Risk Models 1 & 2

CS2-19: Risk models 1 Page 1
Risk models 1
Syllabus objectives
1.2 Compound distributions and their application in risk modelling
1.2.1 Construct models appropriate for short-term insurance contracts in terms of

the numbers of claims and the amounts of individual claims.
1.2.2 Describe the major simplifying assumptions underlying the models in 1.2.1.
1.2.3 Define a compound Poisson distribution and show that the sum of
independent random variables each having a compound Poisson
distribution also has a compound Poisson distribution.
1.2.4 Derive the mean, variance and coefficient of skewness for compound
binomial, compound Poisson and compound negative binomial random
variables.
The Actuarial Education Company © IFE: 2019 Examinations

Page 2 CS2-19: Risk models 1
0 Introduction
In the first section of this chapter, we describe the main features of general insurance policies.
There is no mathematics in this section, and you should be able to read it through fairly quickly in
order to obtain a good overview of the different types of product available.
In the remaining sections of the chapter we introduce the idea of a compound distribution. We will
define and use the compound Poisson, compound geometric, compound negative binomial and
compound binomial distributions.
We will also start to look at two models, the individual risk model and the collective risk model,
which are used to describe aggregate claims, ie the total claims that arise during a period from a
group of policies.
In the simplest case of a life assurance benefit (often referred to as ‘long-term business’), each
policy can result in at most one claim and claims will be for amounts specified in advance (ie the
sum assured). The benefit level may be the same for all policies or it may vary between policies.
In general insurance (often referred to as ‘short-term business’), policies can give rise to more than
one claim and the amounts will not usually be known in advance.
In this chapter we look at the theory of risk models. In the next chapter we explain how to adapt
these models when reinsurance is in place.
© IFE: 2019 Examinations The Actuarial Education Company

1 General features of a product
1.1 Insurable interest

Generally, for a risk to be insurable:
 the policyholder must have an interest in the risk being insured, to distinguish
between insurance and a wager
 a risk must be of a financial and reasonably quantifiable nature.
1.2 Insurable risk

Ideally risk events also need to meet the following criteria if they are to be insurable:
 Individual risk events should be independent of each other.

In practice we won’t often get strict independence but a low correlation is desirable.
 The probability of the event should be relatively small. In other words, an event that
is nearly certain to occur is not conducive to insurance.
For example, a house would not be insured if it stood on the edge of a crumbling cliff.
 Large numbers of potentially similar risks should be pooled in order to reduce the
variance and hence achieve more certainty.
The similar risks should still be independent.
 There should be an ultimate limit on the liability undertaken by the insurer.

This would help the risk event meet the above criteria that it must be of a reasonably
quantifiable nature.
 Moral hazards should be eliminated as far as possible because these are difficult to
quantify, result in selection against the insurer and lead to unfairness in treatment
between one policyholder and another.
Moral hazards occur when a person takes more risks because another party bears the cost
of those risks.
However, the desire for income means that an insurer or reinsurer will usually be found to
provide cover when these ideal criteria are not met.
Other characteristics that most general insurance products share are:
 Cover is normally for a fixed period, most commonly one year, after which it has to
be renegotiated. There is normally no obligation on the insurer or insured to
continue the arrangement thereafter although in most cases a need for continuing
cover may be assumed to exist.
 Claims are not of fixed amounts, and the amount of loss as well as the fact needs to
be proved before a claim can be settled.
 A claim occurring does not bring the policy to an end.

 Claims may occur at any time during the policy period. Although there is normally a
contractual obligation on the policyholder to report a claim to the insurer as quickly
as possible, notification may take some time if the loss is not evident immediately.
Settlement of the claim may take a long time if protracted legal proceedings are
needed or if it is not straightforward to determine the extent of the loss. However,
from the moment of the event giving rise to the claim the ultimate settlement amount
is a liability of the insurer. Estimating the amounts of money that need to be
reserved to settle these liabilities is one of the most important areas of actuarial
involvement in general insurance.
Classes of insurance in which claims tend to take a long time to settle are known as
long-tail. Those which tend to take a short time to settle are known as short-tail, although
the dividing line between the two categories is not always distinct.

2 Models for short-term insurance contracts
2.1 The basic model

Many forms of non-life insurance can be regarded as short-term contracts, for example
motor insurance. Some forms of life insurance also fall into this category, for example
group life and one-year term assurance policies.
A short-term insurance contract can be defined as having the following attributes:
 The policy lasts for a fixed, and relatively short, period of time, typically one year.
 The insurance company receives from the policyholder(s) a premium.
 In return, the insurer pays claims that arise during the term of the policy.
At the end of the policy’s term the policyholder may or may not renew the policy. If it is
renewed, the premium payable by the policyholder may or may not be the same as in the
previous period.
The insurer may choose to pass part of the premium to a reinsurer. In return, the reinsurer
will reimburse the insurer for part of the cost of the claims during the policy’s term
according to some agreed formula.
An important feature of a short-term insurance contract is that the premium is set at a level
to cover claims arising during the (short) term of the policy only. This contrasts with life
assurance policies where mortality rates increasing with age mean that the (level) annual
premium in the early years would be more than sufficient to cover the expected claims in
the early years. The excess amount would then be accumulated as a reserve to be used in
the later years when the premium on its own would be insufficient to meet the expected cost
of claims.
Now to be more specific, a short-term insurance contract covering a risk will be considered.
A risk includes either a single policy or a specified group of policies. For ease of
terminology the term of the contract is assumed to be one year, but it could equally well be
any other short period, for example six months. The random variable S denotes the
aggregate claims paid by the insurer in the year in respect of this risk. Models will be
constructed for this random variable S . In Section 3 collective risk models will be studied.
Later, in the next chapter, the idea of a collective risk model is extended to an individual risk
model.
We will see shortly what these terms mean.
A first step in the construction of a collective risk model is to write S in terms of the
number of claims arising in the year, denoted by the random variable N , and the amount of
each individual claim. Let the random variable X i denote the amount of the i th claim.
Then:
N
S  Xi (19.1)
i 1
where the summation is taken to be zero if N is zero.

This decomposition of S allows consideration of claim numbers and claim amounts

separately. A practical advantage of this is that the factors affecting claim numbers and
claim amounts may well be different. Take motor insurance as an example. A prolonged
spell of bad weather may have a significant effect on claim numbers but little or no effect on
the distribution of individual claim amounts. On the other hand, inflation may have a
significant effect on the cost of repairing cars, and hence on the distribution of individual
claim amounts, but little or no effect on claim numbers.
This approach is referred to as a collective risk model because it is considering the claims arising
from a group of policies taken as a whole, rather than by considering the claims arising from each
individual policy.
The random variable S is the sum of a random number of random quantities, and is said to have a
compound distribution.
Because compound distributions arise commonly in general insurance examples, the random
variable N is often referred to as the ‘number of claims’ and the distribution of the random
variables X1 , X2 ,  is referred to as the ‘individual claim size distribution’, even where the
compound distribution arises in another context.
To define a compound distribution, we need to know:

 the distribution of N (which is a discrete distribution) and
 the distribution of the X i ’s (which may be any distribution).
If the distribution of the X i ’s is continuous, then S will have a mixed distribution, ie partly discrete
and partly continuous. This is because of the possibility that N  0 .
The problems that will be studied are the derivation of the moments and distribution of S in
terms of the moments and distributions of N and the X i ’s. Both will be studied with and
without simple forms of reinsurance. The corresponding problems for the reinsurer will
also be studied, ie the derivation of the moments and distribution of the aggregate claims
paid in the year in respect of this risk by the reinsurer.
2.2 Discussion of the simplifications in the basic model

The model for short-term insurance described in the previous subsection contains a
number of simplifications as compared to a real insurance operation. The first of these is
that it is usually assumed that the moments, and sometimes the distributions, of N and the
X i ’s are known with certainty. In practice these would probably be estimated from some
relevant data.
For example, we might assume that claim amounts have a Gamma(500,4) distribution.
In practice it may not be possible to make such simple assumptions. For example:
 There may not be an appropriate theoretical distribution that models the distribution of
claim amounts actually paid sufficiently well.
 Even if the shape of the distribution is satisfactory, appropriate parameter values may
change over time, even in the short term.

