
di¤(Gs) # probability mass function

4.5.2 Constructing Arithmetic Distributions

In most cases the claim size is a continuous random variable. We can discret-
ize it, in such a way that the discretized distribution places the probabilities
on multiples of a convenient unit of measurement h; the span (this is equi-
valent to change the original monetary unit to h times the original monetary
unit). Let X d be the new random variable and
fj = PrfX d = jg j = 0; 1; 2; : : :
its pf.

Method of rounding down

f0A = FX (h 0);
(4.12)
fjA = FX (h(j + 1) 0) FX (hj 0); j = 1; 2; : : :

Method of rounding to nearest

h
f0B = FX 2
0 ;
(4.13)
1 1
fjB = FX h j + 2
0 FX h j 2
0 ; j = 1; 2; : : : :

Method of rounding up
f0C = 0;
(4.14)
fjC = FX (hj + 0) FX (h (j 1) + 0) ; j = 1; 2; : : : :
Denoting by F B (y); F B (y) and F C (y) the corresponding distribution
functions, it is obvious that
F A (y) FX (y) F C (y); y 0; (4.15)
CHAPTER 4. AGGREGATE LOSS MODELS 35

and that
F A (y) F B (y) F C (y); y 0: (4.16)
Consequently
F A (y) F (y) F C (y); y 0; (4.17)
and
F A (y) F B (y) F C (y); y 0: (4.18)

Method that matches the mean of the distribution

Let fjD be such that
Z j+1
f0D + f1D + ::: + fjD = FX (hy)dy; j = 0; 1; 2; : : : ; (4.19)
j

i.e. Z j+1
D
F (j) = FX (hy)dy; j = 0; 1; 2; : : : : (4.20)
j

If (4.20) holds than

Z j+1
D
1 F (j) = (1 FX (hy))dy; j = 0; 1; 2; : : :
j

and
X
k Z k+1
D
(1 F (j)) = (1 FX (hy))dy:
j=0 0

Calculating the limit as k goes to in…nity

X
1 Z 1
D
(1 F (j)) = (1 FX (hy))dy;
j=0 0

i.e. the expected value of the discretized distribution is equal to the expected
value of X=h.

Example 4.4 Consider a compound Poisson distribution with = 1 and

FX (x) = 1 1= (1 + x)4 with x > 0 (FX (x) is a Pareto with = 1 and
= 4). Determine a lower and an upper limit for the aggregate claim
CHAPTER 4. AGGREGATE LOSS MODELS 36

amount, as well as two approximations, using a span of h = 0:25. We obtain

4 4
1 1
fjA = 1+jh 1+jh+h
; j = 0; 1; 2; : : :
8 4 4
>
> 1 1
; j = 1; 2; : : :
< 1+jh h=2 1+jh+h=2
B
fj =
>
> 4
: 1 1
; j=0
8 1+h=2
> 4 4 (4.21)
< 1 1
; j = 1; 2; : : :
1+jh h 1+jh
fjC =
>
: 0; j = 0
8 1 1 2
< 3h(1+jh+h)3 + 3h(1+jh h)3 3h(1+jh)3 ; j = 1; 2; : : :
fjD =
: 1+ 1 1
; j=0
3h(1+h)3 3h

With h = 0:25 we obtain table 4.1. Applying (4.10) we obtain Table 4.2. J

R code with actuar

#Upper, lower and …rst moment matching
require(actuar)
#Upper, lower and …rst moment matching
fux<- discretize(ppareto(x,4,1),method="upper", from=0, to =50,step=0.25)
‡x<- discretize(ppareto(x,4,1),method="lower", from=0, to =50,step=0.25)
fcx<- discretize(ppareto(x,4,1),method="unbiased",lev=levpareto(x,4,1),
from=0, to =50,step=0.25)
par(col="black")
curve(ppareto(x,4,1),xlim=c(0,5))
x<- seq(0,50,0.25)
par(col="blue")
plot(stepfun(head(x,-1),di¢ nv(fux)),pch=19,add=TRUE)
par(col="green")
plot(stepfun(x,di¢ nv(‡x)),pch=19,add=TRUE)
par(col="red")
plot(stepfun(x,di¢ nv(fcx)),pch=19,add=TRUE)
Fs<- aggregateDist(method = "recursive", model.freq = "poisson", model.sev
= fcx, x.scale = 0.25, lambda = 1)
for(u in x) print(Fs(u))
CHAPTER 4. AGGREGATE LOSS MODELS 37

Table 4.1: Discretization of the claim size distribution

j gjA gjB gjC gjD

0 0.5904000 0.3757049 0.0000000 0.3493333

1 0.2120691 0.3445328 0.5904000 0.3630617
2 0.0909086 0.1363500 0.2120691 0.1413284
3 0.0441222 0.0625037 0.0909086 0.0641579
4 0.0234816 0.0318671 0.0441222 0.0325072
5 0.0134184 0.0176115 0.0234816 0.0178893
6 0.0081149 0.0103689 0.0134184 0.0105009
7 0.0051395 0.0064243 0.0081149 0.0064917
8 0.0033824 0.0041511 0.0051395 0.0041876
9 0.0022994 0.0027784 0.0033824 0.0027992
10 0.0016071 0.0019162 0.0022994 0.0019285
11 0.0011505 0.0013560 0.0016071 0.0013636
12 0.0008412 0.0009813 0.0011505 0.0009862
13 0.0006264 0.0007243 0.0008412 0.0007275
14 0.0004743 0.0005440 0.0006264 0.0005461
15 0.0003644 0.0004150 0.0004743 0.0004164
16 0.0002837 0.0003210 0.0003644 0.0003220
17 0.0002235 0.0002514 0.0002837 0.0002522
18 0.0001780 0.0001992 0.0002235 0.0001997
19 0.0001432 0.0001595 0.0001780 0.0001599
20 0.0001162 0.0001289 0.0001432 0.0001292
21 0.0000952 0.0001051 0.0001162 0.0001053
22 0.0000785 0.0000863 0.0000952 0.0000865
23 0.0000652 0.0000715 0.0000785 0.0000716
24 0.0000545 0.0000596 0.0000652 0.0000597
25 0.0000459 0.0000500 0.0000545 0.0000501
26 0.0000388 0.0000422 0.0000459 0.0000423
27 0.0000331 0.0000358 0.0000388 0.0000359
28 0.0000283 0.0000306 0.0000331 0.0000306
29 0.0000243 0.0000262 0.0000283 0.0000262
30 0.0000210 0.0000226 0.0000243 0.0000226
31 0.0000182 0.0000195 0.0000210 0.0000195
32 0.0000158 0.0000170 0.0000182 0.0000170
33 0.0000138 0.0000148 0.0000158 0.0000148
34 0.0000121 0.0000129 0.0000138 0.0000129
35 0.0000107 0.0000114 0.0000121 0.0000114
36 0.0000094 0.0000100 0.0000107 0.0000100
37 0.0000083 0.0000088 0.0000094 0.0000089
38 0.0000074 0.0000078 0.0000083 0.0000078
39 0.0000066 0.0000070 0.0000074 0.0000070
40 0.0000059 0.0000062 0.0000066 0.0000062
41 0.0000053 0.0000056 0.0000059 0.0000056
42 0.0000047 0.0000050 0.0000053 0.0000050
43 0.0000042 0.0000045 0.0000047 0.0000045
44 0.0000038 0.0000040 0.0000042 0.0000040
45 0.0000034 0.0000036 0.0000038 0.0000036
46 0.0000031 0.0000033 0.0000034 0.0000033
47 0.0000028 0.0000030 0.0000031 0.0000030
48 0.0000026 0.0000027 0.0000028 0.0000027
49 0.0000023 0.0000024 0.0000026 0.0000025
50 0.0000021 0.0000022 0.0000023 0.0000022
CHAPTER 4. AGGREGATE LOSS MODELS 38