 There may not be sufficient homogeneity in the portfolio. For example, different policies
may produce claim amounts that have different sizes. This leads to the idea of a mixture
distribution.
Another simplification is to assume, at least implicitly, that claims are settled more or less
as soon as the incident causing the claim occurs, so that, for example, the insurer’s profit is
known at the end of the year. In practice, there will be at least a short delay in the
settlement of claims and in some cases the delay can amount to many years. This will be
especially true when the extent of the loss is difficult to determine, for example if it is to be
decided in a court of law.
Delays will often lead to higher payments being made, owing to inflationary factors. (The relevant
inflation rate may well be very different from that normally used to measure inflation.)
The model does not in general include any mention of expenses. The premium is assumed
to pay the claims and include a loading for profit. In practice, the premium paid by the
policyholder(s) will also include a loading for expenses. It is possible to include expenses
in the model in a very simple way.
The simplest way to allow for expenses would be to use a claim size distribution that was
artificially inflated to allow for some sort of claim expense amount (eg adding an extra 20%),
although this might not give the right ‘shape’. Alternatively we might express the random
variable X as the sum of two other random variables, one to represent the actual claim amount
and the other to represent the corresponding claim expense.
An important element in models for long-term insurance is a rate of interest since, as

explained above, excess premium income would be invested to build up reserves. Interest
is a relatively less important, but still important, feature of short-term insurance. It is
possible to include interest in models for short-term insurance but it is more usual to ignore
it, at least in elementary models.
We will ignore interest in the models used in this chapter.
There are a number of additional elements included when setting the premium to be
charged to policyholders, including the policyholders’ previous claims record and these are
covered in Subject CP1 – Actuarial Practice. The allowance for policyholders’ claim
experience could be based on claim frequency or total claim amounts. This is beyond the
scope of this subject.
In fact, we have already looked at models that make allowance for policyholders’ claims
experience in Chapter 2.
2.3 Notation and assumptions

Throughout this chapter the following two important assumptions will be made:
 the random variables { X i }N

i 1 are independent and identically distributed
 the random variable N is independent of { X i }N

i 1 .

In words these assumptions mean that:

1. the number of claims is not affected by the amount of individual claims
2. the amount of a given individual claim is not affected by the amount of any other
individual claim
3. the distribution of the amounts of individual claims does not change over the (short)
term of the policy.
Point 1 follows from the second of the two assumptions above. Points 2 and 3 follow from the
first.
Throughout this chapter it will be assumed that all claims are for non-negative amounts, so
that P ( X i  x )  0 for x  0 . Many of the formulae in this chapter will be derived using the
moment generating functions (from now on abbreviated to MGFs) of S , N and X i . These
MGFs will be denoted MS (t ) , MN (t ) and M X (t ) , respectively, and will be assumed to exist
for some positive values of the dummy variable t . The existence of the MGF of a
non-negative random variable for positive values of t cannot generally be taken for
granted; for example the MGFs of the Pareto and of the lognormal distributions do not exist
for any positive value of t . However, all the formulae derived in this chapter with the help
of MGFs can be derived, although less easily, without assuming the MGFs exist for positive
values of t .
One method would be to use characteristic functions, E (eitX ) , which don’t have the same
convergence problems as MGFs. However the Core Reading does not cover these.
G( x ) and F ( x ) shall denote the distribution functions of S and X i , respectively, so that:
G ( x )  P (S  x ) and F ( x )  P ( X i  x )
For convenience it will often be assumed that the density of F ( x ) exists and it will be
denoted f ( x ) . In cases where this density does not exist, so that X i has a discrete or a
mixed continuous/discrete distribution, expressions such as:

0 x f ( x ) dx
should be interpreted appropriately. The meaning should always be clear from the context.
The k th moment, ( k  1,2,3... ) of X i about zero will be denoted mk .
Using this notation:
E ( X i )  m1 and var( X i )  m2  m12

3 The collective risk model
3.1 The collective risk model

Recall from Section 2.1 that S is represented as the sum of N random variables X i , where
X i denotes the amount of the i th claim. Thus:
S  X1  X 2    X N
and S  0 if N  0 .
S is said to have a compound distribution.
Note that it is the number of claims, N , from the risk as a collective (as opposed to
counting the number of claims from individual policies) that is being considered and this
gives the name ‘collective risk model’. Within this framework, expressions in general terms
for the distribution function, mean, variance and MGF of S can be developed.
3.2 Distribution functions and convolutions

An expression for G( x ) , the distribution function of S , can be derived by considering the
event {S  x } . Note that if this event occurs, then one, and only one, of the following
events must occur:
{S  x and N  0} (ie no claims)
or {S  x and N  1} (ie one claim of amount  x )
or {S  x and N  2} (ie two claims which total  x )
or {S  x and N  r } (ie r claims which total  x )
and so on. These events are mutually exclusive and exhaustive.
Thus:

{S  x }   {S  x and N  n }
n0
and hence:

P (S  x )   P (S  x and N  n)
n0

  P (N  n) P (S  x | N  n)
n0

Question
A group of policies can give rise to at most two claims in a year. The probability function for the
number of claims is as follows:
Number of claims, n 0 1 2
P(N  n) 0.6 0.3 0.1
Each claim is either for an amount of 1 or an amount of 2, with equal probability. Claim amounts
are independent of one another and are independent of the number of claims.
Determine the distribution function of the aggregate annual claim amount, S .
Solution
S can take the values 0, 1, 2, 3 or 4.
S will only equal 0 if N  0 and this has probability 0.6. So:
P(S  0)  P(S  0)  P(N  0)  0.6
S will equal 1 if there is one claim for amount 1. So:
P(S  1)  P(N  1, X  1)
 P(N  1)P( X  1) by independence

 0.3  0.5
 0.15
and:
P(S  1)  P(S  0)  P(S  1)  0.6  0.15  0.75
The other values of the CDF can be calculated in a similar way and are given below:
P(S  2)  P(S  1)  P(S  2)  0.75  0.3  0.5  0.1  0.52  0.925
P(S  3)  P(S  2)  P(S  3)  0.925  0.1  2  0.52  0.975

P(S  4)  P(S  3)  P(S  4)  1
The distribution of a sum of independent random variables can be found using convolutions.
If Z  X  Y , where X and Y are independent random variables with PDFs (or PFs) fX (x) and
fY (y), then fZ (z) , the PDF (or PF) of Z , is called the convolution of X and Y .
This is written mathematically as fZ  fX * fY .

A formula for a convolution can be found by summing over all possible values of x and y that give
a particular value z .
Finding a convolution
fZ (z)   fX (x) fY (z  x) for discrete random variables
x
fZ (z)   fX  x  fY (z  x) dx for continuous random variables
‘Sum or integrate over all values of x that could lead to a total of z .’
Similar formulae can be used to find the distribution function of a sum.
FZ (z)   fX (x)FY (z  x) or  FX (x) fY (z  x)

x x
FZ  z    fX (x)FY (z  x)dx or FZ  z    FX (x) fY (z  x)dx
Question
Suppose that N  Poisson( ) , M  Poisson( ) , and N and M are independent.
Use a convolution approach to derive the probability function of N  M .
Solution
Let V  N  M . Then, for v  0,1,2,... :
v
P(V  v)   P(N  n)P(M  v  n)
n0
v
 n e   v  n e  
  n! (v  n)!
n0
e (   ) v v!
   n  v n
v ! n  0 n!(v  n)!
e (   ) v  v  n v  n
    
v ! n0  n 
e (   )
 (   )v
v!
This is the probability function of the Poisson(   ) distribution. So N  M  Poisson(   ) .
This result can also be proved using MGFs.

We can now consider the distribution function for a compound distribution.
For convolutions of distribution functions, suppose that { X i }ni 1 are independent and
identically distributed (IID) random variables with common distribution function F ( x ) .
n
Then the distribution function of  Xi is denoted by F n * ( x ) , so that:
i 1
F n* (x )  P(X1  X 2    X n  x )
Now note that if N  n , then S is the sum of a fixed number n , of random variables,
{ X i }ni  1, and hence:
P (S  x | N  n )  F n * ( x )
where F n * ( x ) is the n-fold convolution of the distribution F ( x ) .
In other words, F 3* (x) would be the convolution F (x)* F (x)* F (x) etc.
(Note that F 1* ( x ) is just F ( x ) and, for convenience, F 0* ( x ) is defined to equal 1 for all
non-negative values of x . Otherwise F 0* ( x )  0 .) Thus:

G( x )  P (S  x )   P (N  n) F n * ( x ) (19.2)
n0
Formula (19.2) is a general expression for the distribution function of S . Neither the
distribution of N nor of X i has been specified.
Note that when X i is distributed on the positive integers it is easy to calculate P (S  x ) for
x  1,2,3,... since:
P (S  x )  G ( x )  G ( x  1)

  P (N  n) F n * ( x )  F n * ( x  1)
n 1

ie P (S  x )   P (N  n)fxn * (19.3)
n 1
n
where f n* n* n*
x  F ( x )  F ( x  1) is the probability function of  Xi .
i 1
n
This is just saying that fxn*  P( X i  x) .
i 1
As in the case when X i is a continuous random variable, P (S  0)  P (N  0) .

When the number of claims is large, and the claim amount distribution is not too skewed, we can
approximate the distribution of S by a normal distribution with mean E (S) and variance var(S) .
We explain how to calculate moments of S in the next section.
3.3 Moments of compound distributions

To calculate the moments of S , conditional expectation results are used, conditioning on
the number of claims, N . To find E [S ] , apply the identity:
E [S ]  E [E [S | N ]]
Here we are using the conditional expectation formula, which is given on page 16 of the Tables.
n
Now E [S | N  n ]   E[ X i ]  nm1 . Hence:
i 1
E [S | N ]  Nm1
and:
E [S ]  E [Nm1 ]  E [ N ]m1 (19.4)
This can also be written as follows:
E (S)  E (N)E ( X )
where E ( X )  E ( X i ) , i  1,2,..., N .
Formula (19.4) has a very natural interpretation. It says that the expected aggregate claim
amount is the product of the expected number of claims and the expected individual claim
amount.
This formula is also given on page 16 of the Tables.
Question
If X has a Pareto distribution with parameters   400 and   3 , and N has a Poisson(50)
distribution, calculate the expected value of S .
Solution
The expected value is:
400
E (S)  E (N)E ( X )  50   10,000
31

Variance
To find an expression for var[S ] , apply the identity:
var[S ]  E [var[S | N ]]  var[E [S | N ]]
Here we are using the conditional variance formula, which is given on page 16 of the Tables.
Since E (S |N)  Nm1 , we have:
var[ S ]  E[var[ S | N ]]  var[Nm1 ]
var[S | N ] can be found by using the fact that individual claim amounts are independent.
Now:
n  n
var[S | N  n ]  var  
 i 1
Xi   
 i 1
var[ X i ]  n(m2  m12 )
and so var[S | N ]  N (m2  m12 ) . Hence:
var[S ]  E [N (m2  m12 )]  var[Nm1 ]
ie: var[S ]  E [N ](m2  m12 )  var[N ]m12 (19.5)
Alternatively, writing this solely in terms of means and variances:
var(S)  E (N)var( X )  var(N)[E ( X )]2
where var( X )  var( X i ) , i  1,2,..., N .
This formula can also be found on page 16 of the Tables. We will use it to determine the
variances of the various compound distributions in the next few sections.
Unlike the expression for E [S ] , formula (19.5) does not have a natural interpretation. The
variance of S is expressed in terms of the mean and variance of both N and X i .
However, this formula shows that the variability of the overall aggregate claim distribution is a
function of both the variability in the number of claims and the variability in the claim amounts.
Moment generating function

The MGF of S is also found using conditional expectation. By definition,
MS (t )  E [exp(tS )] , so:
MS (t )  E E [exp(tS ) | N ] (19.6)
Again, we are conditioning on the number of claims, exactly as we did before.

Now E [exp(tS )| N  n ]  E [exp(tX 1  tX 2    tX n )] , and as { X i }ni  1 are independent random

variables:
n
E [exp(tX 1  tX 2    tX n )]   E [exp(tX i )]
i 1
Also, since { X i }ni  1 are identically distributed, they have common MGF, M X (t ) , so that:
n n
 E [exp(tX i )]   M X (t )  [ M X (t ) ]n
i 1 i 1
Hence:
E [exp(tS ) | N ]  [ M X (t )]N (19.7)
These conditional expectations are random variables because they are functions of N .
Hence, inserting (19.7) in (19.6):
MS (t )  E [ M X (t )N ]  E [exp(N log M X (t ))]  MN (log M X (t )) (19.8)
We can see this last step by observing that E[exp(N log M X (t ))] is of the same form as E (eN t ) but
with t replaced by log M X (t ) . So it is the MGF of N evaluated at log M X (t ) .
Again, this is given on page 16 of the Tables.
Thus, the MGF of S is expressed in terms of the MGFs of N and of X i . As with the
previous results, the distributions of neither N nor of X i have been specified.
A summary of the results for the mean, variance and MGF of S is given below.
Mean, variance and MGF of S

E (S)  E (N)E ( X )
var(S)  E (N)var( X )  var(N)[E ( X )]2
MS (t )  MN log M X (t ) 
There is one special case that is of some interest. This is when all claims are for the same
fixed amount.
Example
Consider a portfolio of one-year term assurances each with the same sum assured.
Suppose that the amount of a claim is B with probability one (assuming that a claim occurs at
all), ie P ( X i  B)  1 so that m1  B and m2  B2 .

B is a constant here, not a random variable. So the expected value of an individual claim amount
is B and its variance is 0.
Then S is distributed on 0, B,2B,... . In fact, S  BN so:
P (S  Bx )  P (N  x )
Formulae (19.4) and (19.5) give the mean and variance of S , but as S  BN it is easier to
note that E [S ]  E [N ]B and var[S ]  var[N ]B2 .
The next three sections consider compound distributions using various models for the
number of claims, N .
3.4 The compound Poisson distribution

First consider aggregate claims when N has a Poisson distribution with mean  denoted
N  Poi ( ) . S then has a compound Poisson distribution with parameter  , and F ( x ) is
the CDF of the individual claim amount random variable.
S is sometimes referred to as a compound Poisson random variable.
The results required for this distribution for N are:
E [N ]  var[N ]  
MN (t )  exp   (e t  1)
 
Note that these results are given in the Tables.
These results can be combined with those of Section 3.1 as follows.
From (19.4):
E (S)  E (N)E ( X )  E ( X )
ie:
E [S ]   m1 (19.9)
From (19.5):
var(S)  E (N)var( X )  var(N)[E ( X )]2   var( X )  [E ( X )]2  E ( X 2 )
ie:
var[S ]   m2 (19.10)
and from (19.8):
MS (t )  exp   (M X (t )  1) (19.11)

The results for the mean and variance have a very simple form. Note that the variance of S
is expressed in terms of the second moment of X i about zero (and not in terms of the
variance of X i ).
Note also that the formula for the skewness of S has a simple form when S is a compound
Poisson random variable:
skew [S ]   m3 (19.12)
ie:
skew(S)  E ( X 3 )
The easiest way to show that the third central moment of S is  m3 is to use the cumulant
generating function:
CS (t )  log MS (t )
To determine the skewness, we differentiate it three times with respect to t and set t  0 ,
ie:
d3
E [(S   m1)3 ]  log MS (t )
dt 3 t 0
In other words:
skew (S)  C S(0)
Recall also that:
E (S)  C S (0)
var(S)  C S (0)
Since MS (t )  exp   (M X (t )  1) , it follows that:
log MS (t )    M X (t )  1
So:
d3  d3 
3
log MS (t )    3 M X (t )  1   m3
dt  dt  t  0
ie E [(S   m1)3 ]   m3
This is because M X (0)  E ( X 3 ) for any random variable.

The coefficient of skewness of S is given by:
skew(S)
[var(S)]3/2
Hence the coefficient of skewness   m3 / ( m2 )3/2 .
This result shows that the distribution of S is positively skewed, since m3 is the third
moment about zero of X i and hence is greater than zero because X i is a non-negative
valued random variable. Note that the distribution of S is positively skewed even if the
distribution of X i is negatively skewed. The coefficient of skewness of S is
 m3 / ( m2 )3/2 , and hence goes to 0 as    . Thus for large values of  , the distribution
of S is almost symmetric.
Mean, variance and skewness of a compound Poisson random variable

If N  Poisson( ) , then S is a compound Poisson random variable and:
E (S)   E ( X )   m1
var(S)   E ( X 2 )   m2
skew(S)   E ( X 3 )   m3
These results are all given on page 16 of the Tables.
Sums of independent compound Poisson random variables

A very useful property of the compound Poisson distribution is that the sum of independent
compound Poisson random variables is itself a compound Poisson random variable. A
formal statement of this property is as follows.
Let S1, S2 ,..., Sn be independent random variables. Suppose that each Si has a compound
Poisson distribution with parameter i , and that the CDF of the individual claim amount
random variable for each Si is Fi ( x ) .
Define A  S1  S2    Sn . Then A has a compound Poisson distribution with parameter

 , and F ( x ) is the CDF of the individual claim amount random variable for A , where:
n n
1
  i and F ( x ) 

 i Fi ( x )
i 1 i 1
Recall that  is the capital form of the Greek letter  .
This is a very important result.