Table 4.2: Approximating F

i FiA FiB FiC FiD

0 0.663916 0.535639 0.367880 0.521698

1 0.804712 0.720184 0.585076 0.711106
2 0.879997 0.825009 0.727208 0.819221
3 0.923145 0.887302 0.819330 0.883622
4 0.949104 0.925535 0.879038 0.923180
5 0.965320 0.949636 0.917961 0.948113
6 0.975773 0.965194 0.943566 0.964196
7 0.982693 0.975457 0.960600 0.974794
8 0.987386 0.982365 0.972074 0.981919
9 0.990637 0.987103 0.979906 0.986798
10 0.992934 0.990412 0.985324 0.990200
11 0.994585 0.992761 0.989124 0.992612
12 0.995792 0.994456 0.991825 0.994349
13 0.996688 0.995696 0.993772 0.995619
14 0.997362 0.996617 0.995194 0.996560
15 0.997876 0.997310 0.996245 0.997268
16 0.998272 0.997837 0.997032 0.997806
17 0.998582 0.998244 0.997629 0.998220
18 0.998825 0.998561 0.998086 0.998542
19 0.999020 0.998810 0.998440 0.998796
20 0.999176 0.999008 0.998717 0.998997
21 0.999302 0.999167 0.998936 0.999158
22 0.999405 0.999296 0.999110 0.999289
23 0.999490 0.999401 0.999251 0.999395
24 0.999560 0.999487 0.999365 0.999482
25 0.999619 0.999558 0.999458 0.999554
26 0.999668 0.999617 0.999535 0.999614
27 0.999710 0.999667 0.999598 0.999665
28 0.999745 0.999709 0.999652 0.999707
29 0.999775 0.999745 0.999696 0.999743
30 0.999801 0.999775 0.999734 0.999773
31 0.999823 0.999801 0.999766 0.999799
32 0.999842 0.999823 0.999793 0.999822
33 0.999859 0.999842 0.999817 0.999841
34 0.999873 0.999859 0.999837 0.999858
35 0.999886 0.999874 0.999854 0.999873
36 0.999897 0.999886 0.999870 0.999886
37 0.999907 0.999897 0.999883 0.999897
38 0.999916 0.999907 0.999895 0.999907
39 0.999923 0.999916 0.999905 0.999915
40 0.999930 0.999924 0.999914 0.999923
41 0.999936 0.999930 0.999922 0.999930
42 0.999942 0.999936 0.999929 0.999936
43 0.999947 0.999942 0.999935 0.999942
44 0.999951 0.999947 0.999941 0.999946
45 0.999955 0.999951 0.999946 0.999951
46 0.999958 0.999955 0.999950 0.999955
47 0.999962 0.999959 0.999954 0.999958
48 0.999964 0.999962 0.999958 0.999962
49 0.999967 0.999965 0.999961 0.999965
50 0.999969 0.999967 0.999964 0.999967
CHAPTER 4. AGGREGATE LOSS MODELS 39

4.6 Approximations
For large portfolios the probability PrfS = 0g is very small, which implies
that the recursive method will have some problems. On the other hand we
may not know the severity distribution, but only the knowledge of the …rst
few moments. In these cases when S is small, the normal approximation
gives good results. This does not happen if S is big. For S > 0:1 the
errors may be signi…cant on the tail of the distribution. In this case it is
preferable to use the NP approximation (normal power) or the translated
gamma approximation.

4.6.1 The NP approximation

Let
S S
Z=
S

FZ (z) = FS ( S +z S) :

The NP approximation is based on a formula known as the Edgeworth series,

2
E(Z 3 ) (3) E(Z 4 ) 3 (4) [E(Z 3 )] (6)
FZ (z) = (z) (z) + (z) + (z) + :::;
6 24 72
where (z) is the distribution function of the standard normal. We will
consider only two terms
S (3)
FZ (z) (z) (z)
6
00
As (z) = z (z) and (3) (z) = (z 2 1) (z), where (z) is the density
function of the standard normal, then
S
FZ (z) (z) (z 2 1) (z);
6
and
FZ (z + z) FZ (z) + FZ0 (z) z
(z) 6
S
(z 2 1) (z) + (z) z
= (z) 6
S
(z 2 1) z (z):
Let
S
z= (z 2 1);
6
from where
S
FZ z + (z 2 1) (z); (4.22)
6
CHAPTER 4. AGGREGATE LOSS MODELS 40

i.e. solving
S
z+ (z 2 1) = y
6
in z, s !
3 9 6
FZ (y) + 2
+1+ y :
S S S

Which is equivalent to
s !
3 9 6 x S
FS (x) + 2
+1+ : (4.23)
S S S S

x
This approximation should be applied only if S
S
> 1. For other values see
(Beard, Pentikäinen & Pesonen 1984).

4.6.2 Translated Gamma Approximation

It consists in approximating S by a r.v.. k + Y , where k is a constant and Y
is a Gamma with parameters and , with k, and chosen in such a way
that S and k + Y have the same mean, variance and skewness coe¢ cient, i.e.
k, and solve the following system:

S =k + ;

2 2
S= ; (4.24)

p2
S= :

Example 4.5 S is compound Poisson with = 50 and FX (x) = 1

1= (1 + x)4 , x > 0. Determine s such that

PrfS sg 0:95;

using

a) the normal approximation

b) the NP approximation

c) the Translated Gamma approximation

CHAPTER 4. AGGREGATE LOSS MODELS 41

The kth raw moment of a Pareto (4,1) is

(4 k)
ak = k! ; > k:
(4)

Then
50
S= 3 ;

2 50
S= 3 ;

q
27
S= 50
:
Hence,

a) let
S S
Z=
S
Using the normal approximation

FZ (z) (z)

and denoting by z the value of z such that (z) = 1 we have that

s= S +z S:

As 1 = 0:95, z = 1:645, from where s=23.38235.

b) in this case
S
FZ z + (z 2 1) (z)
6
so
S
s= S + z + (z 2 1) S:
6
from where s = 24:23536.

c) attending to (4.24) we have that = 200=27, = 3=2 and k = 50=9.