To prove the result, first note that F ( x ) is a weighted average of distribution functions and
that these weights are all positive and sum to one. This means that F ( x ) is a distribution
function and this distribution has MGF:

M(t )   etx f (x) dx
0
where f (x)  F (x) is the PDF of the individual claim amount random variable for A .
So:
 n
1
M (t )  0 exp(tx ) 
 i 1
i fi ( x ) dx
where fi ( x ) is the density of Fi ( x ) . Hence:
n n
1  1
M (t ) 

 i 0 exp{tx } fi ( x ) dx    i Mi (t ) (19.13)
i 1 i 1
where Mi (t ) is the MGF for the distribution with CDF Fi ( x ) .
Let M A (t ) denote the MGF of A . Then:
M A (t )  E [exp(tA)]  E [exp(tS1  tS2    tSn )]
By independence of {Si }ni 1 :
n
M A (t )   E (exp(tSi ))
i 1
As Si is a compound Poisson random variable, its MGF is of the form given by formula
(19.11), so:
E [exp(tSi )]  exp  i (Mi (t )  1)
Thus:
n  n 
M A (t )   exp{i (Mi (t )  1)}  exp   i (Mi (t )  1) 
i 1  i 1 
ie:
M A (t )  exp{  (M (t )  1)} (19.14)
where:
n n
1
  i and M (t )  
 i 1
i Mi (t )
i 1

By the one-to-one relationship between distributions and MGFs, formula (19.14) shows that
A has a compound Poisson distribution with Poisson parameter  . By (19.13), the
individual claim amount distribution has CDF F(x).
Question
The distributions of aggregate claims from two risks, denoted by S1 and S2 , are as follows:
 S1 has a compound Poisson distribution with parameter 100 and distribution function
F1 (x)  1  exp( x /  ) , x  0 .
 S2 has a compound Poisson distribution with parameter 200 and distribution function
F2 (x)  1  exp( x /  ), x  0 .
Assuming that S1 and S2 are independent, determine the distribution of S1  S2 .
Solution
Let S  S1  S2 . Then S has a compound Poisson distribution with parameters   300 and F (x) ,
where:
F (x)  13 F1 (x)  23 F2 (x)  1  13 exp( x /  )  23 exp( x /  )
We can use R to simulate values from a compound Poisson distribution.
The R code to simulate 10,000 values from a compound Poisson distribution with parameter
1,000 and a gamma claims distribution with   750 and   0.25 is:
set.seed(123)
n <- rpois(10000,1000)
s <- numeric(10000)
for(i in 1:10000)
{x <- rgamma(n[i],shape=750,rate=0.25)
s[i] <- sum(x)}
We can obtain a mean of 2,997,651, a standard deviation of 93,719.71 and a coefficient of

skewness of 0.02655921 as follows:
mean(s)
sd(s)
skewness<-sum((s-mean(s))^3)/length(s)
coeff.of.skew<-skewness/var(s)^(3/2)
We can estimate P(S>3,000,000) to be 0.4881 as follows:
length(s[s>3000000])/length(s)
Finally we could estimate the 90th percentile to be 3,115,719 as follows::
quantile(s,0.9)

We can plot a histogram of the compound distribution using the hist function and an
empirical density function using density in the plot function. We can then superimpose a
normal or other distribution to see if they provide a good approximation.
However, a better way to check the fit with a normal distribution is to use the qqnorm
function:
qqnorm(<simulated values>)
or the qqplot function to compare the sample data to simulated values from a fitted model
distribution:
qqplot(<simulated theoretical values>,

<simulated compound distribution values>)
Note we have used set.seed(123) so you can obtain the same values as this example.
3.5 The compound binomial distribution

Under certain circumstances, the binomial distribution is a natural choice for N . For
example, under a group life insurance policy covering n lives, the distribution of the
number of deaths in a year is binomial if it is assumed that each insured life is subject to the
same mortality rate, and that lives are independent with respect to mortality.
The notation N  Bin(n, p) is used to denote the binomial distribution for N . The key
results for this distribution are:
E [N ]  np
var[N ]  np(1  p )
MN (t )  ( pe t  1  p )n
Note that these results are given in the Tables.
However, the notation for the MGF is slightly different.
When N has a binomial distribution, S has a compound binomial distribution. One

important point about choosing the binomial distribution for N is that there is an upper
limit, n , to the number of claims.
Expressions for the mean, variance and MGF of S are now found in terms of n , p , m1 , m2
and M X (t ) when N  Bin(n, p ) .
There is no need to memorise the formulae in this section. However, it is important to be able to
derive them.
Formula (19.4) gives the mean:
E (S)  E (N)E ( X )
 E [S ]  npm1 (19.15)

Formula (19.5) gives the variance:
var(S)  E (N) var( X )  var(N)[E ( X )]2
 var[S ]  np(m2  m12 )  np(1  p) m12

(19.16)
 npm2  np2 m12
Lastly, formula (19.8) gives the MGF:
MS (t )  MN log M X (t ) 
 MS (t )  ( pM X (t )  1  p)n
We can also find expressions for the skewness and the coefficient of skewness.
The third central moment is found from the cumulant generating function:
C S (t )  ln MS (t )
In the next few steps, liberal use is made of the chain rule dx  du  du
dx  dy dy

and the product rule for
differentiation dx  dx dx 
d (uv)  du v  u dv . The third derivative of the cumulant generating function is:
d3 d3
3
log MS (t )  n log  pM X (t )  q  where q  1  p
dt dt 3
d2   d  
 2 
np  M X (t )  ( pM X (t )  q )1 
dt   dt  
d   d 2   d 
2 
1
  np  2 M X (t )  ( pM X (t )  q )  n  p M X (t )  ( pM X (t )  q )2 
dt   dt   dt  
 d3 
 np  3 M X (t )  ( pM X (t )  q )1
 dt 
 
 d2  d 
 3np2  2 M X (t )  ( pM X (t )  q )2  dt M X (t ) 
 dt   
 
3
 d 
 2n  p M X (t )  ( pM X (t )  q )3
 dt 
Setting t = 0 gives:
skew(S)  E [(S  npm1)3 ]  npm3  3np2 m2 m1  2np3 m13 (19.17)
The coefficient of skewness is then given by:
skew(S) npm3  3np2 m2 m1  2np3 m13


[var(S)]3/2 (npm2  np2 m12 )3/2

It can be deduced from formula (5.17) that it is possible for the compound binomial
distribution to be negatively skewed. The simplest illustration of this fact is when all claims
are of (a fixed) amount B . Then S  BN and:
E (S  E [S ])3   B3E (N  E [N ])3 

   
ie:
skew(S)  B3 skew(N)
So the coefficient of skewness of S is a multiple of that for N .
In fact:
skew(S) B3 skew(N)
coeff of skew(S)    coeff of skew(N)
[var(S)]3/2 [B2 var(N)]3/2
If p  0.5 , then the binomial distribution for N is negatively skewed.
So the coefficient of skewness of S will also be negative in this case.
Question
Determine an expression for the MGF of the aggregate claim amount random variable if the number
of claims has a Bin(100, 0.01) distribution and individual claim sizes have a Gamma(10, 0.2)
distribution.
Solution
Since N  Bin(100, 0.01) and X  Gamma(10, 0.2) , we have:
10
 t 
MN (t )  (0.99  0.01e )t 100
and MX (t )   1    (1  5t )10
 0.2 
So:
MS (t )  MN [log M X (t )]
100
 0.99  0.01elog M X (t ) 
 
  0.99  0.01M X (t )
100
100
 0.99  0.01(1  5t )10 
 

3.6 The compound negative binomial distribution

The final choice of distribution for N is the negative binomial distribution, which has
probability function:
 k  n  1 k n
P (N  n)    p q for n  0,1,2, 
 n 
This is the Type 2 negative binomial distribution. See page 9 of the Tables.
The Type 1 negative binomial distribution has probability function:
 n  1  k nk
P(N  n)   p q for n  k , k  1, k  2, ...
 k  1
It is not likely to be appropriate here, unless there is a specific reason why the number of claims
must be at least k .
The parameters of the distribution are k (  0 ) and p , where p  q  1 and 0  p  1 . This

distribution is denoted by NB(k , p) . When N  NB(k , p) :
kq
E [N ] 
p
kq
var[N ] 
p2
MN (t )  p k (1  qe t )k
The special case k  1 leads to the geometric distribution. Once again, note that these
results are given in the Tables.
The negative binomial distribution is an alternative to the Poisson distribution for N .
This is because the negative binomial distribution can take any non-negative integer value, unlike
the binomial distribution which has an upper limit.
One advantage that the negative binomial distribution has over the Poisson distribution is
that its variance exceeds its mean. These two quantities are equal for the Poisson
distribution. Thus, the negative binomial distribution may give a better fit to a data set
which has a sample variance in excess of the sample mean. This is often the case in
practice. In the next chapter a situation leading to the negative binomial distribution for N
is discussed. When N has a negative binomial distribution, S has a compound negative
binomial distribution.

Expressions for the mean, variance and MGF of S when N  NB(k , p) come immediately
from formulae (19.4), (19.5) and (19.8):
kq
E (S)  E (N)E ( X )  E [S ]  m
p 1
kq kq
var(S)  E (N)var( X )  var(N)[E ( X )]2  var[S ]  (m2  m12 )  2 m12
p p
Multiplying out the brackets and regrouping the terms, we see that:
kq kq kq kq kq
(m2  m12 )  2 m12  m2  m12  2 m12
p p p p p
kq kpq kq
 m2  2 m12  2 m12
p p p
kq kq  kpq 2
 m2  m1
p p2
kq kq(1  p) 2
 m2  m1
p p2
kq kq2
 m2  2 m12
p p
So:
kq kq 2
var[S ]  m2  2 m12
p p
and:
pk
MS (t )  MN log MX (t )   MS (t ) 
(1  qM X (t ))k
As before, the third central moment of S can be found from the cumulant generating
function of S , as follows:
d d
dt
log MS (t ) 
dt

k log p  k log 1  qM X (t ) 
kq d 
  dt M X (t ) 
1  qM X (t )  

Then:
2  d2 
d2 d  1 kq
log MS (t )  kq 2  M X (t )    2 M X (t ) 
dt 2
 dt  (1  qM X (t ))2 1  qM X (t )  dt 
 
and:
d3  d 
2
2 d 1
log MS (t )  3 kq  M X (t )   M X (t ) 
3
 dt 
  dt 2  2
dt  (1  qM X (t ))
3  d3 
2kq 3 d  kq
 M X (t )    3 M X (t ) 
3  dt 1  qM X (t )  dt 
(1  qM X (t ))    
Setting t  0 in the third derivative gives:
3kq 2 m1 m2 2kq 3 m13 kqm3

skew(S)  E [(S  E [S ])3 ]  2
 3
 (19.18)
p p p
The parameters k and p are positive, as are the moments of X . It therefore follows from
formula (19.18) that the compound negative binomial distribution is positively skewed. The
coefficient of skewness can be found from E ((S  E (S ))3 ) / (var(S ))3/2 .
Question
The distribution of the number of claims from a motor portfolio is negative binomial with
parameters k  4,000 and p  0.9 . The claim size distribution is Pareto with parameters   5
and   1,200 . Calculate the mean and standard deviation of the aggregate claim distribution.
Solution
The first two moments of the Pareto distribution are:
 1,200
m1  E ( X )    300
 1 4
 2 5  1,2002
m2  E ( X 2 )   m12   3002  240,000
(  1) (  2)
2
4 32
So, using the formulae for the mean and variance of a compound negative binomial distribution:
kq 4,000  0.1
E (S )   m1   300  133,333
p 0.9
kq kq2 2 4,000  0.1 4,000  0.12

var(S)  m2  m1   240,000   3002  111,111,111
p 2 0.9 2
p 0.9
So the standard deviation is 10,541.

4 Appendix
There is some repetition in the Core Reading in Section 3.3. To improve the flow of the chapter,
we have removed the repeated section from the main part of the text and placed it below.
Let S  X 1  X 2    X N (and S  0 if N  0 ) where the X i ’s are independent, identically

distributed (as a variable X ) and are also independent of N . S is said to have a
compound distribution.
Illustration: N is the number of claims which arise in a portfolio of business and X i is the
amount of the i th claim. S is the total claim amount.
The mean and variance of S are easily found:
E (S | N  n )  E ( X 1    X N | N  n)  E ( X 1    X n )  n E ( X )
Similarly:
var(S | N  n )  n var( X )
Therefore:
E (S )  E E (S | N )]  E NE ( X )  E (N )E ( X )
ie:
S  N  X
and:
var(S )  E  var(S | N )  var E (S | N )
 E N var( X )  var NE ( X )
 E (N ) var( X )  var(N ) E ( X )
2
ie:
 S2  N  X2   N2  X2
The MGF of S is given by:

MS (t )  E (e tS )  E E e tS | N 
 
and:
E (e tS | N  n)  E  exp  t ( X 1  X 2    X N )  | N  n 
 E  exp  t ( X 1  X 2    X n )  
 E exp(tX i )  M X (t )
n


Therefore:
MS (t )  E  M X (t )    E  exp  N log M X (t )    MN log M X (t ) 

N
 
Compound Poisson distribution

An important illustration is provided by the compound Poisson distribution, which is the
2
case in which N  Poisson( ) . In this case N   N .
Properties:
E (S )   E ( X )
var(S )   var( X )   E ( X )   E ( X 2 )
2
MN (t )  exp  (et  1)

 
so:
MS (t )  exp    M X (t )  1 
from which the mean and variance can be obtained and the results above verified.

Chapter 19 Summary
Insurable risks
For a risk to be insurable the policyholder should have an interest in the risk being insured to
distinguish between insurance and a wager, and it should be of a financial and reasonably
quantifiable nature. Ideally, risk events should:
 be independent
 have low probability of occurring
 be pooled with similar risks
 have an ultimate liability
 avoid moral hazards.
Characteristics of general insurance products

Most general insurance contracts share the following characteristics:
 Cover is normally for a fixed period, typically a year, after which it needs to be
renegotiated.
 There is usually no obligation to continue cover although in most cases a need for
continuing cover may be assumed to exist.
 Claims are not of fixed amounts.
 The existence of a claim and its amount have to be proved before a claim can be
settled.
 A claim occurring does not bring the policy to an end.
 Claims that take a long time to settle are known as long-tailed and those that take a
short time to settle are known as short-tailed.
Features of short-term insurance contracts

A short-term insurance contract can be defined as having the following attributes:
 The policy lasts for a fixed, and relatively short time period, typically one year.
 The insurance company receives a premium from the policyholder.
 In return the insurer pays claims that arise during the term of the policy.
 At the end of the policy term, the policyholder may or may not renew the policy. If it
is renewed, the premium may or may not be the same as in the previous period.
 The insurer may pass part of the premium to a reinsurer, who, in return, will
reimburse the insurer for part of the claims cost.

Collective risk model

Aggregate claim amounts may be modelled using a compound distribution. The aggregate
claim amount S is the sum of a random number of IID random variables:
S  X1  X 2    X N
where S is taken to be zero if N  0 . We assume that the random variable N is independent

of the random variables X i so that the distributions of the claim numbers and the individual
claim amounts can be analysed separately. The distribution of S is said to be a compound
distribution.
Other simplifying assumptions include:

 The moments (and sometimes the distributions) of N and X i are known.
 Claims are settled more or less as soon as the claims occur.

 Expenses and investment returns are ignored.
Specific types of compound distributions include the compound Poisson, compound binomial,
compound negative binomial, and compound geometric. Formulae for the MGF and the
moments of a compound random variable are given on page 16 of the Tables.
Convolutions
If Z  X  Y , and X and Y are independent, then:
P(Z  z)   P( X  x)P(Y  z  x) if X and Y are discrete

x

fZ (z)  fX  fY (z)   f (x) fY (z  x) dx if X and Y are continuous
 X
Sums of independent compound Poisson random variables

Let S1 , S2 ,..., Sn be a set of independent random variables where Si has a compound Poisson
distribution with parameter i and Fi ( x ) is the CDF of the individual claim amount random
variable for Si . Then A  S1    Sn is compound Poisson with parameter    i . The
CDF of the individual claim amount random variable for A is:
1
F (x) 

 i Fi (x)
The MGF of the individual claim amounts for A is:
1
M(t ) 

 i Mi (t)
where Mi (t ) is the MGF of the individual claim amounts for Si .

Chapter 19 Practice Questions

19.1 (i) State the two conditions that must hold for a risk to be insurable.
(ii) List five other risk criteria that would be considered desirable by a general insurer.
19.2 A group of policies can give rise to at most 2 claims. The probabilities of 0, 1 or 2 claims are ½, ¼
and ¼ respectively. Claim amounts are IID U(0,10) random variables. Let S denote the
aggregate claim amount random variable.
Sketch the frequency distribution of S .
19.3 The random variable S has a compound Poisson distribution with Poisson parameter 4. The
individual claim amounts are either 1, with probability 0.3, or 3, with probability 0.7.
Calculate the probability that S  4 .
19.4 A compound random variable S  X1  X2    X N has claim number distribution:
P(N  n)  9(n  1)4 n 2 , n  0,1,2,
The individual claim size random variable, X , is exponentially distributed with mean 2.
Calculate E (S) and var(S) .
19.5 Write down a formula for the MGF of a compound Poisson distribution with individual claim size
distribution Gamma( ,  ) and Poisson parameter  .
19.6 S1 and S2 are independent random variables each with a compound Poisson distribution. The
distribution of Si , i  1,2 , has Poisson parameter i and individual claim amount distribution
Fi (x) .
Which one of the following statements about the distribution of S1  S2 is correct?
A S1  S2 has a compound Poisson distribution with Poisson parameter 12 and individual
claim amount distribution F1 (x)  F2 (x ) .
B S1  S2 has a compound Poisson distribution with Poisson parameter (1  2 ) and

individual claim amount distribution  F1 (x)  F2 (x)  2 .
C S1  S2 has a compound Poisson distribution with Poisson parameter (1  2 ) and

individual claim amount distribution  1F1 (x)  2F2 (x)   1  2  .
D S1  S2 does not have a compound Poisson distribution.