PrfS sg Prfk + Y sg:

Then s = ! 50=9 where ! is the quantil 0.95 of a Gamma with

parameters 200/27 and 3/2. To calculate the quantil we can use Exel,
or the relation between the Gamma and the qui-square ( 2Y 2
2 ).
Using Exel ! = 18:5689, from where s = 24:124455. J
Chapter 5

Premium principles

In the insurance industry the production cycle is completely inverted. Insur-

ance is sold in the …rst place, under the payment of a premium, after which
claims may or not occur, which de…nes the cost. The insured is willing to
pay a price, usually greater than the mean expected cost of the claims that
he will produce, to prevent himself of big losses. The premium of a policy
(the price), excluding taxes, has a loading for expenses, and a loading for
deviations from the expected losses. In this course we do not deal with the
loading for expenses. In what follows we will call premium, the premium net
of expenses (including acquisition) and taxes.
In most situations the insurer does not have enough data from each policy
to be able of estimating accurately the distribution function of its losses.
Usually the insurer has data from a set of similar policies. Usually it is
supposed that the payments with respect to a given policy have a certain
distribution depending on one or more unknown parameters. This parameter
or set of parameters changes in the collective. The di¤erence between a given
risk and its collective is of fundamental importance in the theoretical
premium calculation principles.
In most applications it is assumed that there is an unknown parameter
associated to the number of claims. The number of claims, in a third party
motor insurance policy, each policy, in a collective of similar policies, origin-
ates claims according to a Poisson distribution with parameter . varies in
the collective , according to a given distribution U which can be estimated,
since U (x) represents the percentage of drivers with expected claim frequency
smaller or equal to x. Hence N is the number of claims of risk and N the
number of claims of a policy randomly chosen from the collective , i.e.
Z 1
PrfN = kg = PrfN = kgdU ( ):
0

42
CHAPTER 5. PREMIUM PRINCIPLES 43

If we considered that the severity depends also on a parameter , depending

on the risk, we would have = f ; g in which case is a. bidimensional ran-
dom variable. may be unidimensional, bidimensional, tridimensional,etc.
The aggregate claim amount, in a given period of time, of risk is de-
noted S and the aggregate claim amount of a risk randomly chosen from the
collective is denoted S. The distribution function of S is denoted FS and
the distribution of S by FS .
Loosely speaking a premium calculation principle, is a rule H that assigns
to each risk a real nonnegative number. Symbolically if the risk is denoted
Y , we obtain the premium H(Y ). In what follows we will assume that H(Y )
depends only on the cumulative distribution function of Y:
Hence the collective premium is

P = H(S)

and the risk premium

P = H(S ):
The collective pure premium of S is the expected value of S. In a similar
way the pure risk premium of S is the expected value of S . From the
actuarial point of view the collective premium should be greater than the
pure collective premium. The di¤erence between the premium and the pure
premium is called security loading.
Theoretically, the "correct" premium of a risk with parameter is the risk
premium P . In most cases FS is unknown or at least some of its parameters
are, the same happens with the risk premium. Hence the premium calculated
a priori is usually the collective premium, which can be modi…ed later on
according to the a posteriori information.

5.1 Some premium calculation principles

We will present some premium calculation principles, for the collective premium.
For the risk premium we just have to replace the distribution of S by the
distribution of S :

Expected value principle

P = (1 + )E(S); 0: (5.1)

Standard deviation principle

p
P = E(S) + Var[S]; 0: (5.2)
CHAPTER 5. PREMIUM PRINCIPLES 44

Variance principle

P = E(S) + Var[S]; 0: (5.3)

Zero utility principle

The collective premium P is the solution of

u(x) = E[u(x + P S)] (5.4)

where u(:) is a utility function such that u0 (x) > 0 and u00 (x) 0,
which is to say the utility of the initial wealth x must be equal to the
expected utility of the wealth resulting of insuring S. P is a function
of x, of the utility function u and of the distribution function of S.
The most important particular case is the exponential principle, which
is obtained for the exponential utility, i.e.
1 x
u(x) = 1 e :

In this case
1 S 1
P = ln E e = ln MS ( ): (5.5)
P is in this particular case independent of x. represents the risk
aversion coe¢ cient ( = u00 (x)=u0 (x)). P is increasing with .

Risk adjusted premium principles

Given any concave and increasing function g, such that g(0) = 0 and
g(1) = 1; we can de…ne a premium principle by
Z 1
P = g(1 FS (x))dx: (5.6)
0

Note that (5.6) is the expected value of a random variable with distri-
bution function 1 g(1 FS (x)):
Some particular cases are

– PH (Proportional Hazard) transform premium principle

g(x) = x1= ; 1; (g 0 (0) = 1);

– Dual power function premium principle

g(x) = 1 (1 x) ; 1; (g 0 (0) = );
CHAPTER 5. PREMIUM PRINCIPLES 45

– Denneberg’s absolute deviation premium principle

(1 + )x; 0 x < 0:5; 0 1
g(x) =
+ (1 )x; 0:5 x 1; (g 0 (0) = 1 + )

– Gini’s premium principle

g(x) = (1 + )x x2 ; 0 1; (g 0 (0) = 1 + );

– Exponential function principle

1 e x
; >0
g(x) = 1 e g 0 (0) = <1 ;
x = 0; 1 e
– Logarithmic function principle
(
log(1+ x)
; >0
g(x) = log(1+ ) g 0 (0) = <1 :
x = 0; log(1 + )

5.2 Properties
P1 - (Non-negative loading) H(S) E(S).
P2 - (No unjusti…ed loading) If PrfS = ag = 1 then
H(S) = a:

P3 - (No rip-o¤)
H(S) minfM jFS (M ) = 1g:
P4 - (Positive Homogeneity) 8b 2 <+
H(bS) = bH(S):

P5 - (Cash invariant) 8 c 2 <

H(S + c) = H(S) + c:

P6 - (Subaditivity) 8 S1 and S2 ,
H(S1 + S2 ) H(S1 ) + H(S2 ):

P7 - (Monotonicity) If PrfS1 S2 g = 1 ) H(S1 ) H(S2 ):

See (Goovaerts, Haezendonck & de Vylder 1983) for other premium prin-
ciples and their properties.
CHAPTER 5. PREMIUM PRINCIPLES 46

Table 5.1: Premium principles and properties

Property Principle
EV SD V Exponential Risk Ajust.
P1 Yes Yes Yes Yes Yes
P2 No Yes Yes Yes Yes
P3 No No No Yes Yes
P4 Yes Yes No No Yes
P5 No Yes Yes Yes Yes
P6 Yes Yes No No Yes
P7 Yes No No Yes Yes
Chapter 6

Risk measures

6.1 Coherent risk measures

In this course, as risks are modelled as non-negative r.v.’s, measuring risk is
equivalent to establishing a correspondence between the space of random
variables and the non-negative real numbers, representing the cash which has
to be added to X to make it acceptable. Premium calculation principles are
examples of possible risk measures.

De…nition 6.1 A risk measure that is cash invariant, positive homogeneous,

subaditive and monotone is called coherent.

See (Artzener, Delbaen, Eber & Heath 1999).

6.2 Value-at-Risk
Remember the de…nition of quantile:

De…nition 6.2 The pth quantile of a random variable X or of its cor-

responding distribution is denoted by p and de…ned as any value satisfying
FX ( p ) p FX ( p ):

De…nition 6.3 Given a risk X and a probability level p 2 (0; 1); the corres-
ponding VaR, denoted by VaRp (X) is p th quantile of X:

VaR is the percentile premium principle proposed by (Goovaerts et al.

1983).
Properties: VaR does not necessarily entail non-negative loading; does
not cause unjusti…ed loading; is no-ripo¤; it is cash invariant; positively
homogeneous; it is not subadditive; it is monotone.