19.7 Claims on a group of policies of a certain type arise as a Poisson process with parameter 1 .
Exam style Claims on a second, independent, group of policies arise as a Poisson process with parameter 2 .
The aggregate claim amounts on the respective groups are denoted S1 and S2 .
Using moment generating functions (or otherwise), show that S (the sum of S1 and S2 ) also has
a compound Poisson distribution and hence derive the Poisson parameter for S . [4]
19.8 The aggregate claim amount from a portfolio has a compound negative binomial distribution.
Exam style
(i) Show that if S  X1    X N , then:
MS (t )  MN log MX (t ) [3]
(ii) If N has Type 2 negative binomial distribution with k  2 and p  0.9 , and X has a
gamma distribution with   10 and   0.1 , determine an expression for MS (t ) . [2]
(iii) (a) Calculate the mean and variance of S .
(b) Using a suitable approximation, estimate the aggregate amount which will be
exceeded with probability 0.1%. [4]
(iv) The insurer in fact has 100 identical independent portfolios of this type. Let:
T  S1    S100
(a) Determine the moment generating function for T .
(b) Using a normal approximation, estimate the total aggregate claim amount from
the whole business which will be exceeded with probability 0.1%.
(c) Comment on your answers to parts (iii)(b) and (iv)(b). [4]

[Total 13]

Chapter 19 Solutions
19.1 (i) Criteria for an insurable risk
The two conditions are:

 The policyholder must have an interest in the risk being insured.
 The risk must be of a financial and reasonably quantifiable nature.
(ii) Other desirable features of a risk
Other desirable features are:

 Individual risks should be independent of one another.
 The probability that the insured event will occur should be small.
 Large numbers of similar risks should be pooled in order to reduce the variance and achieve
greater certainty.
 The insurer’s liability should be limited.
 Moral hazards should be eliminated as far as possible since these are difficult to quantify,
result in selection against the insurer and lead to unfairness in the treatment of some
policyholders.
19.2 If N  0 , ie there are no claims, then S  0 . So there is a point mass (or a ‘blob’ of probability) at
S 0.
If N  1 , ie there is exactly one claim, then S has a U(0, 10) distribution. This will happen with
probability ¼.
If N  2 , ie there are exactly 2 claims, then S  X1  X2 where X1 and X2 are independent random
variables with PDFs:
0.1 if 0  x  10
fX ( x )  fX ( x )  
1 2
0 otherwise
The PDF of S is:
 10
fS (s)   fX (s  x) fX (x) dx  0.1 fX (s  x) dx
 1 2 0 1
The integrand is 0.1 if 0  s  x  10 and 0 otherwise. So for 0  s  10 :
s
fS (s)  0.1 0.1 dx  0.01s
0
and for 10  s  20 :
10
fS (s)  0.1 0.1 dx  0.2  0.01s
s 10

So for N  2 , S has a symmetrical triangular shaped distribution on the interval (0, 20).
A graph of the distribution is shown below:
f( s)
1
2
1
20
1
40
s
10 20
This graph is the combination of a blob at zero, a uniform distribution on (0,10) and a triangular
distribution on (0,20).
19.3 We need to consider how we could get an aggregate claim amount of 4. This could happen in two
ways:
(a) 2 claims, one for 1 and one for 3
(b) 4 claims, all for an amount of 1.
The probability of this happening is:
P(S  4)  P(N  2)P( X1  1)P( X2  3)  P(N  2)P( X1  3)P( X2  1)
 P(N  4)P( X1  1)P( X2  1)P( X3  1)P( X4  1)
Since the X i ’s are identical, this simplifies to:
P(S  4)  2P(N  2)P( X  1)P( X  3)  P(N  4)[P( X  1)]4
e 4 42 e 4 4 4
 2  0.3  0.7   0.34
2! 4!
 0.06312
19.4 The probability function of N can be written as:
 n  1
P(N  n)  9(n  1)4 n 2   2
 (3 / 4) (1 / 4)
n
 n 
We can see from this formula that N has a Type 2 negative binomial distribution with parameters
k  2 and p  3 / 4 .

Hence:
kq 2  1 / 4
E (N)   2/3
p 3/4
kq 2  1 / 4
and: var(N)  2  8/9
p (3 / 4)2
The individual claim amounts have an exponential distribution with   ½ . So the mean and
variance of the individual claims are:
1 1
E(X )   2 and var( X )  2  4
 
Hence:
2 4
E (S)  E (N) E ( X )   2 
3 3
2 8 56
and: var(S)  E (N) var( X )  var(N)[E ( X )]2   4   22 
3 9 9
19.5 Since N  Poisson( ) :
MN (u)  exp[ (eu  1)]
So:
 
M S (t )  MN [log M X (t )]  exp   elog M X (t )  1   exp    M X (t )  1  
 
Since X  Gamma( ,  ) :

 t 
M X (t )   1  
  
So:
   
 t 
MS (t )  exp    1    1 
    
  
19.6 Option C is correct.
S1  S2 has a compound Poisson distribution with Poisson parameter (1  2 ) and individual
claim amount distribution  1F1 (x)  2F2 (x)   1  2  .

19.7 Let Ni denote the number of claims on policies of type i and let X i denote the claim amount
random variable for policies of type i , for i  1, 2 . Then:
   
MSi (t )  MNi ln M X i (t )  exp i M X i (t )  1 
By independence:
        
MS (t )  E etS  E e  1 2   E etS1 etS2  E etS1 E etS2  MS1 (t )MS2 (t )
t S S
[1]
Hence:
    
MS (t )  exp 1 M X1 (t )  1 exp 2 MX2 (t )  1 

 exp 1M X1 (t )  2MX2 (t )   1  2  
 exp   (MW (t )  1)  [1]
where:
  1  2 [1]
and:
1M X1 (t )  2M X2 (t )
MW (t )  [1]
1  2
Hence S is a compound Poisson random variable with Poisson parameter   1  2 .
19.8 (i) MGF of S
The MGF of S is:
MS (t )  E (etS )  E E (etS | N)

 
using the standard result for conditional means from page 16 of the Tables.
Looking at the inner expression, we have:
 
E (etS | N  n)  E et ( X1  Xn )  E (etX1 ) E (etXn ) [1]
Now each of these terms is just the MGF of the random variable X .

So:
E (etS |N  n)  M X (t )
n
and hence:
E (etS | N)  [M X (t )]N [1]
So:

N
  
MS (t )  E MX (t )   E eN log MX (t )  MN log MX (t ) [1]
(ii) Compound negative binomial distribution
We first need the individual MGFs. Using results from the Tables, we have:
10
 t 
MX (t )   1    (1  10t )10 [½]
 0.1 
2
 0.9 
and: MN (t )    [½]
 1  0.1et 
Combining these, using the result from part (i):
2
 0.9 
MS (t )  MN log MX (t )   10 
[1]
 1  0.1(1  10t ) 
(iii)(a) Mean and variance of S
We could differentiate this expression to find the mean and variance of S . However, it is much
easier to use the standard compound distribution formulae:
E (S)  E ( X )E (N)
var(S)  E ( X ) var(N)  var( X )E (N)

2
and:
Using the results from the Tables for the individual distributions:
 10
E(X )    100
 0.1
 10
var( X )  2   1,000 [1]
 0.12
kq 2  0.1
E (N)    0.22222
p 0.9
kq 2  0.1
var(N)  2   0.24691 [1]
p 0.92

Using the formulae above:
E (S)  E ( X )E (N)  22.222
var(S)  E ( X ) var(N)  var( X )E (N)  1002  0.24691  1,000  0.22222  2,691.358

2
and: [1]
(iii)(b) Aggregate amount
We now assume that S has an approximate normal distribution with this mean and variance. So,
standardising in the usual way, we have:
 k  E (S ) 
P  S  k   0.001  P  N(0,1)    0.001
 var(S) 
Using the percentage points of the standard normal distribution, we find that:
k  E (S)
 3.0902  k  22.222  3.0902 2,691.358  182.54 [1]
var(S)
(iv)(a) MGF of T
The MGF of T is:
MT (t )  E (etT )  E[et (S1    S100 ) ]  E (etS1 )  E (etS100 )  MS (t )

100
200
 0.9 
  [1]
 1  0.1(1  10t )10
 
The mean and variance of T are 100 times the corresponding results for S , ie:
E (T )  100E (S)  2,222
and: var(T )  100 var(S)  269,135.8 [1]
(iv)(b) Total aggregate amount
So the corresponding figure for the aggregate amount exceeded with probability 0.001 is:
2,222  3.0902 269,135.8  3,825.37 [1]
(iv)(c) Comment
This is substantially less than one hundred times the corresponding answer to part (iii)(b). The
Central Limit Theorem tells us that as the number of portfolios increases, bad experience in some
of the portfolios will be offset by better experience in others, leading to a situation where the
overall variation is relatively smaller. Pooling of similar risks reduces the overall variance. We can
see this happening here. [1]

Risk models 2
Syllabus objectives
1.2 Compound distributions and their application in risk modelling
1.2.1 Construct models appropriate for short-term insurance contracts in terms of

the numbers of claims and the amounts of individual claims.
1.2.2 Describe the major simplifying assumptions underlying the models in 1.2.1.
1.2.4 Derive the mean, variance and coefficient of skewness for compound
binomial, compound Poisson and compound negative binomial random
variables.
1.2.5 Repeat 1.2.4 for both the insurer and the reinsurer after the operation of
simple forms of proportional and excess of loss reinsurance.

0 Introduction
In this chapter we will look at some of the practical applications of risk models. We start by
looking at how the models can be adapted for situations involving reinsurance. A section on the
individual risk model is followed by some more complex examples of the use of risk models in
practice.

1 Aggregate claim distributions under proportional and individual

excess of loss reinsurance
In Chapter 19, we introduced the notation S to denote the aggregate claim amount random
variable, ie:
S  X1  X 2    X N
where N denotes the number of claims and X i denotes the amount of the i th claim.
Here we extend this concept to consider the situation when reinsurance is in force. We will use
the following notation:
 Yi is the amount paid by the insurer in respect of the i th claim
 Z i is the amount paid by the reinsurer in respect of the i th claim
 SI  Y1  Y2    YN is the aggregate claim amount paid by the insurer
 SR  Z1  Z2    ZN is the aggregate claim amount paid by the reinsurer.
The formulae that we derived for the mean, variance and MGF of S can be adapted to cover the
reinsurance situation by replacing X by Y or Z , as appropriate. For example:
E (SI )  E (N)E (Y )
var(SI )  E (N)var(Y )  var(N)[E (Y )]2
MSI (t )  MN (log MY (t ))
1.1 Proportional reinsurance

The distribution of the number of claims involving the reinsurer is the same as the
distribution of the number of claims involving the insurer, as each pays a defined
proportion of every claim.
For a retention level  ( 0    1 ), the i th individual claim amount for the insurer is  X i
and for the reinsurer is (1   ) X i .
In other words:
Yi   Xi
Zi  (1   ) Xi

So:
SI  Y1  Y2    YN
  X1   X2     XN
  ( X1  X2    XN )
S
and:
SR  Z1  Z2    ZN
 (1   ) X1  (1   ) X2    (1   ) XN
 (1   )( X1  X2    XN )
 (1   )S
ie the aggregate claims amounts are  S and (1   )S respectively.
Question
(i) Show that under a proportional reinsurance arrangement where the direct writer retains a
proportion  , the MGF of the net individual claim amount Y paid by the direct insurer is
M X ( t ) .
(ii) Hence give a formula for MSI (t ) when the number of claims follows a Poisson distribution
with mean 25 and individual claim amounts are exponentially distributed with mean 1,000.
Solution
(i) MGF of Y
Under this arrangement, Y   X . So:
MY (t )  E (etY )  E (et X )  E[e(t ) X ]  M X ( t )
(ii) MGF of insurer’s aggregate claim amount
Since N  Poisson(25) :
t
MN (t )  e25(e 1)
and the MGF of SI is:
MSI (t )  MN (log MY (t ))  e25(exp(log MY (t ))1)  e25(MY (t )1)

1 ):
In addition, since X is exponentially distributed with mean 1,000 (ie with parameter 1,000
M X (t )  (1  1,000t )1
and hence:
MY (t )  M X ( t )  (1  1,000 t )1
So the MGF of SI is:
25 (11,000 t )1 1

MSI (t )  e  
1.2 Individual excess of loss reinsurance

The amount that an insurer pays on the i th claim under individual excess of loss
reinsurance with retention level M is Yi  min{ X i , M } .
Equivalently:
 X if X  M
Y 
M if X  M
The amount that the reinsurer pays is Z i  max{0, X i  M } .
We can also write this as:
0 if X  M
Z 
 X  M if X  M
As previously stated, the insurer’s aggregate claims net of reinsurance can be represented
as:
SI  Y1  Y2    YN
and the reinsurer’s aggregate claims as:
SR  Z1  Z2    ZN (20.1)
If, for example, N ~ Poi ( ) , SI has a compound Poisson distribution with Poisson
parameter  , and the i th individual claim amount is Yi . Similarly, SR has a compound
Poisson distribution with Poisson parameter  , and the i th individual claim amount is Zi .

Hence, for this compound Poisson distribution with reinsurance, we have:
E (SI )   E (Y )
var(SI )   E (Y 2 )
skew(SI )   E (Y 3 )
Similar formulae can be obtained for SR by replacing Y with Z .
Note, however that if F (M )  0 , as will usually be the case, then Zi may take the value 0.
Here, F (x) denotes the distribution function of the original claim amount random variable, X .
So:
F (M)  P( X  M)
If this is greater than 0, then there is a non-zero probability that the reinsurer will not be involved
in a claim.
In other words, 0 is counted as a possible claim amount for the reinsurer. From a practical
point of view, this definition of SR is rather artificial. The insurer will know the observed
value of N , but the reinsurer will probably know only the number of claims above the
retention level M , since the insurer may notify the reinsurer only of claims above the
retention level.
Example
The annual aggregate claim amount from a risk has a compound Poisson distribution with
Poisson parameter 10. Individual claim amounts are uniformly distributed on (0,2 000) . The
insurer of this risk has effected excess of loss reinsurance with retention level 1,600.
Calculate the mean, variance and coefficient of skewness of both the insurer’s and
reinsurer’s aggregate claims under this reinsurance arrangement.
Solution
Let SI and SR be as above. To find E [Si ] calculate E [Yi ] . Now:
M
E (Yi )   x f ( x ) dx  M P ( X i  M )
0
where f ( x )  0.0005 is the U(0,2000) density function and M = 1,600.
This gives:
M
 0.0005 x 2 
E [Yi ]     0.2M  960
 2  0

So:
E [SI ]  10E [Yi ]  9,600
To find var[SI ] , we must calculate E [Yi2 ] :
M
E [ Yi2 ]  x
2
f ( x ) dx  M 2 P ( X i  M )
0
M
 0.0005 x 3  2
   0.2 M
 3  0
 1,194,666.7
So:
var[SI ]  10E [Yi2 ]  11,946,667
To find the coefficient of skewness of the insurer’s claims, we must calculate E [Yi3 ] :
M
E [ Yi3 ]  x
3
f ( x ) dx  M 3 P ( X i  M )
0
M
 0.0005 x 4  3
   0.2 M
 4  0
 1,638, 400,000
So:
E  SI  E (SI )    10 E [ Yi3 ]  16,384,000,000

3
 
and the coefficient of skewness of SI is:
16,384,000,000
 0.397
11,946,6673/2
To find E [SR ] , note that the expected annual aggregate claim amount from the risk is
E [S ]   E [ X ]  10  1,000  10,000 . Then:
E [SR ]  10,000  E [SI ]  400

To find var[SR ] , calculate E [ Z i2 ] from:
2,000
E [ Z i2 ]   ( x  M )2 f ( x ) dx
M
2,000  M
  0.0005 y 2 dy where y  x  M
0
2,000  M
 0.0005 y 3 
 
 3  0
 10,666.7
So:
var[SR ]  10E [ Z i2 ]  106,667
To find the coefficient of skewness of the reinsurer’s claims, we need to calculate E [ Z i3 ] :
2,000
E [ Z i3 ]   ( x  M )3 f ( x ) dx
M
2,000  M
  0.0005 y 3 dy where y  x  M
0
 3,200,000
So:
E  SR  E (SR )    10 E [ Zi3 ]  32,000,000

3
 
and the coefficient of skewness of SR is:
32,000,000
 0.92
106,6673/2
Question
Calculate the variance of S , the aggregate claim amount before reinsurance for the example
above and explain why:
var(SI )  var(SR )  var(S)

Solution
We have var(S)  10E ( X 2 ) , where:
2000
x2 4,000,000
E(X 2 )   2000
dx 
3
0
So var(S)  13,333,333 .
This is not equal to 11,946,667  106,667 because SI and SR are not independent.
To simulate the collective risk model with individual reinsurance we can combine the R
code from Chapters 15, 18 and 19.
For example, to simulate 10,000 values for a reinsurer where claims have a compound
Poisson distribution with parameter 1,000 and a gamma claims distribution with   750
and   0.25 under individual excess of loss with retention 2,500 we would use:
set.seed(123)
M <-2500
n <- rpois(10000,1000)
sR <- numeric(10000)
for(i in 1:10000)
{x <- rgamma(n[i],shape=750,rate=0.25)
z <- pmax(0,x-M)
sR[i] <- sum(z)}
We can now find moments, the coefficient of skewness, probabilities and quantiles as
before.
Earlier we mentioned that using SR  Z1    ZN is a bit artificial. We now look at an alternative

way of modelling the reinsurer’s compound claim amount distribution.
The reinsurer’s aggregate claims can also be represented by:
SR  W1  W2    WNR (20.2)
where the random variable NR denotes the actual number of (non-zero) payments made by
the reinsurer.
Here:
Wi  X i  M | X i  M
For example, suppose that the risk above gave rise to the following eight claim amounts in a
particular year:
403 1,490 1,948 443 1,866 1,704 1,221 823
The retention limit is 1,600.