47
CHAPTER 6. RISK MEASURES 48

Example 6.1 VaR is not subadditive:

Let Z be a continuous random variable such that FZ (1) = 0:91; FZ (90) =
0:95 and FZ (100) = 0:96 and consider p = 0:95: Then V aR0:95 (Z) = 90: Let
X; Y : Z = X + Y and
Z Z 100 0 Z 100
X= and Y =
0 Z > 100 Z Z > 100
FX (1) = 0:95; FX (90) = 0:99 and FX (100) = 1 and V aR0:95 (X) = 1:
FY (0) = 0:96 =) V aR0:95 (Y ) 0
=) V aR0:95 (X) + V aR0:95 (Y ) 1; but V aR0:95 (X + Y ) = 90:

De…nition 6.4 The Tail-Value-at-Risk of X at 100p% probability level is

de…ned as R1
p
V aRu (X)du
T V aRp (X) = (6.1)
1 p
and can be seen as the average of all VaR values above the security level p:

If the random variable X is continuous the TVaR is equal to

R1
xfX (x)dx
E[XjX > p ] = p = (6.2)
1 FX ( p )
R1
(x p )fX (x)dx
= p+ p =
1 FX ( p )
= V aRp (X) + eX ( p ):

E[XjX > p ] is, when X is a continuous random variable, also called

conditional tail expectation, CT Ep (X). When X is not continuous the con-
ditional tail expectation is de…ned in a di¤erent way, in such a way that it
is still equal to the T V aRp (X). Another name sometimes given to these
measure is Condicional-Value-at- Risk.
To show that in the continuous
R1 case (6.1) Ris qual to (6.2) make the change
1
of variable x = FX (u) in p V aRu (X)du = p FX 1 (u)du:
1

De…nition 6.5 The expected shortfall ESp (X) = E[(X V aRp (X))+ ]; which
is to say it is the stop loss premium with retention V aRp (X):

ESp (X)
T V aRp (X) = V aRp (X) +
1 p
It is possible to prove that T V aRp (X) is subaditive (or the conditional
tail expectation).
CHAPTER 6. RISK MEASURES 49

Example 6.2 Calculate the VaR and CTE for the N ( ; ):

Homework 6.1 Exercises 3.31; 3.32, 3.33; 3.35; 3.36.

Chapter 7

Reinsurance

7.1 Introduction
This section follows the articles in the Encyclopedia of Actuarial Science by
Gary Patrik, namely:
http://onlinelibrary.wiley.com/doi/10.1002/9780470012505.tar017/pdf
http://onlinelibrary.wiley.com/doi/10.1002/9780470012505.taf029/abstract
http://onlinelibrary.wiley.com/doi/10.1002/9780470012505.taw005/abstract

7.1.1 What is reinsurance?

Reinsurance is a form of insurance. A reinsurance contract is legally an
insurance contract; the reinsurer agrees to indemnify the cedant insurer for
a speci…ed share of speci…ed types of insurance claims paid by the cedant for
a single insurance policy or for a speci…ed set of policies. The terminology
used is that the reinsurer assumes the liability ceded on the subject policies.
The cession, or share of claims to be paid by the reinsurer, may be de…ned
on a proportional share basis (a speci…ed percentage of each claim) or on
nonproportional basis.
The nature and purpose of insurance is to reduce the …nancial cost to
individuals, corporations and other entities arising from the potential occur-
rence of speci…ed contingent events. An insurance company sells insurance
policies guaranting that the insurer will indemnify the policyholders for part
of the …nancial losses stemming from these contingent events. The pooling
of liabilities by the insurer makes the total losses more predictable than is
the case for each individual insured, thereby reducing the risk relative to the
whole. Insurance enables individuals, corporations and other entities to per-
form riskier operations. This increases innovation, competition and e¢ ciency
in a capitalistic marketplace.

50
CHAPTER 7. REINSURANCE 51

The nature and purpose of reinsurance is to reduce the …nancial cost

to insurance companies arising from the potential occurrence of speci…ed
insurance claims, thus further enhancing innovation, competition and e¢ -
ciency in the marketplace. The cession of shares of liability spreads risk
further throughout the insurance system. Just as an individual or company
purchases an insurance policy from an insurer, an insurance company may
purchase fairly comprehensive reinsurance from one or more reinsurers. And
a reinsurer may reduce its assumed reinsurance risk by purchasing reinsur-
ance coverage from other reinsurers, both domestic and international; such
a cession is called a retrocession.
Reinsurance companies are of two basic types: direct writers, which have
their own employed account executives who produce business, and broker
companies, which receive business through reinsurance intermediaries, or
brokers. Some direct writers do receive a part of their business through
brokers, and likewise, some broker reinsurers assume some business directly
from the ceding companies. It is estimated that more than half of U.S. rein-
surance is placed via intermediaries.
The form and wording of reinsurance contracts are not as closely regu-
lated as are insurance contracts, and there is no rate regulation of reinsurance
between private companies. A reinsurance contract is often a manuscript con-
tract setting forth the unique agreement between the two parties. Because
of the many special cases and exceptions, it is di¢ cult to make correct gen-
eralizations about reinsurance. Consequently, whenever you read anything
about reinsurance, you should often supply for yourself the phrases “It is
generally true that . . .”and “Usually . . .”whenever they are not explicitly
stated.
This heterogeneity of contract wordings also means that whenever you are
accumulating, analyzing and comparing various reinsurance data, you must
be careful that the reinsurance coverages producing the data are reasonably
similar.

7.1.2 Functions of Reinsurance

Reinsurance does not change the basic nature of an insurance coverage. On
a long-term basis, it cannot be expected to make bad business good. But it
does provide the following direct assistance to the cedant.

Capacity
Having reinsurance coverage, a cedant can write higher policy limits while
maintaining a manageable risk level. By ceding shares of all policies or just
CHAPTER 7. REINSURANCE 52

larger policies, the net retained loss exposure per individual policy or in
total can be kept in line with the cedant’s surplus. Thus smaller insurers can
compete with larger insurers, and policies beyond the capacity of any single
insurer can be written.
The word “capacity” is sometimes also used in relation to aggregate
volume of business. This aspect of capacity is best considered below in the
general category of …nancial results management.

Stabilization
Reinsurance can help stabilize the cedant’s underwriting and …nancial results
over time and help protect the cedant’s surplus against shocks from large, un-
predictable losses. Reinsurance is usually written so that the cedant retains
the smaller, predictable claims, but shares the larger, infrequent claims. It
can also be written to provide protection against a larger than predicted ac-
cumulation of claims, either from one catastrophic event or from many. Thus
the underwriting and …nancial e¤ects of large claims or large accumulations
of claims can be spread out over many years. This decreases the cedant’s
probability of …nancial ruin.

Financial results management

Reinsurance can alter the timing of income, enhance surplus and improve
various …nancial ratios by which insurers are judged. An insurance company
with a growing book of business whose growth is stressing their surplus can
cede part of their liability to a reinsurer to make use of the reinsurer’s surplus.
This is essentially a loan of surplus from the reinsurer to the cedant until the
cedant’s surplus is large enough to support the new business. We will see
other ways that reinsurance can be used to alter a cedant’s …nancial numbers.
As you might expect in a free market, this aspect of reinsurance has led to
some abuses in its use.

Management advice
Many professional reinsurers have the knowledge and ability to provide an
informal consulting service for their cedants. This service can include advice
and assistance on underwriting, marketing, pricing, loss prevention, claims
handling, reserving, actuarial, investment and personnel issues. Enlightened
self-interest induces the reinsurer to critically review the cedant’s operation,
and thus be in a position to o¤er advice. The reinsurer typically has more
experience in the pricing of high limits policies and in the handling of large
and rare claims. Also, through contact with many similar cedant companies,
CHAPTER 7. REINSURANCE 53

the reinsurer may be able to provide an overview of general issues and trends.
Reinsurance intermediaries may also provide some of these same services for
their clients.

7.2 Forms of reinsurance

Most reinsurance arrangements are placed on a facultative or a obligatory
way. Facultative reinsurance is made by what is called a facultative certi-
…cate. A facultative certi…cate reinsures just one primary policy. Its main
function is to provide additional capacity. It is used to cover part of speci…ed
large, especially hazardous or unusual exposures to limit their potential im-
pact upon the cedant’s net results or to protect the cedant’s ongoing ceded
treaty results in order to keep treaty costs down. The reinsurer is o¤ered,
underwrites and accepts each certi…cate individually; it is very similar to
primary insurance individual risk underwriting. Because facultative reinsur-
ance usually covers the more hazardous or unusual exposures, the reinsurer
must be aware of the potential for antiselection within and among classes
of insureds. Property certi…cate coverage is sometimes written on a propor-
tional basis; the reinsurer reimburses a …xed percentage of each claim on
the subject policy. Most casualty certi…cate coverage is written on an excess
basis; the reinsurer reimburses a share (up to some speci…ed dollar limit)
of the part of each claim on the subject policy which lies above some …xed
dollar attachment point (net retention).
Obligatory reinsurance is placed through a treaty. A treaty reinsures
a speci…ed part of the loss exposure for a set of insurance policies for a
speci…ed coverage period. For ongoing treaty coverage, the claims covered
may be either those occurring during the treaty term or those occurring on
policies written during the term. In the case of claims-made coverage, the
word “occurring” means those claims made to the ceding company during
the term. The premium subject to the treaty corresponds to the types of
claims covered: it is earned premium arising from policies of the speci…ed
type either in force or written during the term of the treaty. Because an
ongoing treaty relationship involves a close sharing of much of the insurance
exposure, it can create a close working partnership between the parties; the
expertise and services of the reinsurer or broker are available to the cedant.
This is especially true for treaties written by a direct writer or where there
is a strong reinsurer leading a brokered treaty.
CHAPTER 7. REINSURANCE 54

7.3 Quota-share treaty

A quota-share treaty reinsures a …xed percentage of each subject policy.
Its main function is …nancial results management, although it also provides
some capacity. The reinsurer usually receives the same share of premium
as claims, and pays the cedant a ceding commission commensurate with the
primary production and handling costs (underwriting, claims, etc.). Quota-
share treaties usually assume in-force exposure at inception. The cedant’s
…nancial results are managed because the ceding commission on the ceded
unearned premium reserve transfers statutory surplus from the reinsurer to
the cedant; we shall see this later. The cession of premium also reduces the
cedant’s net-premium-to-surplus ratio.
The ceding commission on quota-share treaties is often de…ned to vary
within some range inversely to the loss ratio. This allows the cedant to
retain better-than-expected pro…ts, but protects the reinsurer somewhat from
adverse claims experience.

Let a; 0 < a < 1; be the proportion retained of each claim. Than the
retained claim is aX; and its retained distribution FX (x=a):

7.4 Surplus treaty

Follows the article on Surplus in the Encyclopedia of Actuarial Science by
Ana Mata.
http://onlinelibrary.wiley.com/doi/10.1002/9780470012505.tas047/pdf
Surplus treaty is a type of proportional or pro-rata reinsurance treaty
where the ceding company determines the maximum loss that it can retain
for each risk in the portfolio. This amount is de…ned as “a line". Every risk
that provides coverage greater than the retained line is ceded to the surplus
treaty on a proportional basis where the proportion varies with the size of
the risk (hence the name “surplus"). For each risk ceded to the treaty the
losses and the premium are shared in equal proportions.
The main di¤erence between a surplus treaty and quota share rein-
surance (or standard proportional reinsurance) is that in a quota share the
insurer and the reinsurer share in a …xed proportion each and every risk of
the portfolio (losses and premiums), e.g. 80% of every risk may be ceded
to the reinsurer. In a surplus treaty the ceding company retains a …xed
maximum amount for each risk and this amount de…nes the retained pro-
portion depending on the total size of the underlying policy. For example, if
the retained line is $100,000 per risk, for a $500,000 policy limit the ceding
CHAPTER 7. REINSURANCE 55

company retains 20%, while for a $200,000 policy limit it retains 50%. See
Example 1 below.

Example 7.1 An insurance company has a portfolio of policies with limits

as shown in the following Figure. For a 50% quota share the ceding company’s
maximum liability per risk is $500,000 (from the policies with $1m in limit).
Under a surplus treaty with retained line of $100,000 the maximum liability
per risk is $100,000, ceding up to 9 lines to the reinsurer. Note that although
the maximum liability per risk for the ceding company is …xed under a surplus
treaty, the losses and premiums are shared between insurer and reinsurer from
…rst dollar, i.e. it is not an excess of loss treaty. For example, for a risk
with a limit of one million, 90% would be ceded even for a small loss of $100.

The proportion of each risk covered by the reinsurer under a surplus

treaty, (1 aR ), is calculated as follows:

(Policy limit Retained Line)

1 aR = (7.1)
Policy limit
In practice there are many variations on how the surplus treaty may act.
In the example above, there may be a …rst surplus treaty covering 4 lines
and a second surplus covering 5 lines.

Figure 7.1: Quota share vs. Surplus treaty

CHAPTER 7. REINSURANCE 56

7.4.1 Surplus treaty: insurer’s vs. reinsurer’s experi-

ence
Under a regular quota share agreement the ceding company and the reinsurer
would experience the same loss ratio (losses/premium), whereas under a
surplus treaty the reinsurer’s experience might be worse than the ceding
company’s. This is due to the fact that larger risks, for which the reinsurer
has a higher share, are often subject to a smaller pro…t margin. See Example
below.
In order to appropriately price a surplus treaty, the reinsurer requires
more detailed data that under a quota share treaty. The reinsurer needs to
estimate an average annual aggregate loss by policy limit in order to estimate
its own expected loss for the coverage provided. Therefore, the minimum data
required are:

1. Historical aggregate loss development by policy limit.

2. Historical premium by policy limit.

3. Historical changes in premium rates.

4. Historical changes in volume of business written by the ceding company.

In practice such detailed data are rarely available and hence additional
assumptions and adjustments to the available data need to be made in order
to estimate the reinsurer’s expected pro…tability.

Example 7.2 An insurance company has a portfolio of policies as detailed

in the following table. Assume that the ceding company only has capacity to
retain $100,000 per risk.

Limit Premium
100,000 6,500,000
300,000 2,500,000
500,000 3,250,000
1,000,000 750,000

A 9-line surplus treaty is agreed between the ceding company and the a
reinsurer. Hence, the reinsurer covers the percentages shown in Figure 7.1 for
each risk size, see formula (7.1). Assume that using historical development
losses by policy limit we estimate the expected aggregate losses by risk size
as shown in the following Table. These results do not include commissions
and expenses.
CHAPTER 7. REINSURANCE 57

Before Reinsurance (Gross) After Reinsurance (Net) Ceded to Reinsurance

Limit % Ceded Losses Premium LR Losses Premium LR Losses Premium LR
100,000 0% 3,250,000 6,500,000 50% 3,250,000 6,500,000 50% 0 0 N/A
300,000 67% 1,625,000 2,500,000 65% 541,667 833,333 65% 1,083,333 1,666,667 65%
500,000 80% 2,437,500 3,250,000 75% 487,500 650,000 75% 1,950,000 2,600,000 75%
1,000,000 90% 652,500 750,000 87% 65,250 75,000 87% 587,250 675,000 87%
Total 7,965,000 13,000,000 61.3% 4,344,417 8,058,333 53.9% 3,620,583 4,941,667 73.3%

From the results shown in this Table observe how the surplus treaty pro-
tects the insurer against higher risks and how the reinsurer’s pro…tability is
worse (20% worse in this case) than the net pro…tability of the ceding com-
pany. Under a quota share both the insurer and the reinsurer’s loss ratio
would have been 61.27% which is the loss ratio before reinsurance. Never-
theless, the gross, ceded and net loss ratios are the same by policy limit.
This is due to the fact that the loss ratio for policies with greater limits are
worse than for policies with lower limits and this creates an imbalance in the
reinsurer’s pro…tability (as it has larger share of the worse risks).

7.5 Excess of loss covers

An excess treaty reinsures, up to a limit, a share of the part of each claim that
is in excess of some speci…ed attachment point (cedant’s retention). Its main
functions are capacity and stabilization. An excess treaty typically covers
exposure earned during its term on either a losses-occurring or claims-made
basis, but run-o¤ exposure may be added in.
Per risk excess is a form of excess of loss reinsurance in which the retention
and limit of reinsurance apply "per risk" rather than per accident, per event
or in the aggregate. Per risk excess is generally employed by medium-to-
large companies, where surplus relief (recovery of expenses and costs from
reinsurance ceded) is not of great concern from a …nancial standpoint. Per
risk excess is also generally employed on short-tail business in property lines
rather than long-tail, in casualty lines.
For a per occurrence or per event excess treaty, the retention and limit
of reinsurance apply per accident or event. The subject loss is de…ned to be
the sum of all claims arising from one covered loss event or occurrence for
all subject policies. Per-occurrence excess is used for casualty exposures to
provide protection all the way up from working cover layers through clash
layers.
A working cover excess treaty reinsures an excess layer for which claims
activity is expected each year. The signi…cant expected claims frequency
creates some stability of the aggregate reinsured loss. So working covers
CHAPTER 7. REINSURANCE 58

are often retrospectively-rated, with the …nal reinsurance premium partially

determined by the treaty’s loss experience.
A higher exposed layer excess treaty attaches above the working cover(s),
but within policy limits. Thus there is direct single-policy exposure to the
treaty.
A clash treaty is a casualty treaty that attaches above all policy limits.
Thus it may be only exposed by:

1. extra-contractual-obligations (i.e., bad faith claims);

2. excess-of-policy-limit damages (an obligation on the part of the insurer

to cover losses above an insurance contract’s stated policy limit);

3. catastrophic workers’compensation accidents

4. “clash”of claims arising from one or more loss events involving multiple
coverages or policies.

Both higher exposed layers and clash are almost always priced on a …xed
cost basis, with no variable commission or additional premium provision.
A catastrophe cover is a per-occurrence treaty used for property exposure.
It is used to protect the net position of the cedant against the accumulation of
claims arising from one or more large events. It is usually stipulated that two
or more insureds must be involved before coverage attaches. The coverage is
typically of the form of a 90% or 95% share of one or more layers (separate
treaties) in excess of the maximum retention within which the cedant can
comfortably absorb a loss, or for which the cedant can a¤ord the reinsurance
prices.
When reinsuring the layer L xs M; the insurer is ceding to the reinsurer,
for each loss X; the amount
8
< 0 if X M
Z(M; L) = min(L; (X M )+ ) = X M if M < X M + L (7.2)
:
L if X > M + L;

with y+ = maxf0; yg:

The insurer may use a program with several XL treaties. They are made
in such a way that the ceiling M + L of a treaty is the retention limit of the
next treaty. When L = +1 we speack of unlimited cover. The reinsurer
may also limit its responsibility to a proportion of the excess X M:
The retained part for a L xs M layer is
CHAPTER 7. REINSURANCE 59

8
< X if X M
Y (M; L) = X Z(M; L) = M if M < X M + L (7.3)
:
X L if X > M + L:

Note that (7.2) may be written as (X M )+ (X (M + L))+ so

Y (M; L) = X (X M )+ + [X (M + L)]+ .
The k-th raw moment of the individual claims are

Z M Z 1
k k
k
E (Y (M; L)) = x dFX (x) + M (FX (M + L) FX (M )) + (x L)k dFX (x)
0 M +L

Z M Z 1
k 1
=k x (1 FX (x))dx + k (x L)k 1 (1 FX (x))dx;
0 M +L
(7.4)
The distribution function of the retained claims is
FX (x) if x < M
FY (M;L) (x) = (7.5)
FX (x + L) if x M:

The aggregate claim amount is

X
N
e
S(M; L) = Yi (M; L);
i=0

which is a compound distribution, with the same number of claim then before
reinsurance and with individual claim amount FY (M;L) (x) given by (7.5).
The aggregate ceded claims are

X
N X
N
b
S(M; L) = Zi (M; L) = min(L; (Xi M )+ );
i=0 i=0

have a compound distribution, with the same number of claims and individual
losses with distribution function
FX (x + M ) if 0 x<L
FZ(M;L) (x) =
1 if X L

b
Alternatively we could think of writing S(M; L) as a compound distri-
P
bution where the number of payments were N and individual payments
Y P:
CHAPTER 7. REINSURANCE 60

The k-th raw moment of the individual ceded losses are

Z M +L
E (Z(M; L)) =k
(x M )k dFX (x) + Lk (1 FX (M + L))
M
(7.6)
Z M +L
=k (x M )k 1 (1 FX (x))dx:
M

7.6 Aggregate excess, or stop loss covers

For an aggregate excess treaty, also sometimes called a stop loss cover, a loss
is the accumulation of all subject losses during a speci…ed time period, usually
one year. It usually covers all or part of the net retention of the cedant. It
protects net results, providing very strong stabilization. Claims arising from
natural catastrophes are often excluded, or there may be a per-occurrence
maximum limit.
Chapter 8

Ruin Theory

Let Ut , for t 0, be the surplus of a given portfolio at time t: Let St be the

aggregate losses occurred between 0 and t of the same portfolio. U (0)=u, is
the initial surplus, or initial reserve. Let Pt be the premium income received
between 0 and t; net of expenses and let us assume that the premium is
received at a constant rate c > 0, i.e. the premium received up to t is ct.
The the surplus at time t is
Ut = u + ct St ; t 0: (8.1)
We are not considering here income on reserves, dividends, and other
factors.
We say that ruin occurs if the surplus is negative in a given point or points
of time. The surplus level can be analysed in discrete or in continuous time.
The time horizon may be …nite or in…nite, which leads to four concepts of ruin
probabilities. Continuous time and in…nite time horizon, discrete time and
in…nite time horizon, continuous time and …nite time horizon and discrete
time and …nite time horizon.

8.1 Continuous time model

Let
T = infft : t 0 e Ut < 0g
be the time when ruin occurs, with the assumption that T = 1 if Ut 0 for
all t. Then the ruin probability in continuous time and in…nite horizon is
(u) = Pr(T < 1);
and the ruin probability before t in continuous time is
(u; t) = Pr(T < t):

61
CHAPTER 8. RUIN THEORY 62

We must have (u; t) (u); 8t.

We assume that fSt ; t 0g follows a compound Poisson model. Let
c > X (note that if c X then (u) = 1). De…ne the loading coe¢ cient
by the equation c = (1 + ) X . Let be the supremum of the values
for which the moment generating function of FX (x) exists in the interval
( 1; ). We assume that > 0 and that MX (r) goes to 1 when r !
(this may not happen when is …nite).

8.1.1 The adjustment coe¢ cient

The adjustment coe¢ cient is de…ned as the only positive root r = R of

rc = [MX (r) 1]; r< ; (8.2)

which is equivalent to

1 + (1 + ) Xr = MX (r); r< ; (8.3)

or to Z 1
(1 FX (x))
1+ = erx dx; r< ; (8.4)
0 X
when FX (0) = 0; which is equivalent to
Z 1
1+ = erx fe (x)dx
0

where fe (x) = (1 FX (x)) is the equilibrium probability density function of X.

X
The left hand side is a linear function with slope (1+ ) X passing through
the point (0,1); on the other hand limr!1 MX (r) = 1, MX (0) = 1, MX0 (r) =
E XerX , MX0 (0) = X , MX00 (r) = E X 2 erX , which implies that the right
hand side of (8.3) is an increasing, convex and that it also takes the value 1
at the origin. As c > X , which is to say > 0, then the slope of the left
hand side function exceeds the slope of the right hand side function at r = 0.
Then we can conclude that (8.3) has two solutions, one is the trivial solution
and the other is the adjustment coe¢ cient.

Example 8.1 Let the individual claims be exponential with mean X.

1
1 + (1 + ) Xr = (1 X r) ;

R= :
(1 + ) X
J
CHAPTER 8. RUIN THEORY 63

Figure 8.1: Adjustment coe…cient

Example 8.2 Calculate the adjustment coe¢ cient when X is a Gamma with
parameters (2; ) : Let = 2: Then X = 2 and MX (r) = (1 r) 2 : We
get
6 2 r2 11 r + 4 = 0
which is equivalent to
(2r 1)(3r 4) = 0
1
from where R = 2
:

Note that
2 X
R< :
E(X 2 )
This is a good guess for a starting point to calculate the adjustment coe¢ cient
when numerical techniques have to be used.

8.1.2 A Functional equation to the ultimate ruin prob-

ability
Theorem 8.1 (Lundberg inequality) For the classical risk process with initial
capital u (any real number),
Ru
(u) e ; (8.5)

where R is the adjustment coe¢ cient.

CHAPTER 8. RUIN THEORY 64

Proof. Let k (u) be the probability that ruin occurs at or before the k th
claim for any 1 < u < +1. If we prove that k (u) e Ru is true for all
k; than taking limits when k ! 1 we get (8:5). For k = 0; the inequality
holds because 0 (u) = 0; for any u 0 and 0 (u) = 1 for u 0: We are
going to split "ruin at or before the n th claim" as regards time and size of
the …rst claim. Assume that the …rst claim occurs between time t and t + dt:
This event has probability e t dt: Assume that its size is between x and
x + dx; which has probability dFX (x): Then the capital after time t equals
u + ct x: Integrating over x and t yields
Z 1Z 1
k (u) = k 1 (u + ct x)dFX (x) e t dt:
0 0

Now assume by induction that k 1 (u) e Ru for all u: Then

Z 1Z 1
k (u) e R(u+ct x) dFX (x) e t dt =
0
Z0 1 Z 1
Ru t(Rc+ )
= e e dt eRx dFX (x) =
0 0
Ru Ru
= e MX (R) = e :
+ Rc

Homework 8.1 11.3; 11.7

De…nition 8.1 Let G(u; y) = Prfruin occurs with initial reserve u and the
de…cit immediately after ruin occurs is at most y); u 0; y 0; i.e.

G(u; y) = Pr[U (T ) 2 ( y; 0) and T < +1 j U (0) = u]:

Hence
(u) = lim G(u; y):
y !+1

Theorem 8.2 The function G(u; y) satis…es the equation

Z u
@
G(u; y) = G(u; y) G(u x; y)dFX (x) [FX (u + y) FX (u)]
@u c c 0 c
(8.6)

Proof. Conditioning by the time and the size of the …rst claim
Z +1 Z u+ct
t
G(u; y) = G(u + ct x; y)dFX (x) + FX (u + ct + y) FX (u + ct) e dt:
0 0
CHAPTER 8. RUIN THEORY 65

Let make the change of variable t = z c u (z = u + ct), dt = dz=c:

Z +1 Z z
z u
G(u; y) = G(z x; y)dFX (x) + FX (z + y) FX (z) e c dz =

u 0 c
Z +1 Z z
u=c
= e G(z x; y)dFX (x) + FX (z + y) FX (z) e z=c dz
c u 0

Calculating the derivative with respect to u; we get

Z u
@ u=c u=c
G(u; y) = G(u; y)+ e e G(u x; y)dFX (x) + FX (u + y) FX (u)
@u c c 0

from where the result follows.

Theorem 8.3 The function G(0; y) is given by
Z y
G(0; y) = [1 FX (x)] dx; y 0
c 0
Proof. See book.
Theorem 8.4 The survival probability with no initial reserve is

(0) = 1 (0) = (8.7)

1+
R1
Proof. Recall that X = 0 [1 F (x)] dx and note that
Z 1
X 1
(0) = lim G(0; y) = [1 FX (x)] dx = = ;
y!1 c 0 c 1+
from where the result follows.
Theorem 8.5 The probability of ultimate survival (u) satis…es
Z u
0
(u) = (u) (u x)dFX (x); u 0: (8.8)
c c 0
Proof. Making y ! 1 in(8.6) we gert
Z u
0
(u) = (u) (u x)dFX (x) [1 FX (u)] (8.9)
c c 0 c
which is equivalent to
Z u
0
(u) = [1 (u)] [1 (u x)] dFX (x) [1 FX (u)] =
c c 0 c
Z u
(u) + (u x)dFX (x):
c c 0
The result follows.
CHAPTER 8. RUIN THEORY 66

Exercise 8.1 Show that

1 u
(u) = 1 exp ; u 0 (8.10)
1+ (1 + )
x=
satis…es (8.8) if FX (x) = 1 e ; x > 0 or that

1 u
(u) = exp ; u 0
1+ (1 + )
satis…es (8.9).

8.1.3 The maximum aggregate loss

We now derive a general solution to (8.8) subject to the boundary condition
(8.7) and (1) = 1.
The probability that the surplus will ever fall below the initial reserve u is
(0); because the surplus process has stationary and independent increments.

Theorem 8.6 The random variable Y representing the amount of a drop

below the initial reserve u; given that there is a drop, has probability density
function fe (y) = [1 FX (y)]= X :

Proof. Recall that G(u; y) = Prfruin occurs with initial reserve u and
the de…cit immediately after ruin occurs is at most y); u 0; y 0: As the
process has stationary and independent increments G(0; y) also respresents
the probability that the surplus ever drops below its original value and the
amount of this drop is at most y:

G(0; y)
Pr[Y y] = =
(0)
Z y
= [1 FX (x)] dx =
c (0) 0
Z y
1
= [1 FX (x)] dx:
X 0

Hence
1 FX (y)
fY (y) = = fe (y): (8.11)
X
CHAPTER 8. RUIN THEORY 67

Note that this density is independent of u: If there is a drop of y; the

surplus after the drop is u y; and because the surplus has stationary and
independent increments, ruin occurs thereafter with probability (u y)
provided u y 0; otherwise ruin would have already occurred. The prob-
ability of a second drop is again (0); and the amount of the second drop
has also density fe (y) and is independent of the …rst drop. The total number
of drops K is geometrical distributed, that is
k
k 1
PrfK = kg = [1 (0)] (0) = ; k = 0; 1; 2; :::
1+ 1+
(the geometric parameter is 1/ ):
The Maximum aggregate loss is
L = Y1 + Y2 + ::: + YK
with L = 0 if K = 0 and
(u) = PrfL ug; u 0:
Hence
X
1
1
k
(u) = Fe k (u); 0:
k=0
1+ 1+
CHAPTER 8. RUIN THEORY 68

This is known as the Beekman convolution formula.

Theorem 8.7 The moment generating function of L; is given by

Xr
ML (r) = : (8.12)
1 + (1 + ) Xr MX (r)

Proof. Considering (8.11) and using integration by parts we have that

R1
MY (r)= 1 0 ery [1 FX (y)]dy
X
(8.13)
= 1 r [MX (r) 1]:
X

The moment generating function of K is

MK (r) = ;
1+ er
so the moment generating function of L is

ML (r)=MK (ln MY (r))

= 1+ MY (r)

= 1+ 1=( X r)[MX (r) 1]

Xr
= 1+(1+ )
:
Xr MX (r)

An equivalent equation, which re‡ects the probability mass at zero is

1 [MX (r) 1]
ML (r) = + : (8.14)
1+ 1 + 1 + (1 + ) X r MX (r)

As (u) is the distribution function of L, we have that

R1
ML (r)=1 (0) + 0 eur [ 0 (u)]du
R1 0
= 1+ + 0
eur [ (u)]du

so (8.14) is equivalent to
Z 1
1 [MX (r) 1]
eur [ 0 (u)]du = : (8.15)
0 1 + 1 + (1 + ) X r MX (r)
CHAPTER 8. RUIN THEORY 69

This equation may be used in some cases to calculate (u). This is the case
when X is a mixture of exponentials. Let
X
n
1 1
x
fX (x) = Ai e i ; x>0 (8.16)
i=1 i

Pn
with i > 0, Ai > 0, 8i and i=1 Ai = 1. The moment generating function
of (8.16) is
X
n
1
MX (r) = Ai : (8.17)
i=1
1 r i
(8.17) exists only for r < minf 1 ; : : : ; n g. Extending its domain to all
r 6= 1= i , we can prove that the denominator of (8.15), which is the equation
that de…nes the adjustment coe¢ cient, has n roots called ri , i = 1; : : : ; n.
Replacing (8.17) in (8.15) and as both numerator and denominator are poly-
nomials in r, we may apply the method of partial fractions writing
Z 1 Xn
ur 0 Ci ri
e [ (u)]du = : (8.18)
0 i=1
ri r

But the only function satisfying (8.18) and (1) = 0 is

X
n
ri u
(u) = Ci e : (8.19)
i=1

Example 8.3 Given = 2=5 and

3 3x 7 7x
fX (x) = e + e ; x > 0;
2 2
determine (u). In this case

3=2 7=2
MX (r) = +
3 r 7 r
and
5
X = :
21
Then (8.15) becomes
Z 1
0 6 5 r
eur [ (u)]du = ; (8.20)
0 7 6 7r + r2
CHAPTER 8. RUIN THEORY 70

which is by applying the method of partial fractions

Z 1
6 4=5 1=5
eur [ 0 (u)]du = + ; (8.21)
0 7 1 r 6 r
i.e. r1 = 1, r2 = 6, C1 = 24=35 and C2 = 1=35. Hence
24 u 1 6u
(u) = e + e :
35 35
J

8.2 Discrete time model

let Un be the reserve at the end of period n, n = 0; 1; : : :, with
X
n
Un = u + (c Sk ); (8.22)
k=1

where u = U0 is the initial reserve, c is the premium received in each time

period end Sk are the aggregate claims in period k, with k = 1; 2; : : :, assumed
i.i.d. and identical distributed to a r.v. S. Assume that c > E[S] and that the
moment generating function of S exists and is such that limr ! MS (r) = 1.
Let
T = minfn : Un < 0g
and
(u) = Pr(T < 1):
~ is the only positive root of the equation
The adjustment coe¢ cient R,

E er(S c)
= 1; r< : (8.23)

Let us show that (8.23) has only a positive root. Let

h(r) = E er(S c)
:

Than
h0 (r) = E (S c)er(S c)

and
h00 (r) = E (S c)2 er(S c)
:
As h(0) = 1, h0 (0) = E[S c] is negative by assumption, h00 (r) > 0, 8r 0
and limr! h(r) = 1, h(0) = 1 and goes to 1 when r ! . The results
follows.
CHAPTER 8. RUIN THEORY 71

Figure 8.2: Adjustment coe¢ cient

~ when S is N( ; ).
Example 8.4 Determine R
r+ 12 2 r2
MS (r) = e :

Then
~ = 2(c
R
)
:
2

When c is calculated according to the exponential principle with para-

meter , we have that R ~= .
Note that if S is compound Poisson ( ; FX (:)), (8.23) is equivalent to
~ = R. By this motive we use the letter R, for both models.
(8.2), i.e. R
Bibliography

Adelson, R. M. (1966). Compound Poisson distributions, Operations Re-

search Quaterly 17: 73–75.

Artzener, P., Delbaen, F., Eber, J.-M. & Heath, D. (1999). Coherent meas-
ures of risk, Mathematical Finance 9(3): 203–228.

Beard, R. E., Pentikäinen, T. & Pesonen, E. (1984). Risk Theory, third edn,
Chapman and Hall, London.

Goovaerts, M. J., Haezendonck, J. & de Vylder, F. (1983). Insurance Premi-

ums, North-Holland, Amsterdam.

Klugman, S., Panjer, H. & Willmot, G. (2008). Loss Models- from Data to
Decisions, third edn, John Wiley & Sons, New York.

Mata, A. (2006). Surplus treaty, in J. Teugels & B. Sundt (eds), Encyclopedia

of Actuarial Science, Vol. 3, John Wiley & Sons, pp. 1650–1652.

Panjer, H. H. (1981). Recursive evaluation of a family of compound distri-

butions, ASTIN Bulletin 12: 22–26.

Patrik, G. (2006a). Reinsurance, in J. Teugels & B. Sundt (eds), Encyclopedia

of Actuarial Science, Vol. 3, John Wiley & Sons, pp. 1400–1403.

Patrik, G. (2006b). Reinsurance, functions and values, in J. Teugels &

B. Sundt (eds), Encyclopedia of Actuarial Science, Vol. 3, John Wiley
& Sons, pp. 1414–1416.

Patrik, G. (2006c). Working covers, in J. Teugels & B. Sundt (eds), Encyc-

lopedia of Actuarial Science, Vol. 3, John Wiley & Sons, pp. 1761–1763.

Sundt, B. & Jewell, W. (1981). Futher results on recursive evaluation of

compound distributions, ASTIN Bulletin 12: 27–39.

72