Then in formula (20.1) the observed value of N is 8, and the third, fifth and sixth claims
require payments from the reinsurer of 348, 266 and 104 respectively. The reinsurer makes
a ‘payment’ of 0 on the other five claims.
In formula (20.2), the observed value of NR is 3 and the observed values of W1 , W2 and
W3 are 348, 266 and 104 respectively. Note that the observed value of SR is the same
(ie 718) under each definition.
Wi has density function:
f (x  M)
g (w )  , w 0
1  F (M )
We saw this result in Section 1.2 of Chapter 18. It can also be written as:
f (w  M)
fW (w)  X
1  FX (M)
To specify the distribution for SR as given in formula (20.2) the distribution of NR is

needed.
In some contexts it may be obvious what this distribution is, but here is a general method for
establishing the distribution.
This is found as follows. Define:
NR  I1  I2    IN
where N denotes the number of claims from the risk (as usual). I j is an indicator random
variable which takes the value 1 if the reinsurer makes a (non-zero) payment on the j th
claim, and takes the value 0 otherwise. Thus NR gives the number of payments made by
the reinsurer.
From its definition, we see that NR is a compound random variable. However, instead of being
the sum of individual claims amounts, NR is a sum of indicator random variables.
Since I j takes the value 1 only if X j  M :
P (I j  1)  P ( X j  M )   , say
and:
P (I j  0)  1  
In other words, I j has a Binomial(1,  ) distribution.
Further, I j has MGF:
MI (t )   e t  1  
The formula for the MGF of a binomial distribution is given on page 6 of the Tables.

By formula (19.8) in Chapter 19 (the formula for the MGF of a compound random variable), NR
has MGF:
MNR(t) = MN (log MI (t))
Question
If N has a Poisson( ) distribution and P( X  M)  ½ , show that NR has a Poisson(½ )

distribution.
Solution
Here   ½ , so:
MI (t )  ½et  ½
and:
  
MNR (t )  MN [log MI (t )]  exp [MI (t )  1]  exp  (½et  ½)  exp ½ (et  1)
This is the MGF of the Poisson(½ ) distribution. By the uniqueness property of MGFs, it follows
that NR  Poisson(½) .
We now continue the above Core Reading example where the annual aggregate claim amount
from a risk has a compound Poisson distribution with Poisson parameter 10, individual claim
amounts are uniformly distributed on (0,2000) , and the insurer of this risk has effected excess of
loss reinsurance with retention level 1,600.
Example
Continuing the above example and using formula (20.2) as the model for SR , it can be seen
that SR has a compound Poisson distribution with Poisson parameter 0.2  10  2 .
Individual claims, Wi , have density function:
f (w  M ) 0.0005
g (w )    0.0025 , for 0  w  400
1  F (M ) 0.2
ie Wi is uniformly distributed on (0,400).
Using the formulae for the moments of a continuous uniform distribution from page 13 of the
Tables, we have:
E [ Wi ]  200 , E [ Wi2 ]  53,333.33 and E [ Wi3 ]  16,000,000
giving the same results as before.

If we multiply these figures by 2 (the Poisson parameter of SR ), we get E (SR )  400 ,

var(SR )  106,667 and skew(SR )  32,000,000 , which agree with the answers obtained
previously.
Thus, there are two ways to specify and evaluate the distribution of SR .
1.3 Aggregate excess of loss reinsurance

Under an aggregate excess of loss arrangement with retention limit M , the insurer pays all the
claims if the total claim amount is less than or equal to the retention limit. The maximum
payment made by the insurer is M . So the insurer’s aggregate claim payment is:
 S if S  M
SI  
M if S  M
The reinsurer’s aggregate claim payment is:
0 if S  M
SR  
 S  M if S  M
Calculations involving aggregate excess of loss reinsurance are done using a first principles
approach.
Question
The annual number of claims from a small group of policies has a Poisson distribution with a mean
of 2. Individual claim amounts have the following distribution:
Amount 200 400

Probability 0.7 0.3
Individual claim amounts are independent of each other and are also independent of the number
of claims. The insurer has purchased aggregate excess of loss reinsurance with a retention limit
of 600.
Calculate the probability that the reinsurer is involved in paying the claims that arise in the next
policy year.
Solution
The reinsurer will be involved if the total claim amount is more than 600. Since the total claim
amount must be a multiple of 200, the probability is:
P(S  600)  1  P(S  0)  P(S  200)  P(S  400)  P(S  600)

The total claim amount will be 0 only if there are no claims. So:
e 2 20
P(S  0)  P(N  0)   e 2
0!
Using the assumption that individual claim amounts are independent of the number of claims, we
have:
e 2 21
P(S  200)  P(N  1, X1  200)  P(N  1)P( X1  200)   0.7  1.4 e 2
1!
P(S  400)  P(N  2, X1  200, X2  200)  P(N  1, X1  400)
e 2 22 e 2 21
  0.72   0.3
2! 1!
 1.58 e 2
and:
P(S  600)  P(N  3, X1  200, X2  200, X3  200)
P(N  2, X1  200, X2  400)  P(N  2, X1  400, X2  200)
e 2 23 e 2 22
  0.73  2   0.7  0.3
3! 2!
 1.29733 e 2
So:
P(S  600)  1  e 2 (1  1.4  1.58  1.29733)  0.28579
We can also simulate the collective risk model with aggregate reinsurance. For example to
simulate 10,000 values for a reinsurer where claims have a compound Poisson distribution
with parameter 1,000 and a gamma claims distribution with   750 and   0.25 under
aggregate excess of loss with retention 3,000,000 we would take our S from Section 3.4 of
Chapter 19 and then use:
sR <- pmax(0,s-3000000)

2 The individual risk model

Under this model a portfolio consisting of a fixed number of risks is considered. It will be
assumed that:
 these risks are independent
 claim amounts from these risks are not (necessarily) identically distributed random
variables
 the number of risks does not change over the period of insurance cover.
As before, aggregate claims from this portfolio are denoted by S . So:
S  Y1  Y2    Yn
where Y j denotes the claim amount under the j th risk and n denotes the number of risks.
It is possible that some risks will not give rise to claims. Thus, some of the observed values
of {Y j }nj 1 may be 0.
In fact in most forms of insurance most policies would not give rise to any claims during a given
year.
This approach is referred to as an individual risk model because it is considering the claims arising
from each individual policy.
For each risk, the following assumptions are made:
 the number of claims from the j th risk, N j , is either 0 or 1 (20.3)
 the probability of a claim from the j th risk is q j . (20.4)
If a claim occurs under the j th risk, the claim amount is denoted by the random variable
X j . Let F j ( x ) ,  j and  2j denote the distribution function, mean and variance of X j

respectively.
Assumption (20.3) is very restrictive. It means that a maximum of one claim from each risk
is allowed for in the model. This includes risks such as one-year term assurance (since a
policyholder can only die once), but excludes many types of general insurance policy. For
example, there is no restriction on the number of claims that could be made in a policy year
under household contents insurance.
There are three important differences between this model and the collective risk model:
(1) The number of risks in the portfolio has been specified. In the collective risk model,
there was no need to specify this number, nor to assume that it remained fixed over
the period of insurance cover (not even when it was assumed that N ~ Bin(n, q ) ).
On the other hand we could argue that there was an implicit assumption of a constant number of
risks in the very fact that we were using a binomial distribution to model the number of claims.
(2) The number of claims from each individual risk has been restricted. There was no
such restriction in the collective risk model.

(3) It is assumed that individual risks are independent. In the collective risk model it
was individual claim amounts that were independent.
The contrast here is between the occurrence of claims and the size of claims.
Assumptions (20.3) and (20.4) say that N j ~ Bin(1, q j ) . Thus, the distribution of Y j is
compound binomial, with individual claim amount random variable X j . From formulae
(19.15) and (19.16) in Chapter 19 (for the mean and variance of a compound random variable) it
follows immediately that:
E [Y j ]  q j  j (20.5)
var[Y j ]  q j  2j  q j (1  q j ) 2j (20.6)
S is the sum of n independent compound binomial random variables. The distribution of

S can be stated only when the compound binomial variables are identically distributed, as
well as independent. It is possible, but complicated, to compute the distribution function of
S under certain conditions.
However, it is easy to find the mean and variance of S :
 n  n n
E[S ]  E 
 j 1
Yj  
 j 1

E [Y j ]  
qjj
  j  1
(20.7)
The assumption that individual risks are independent is needed to write:
 n  n n
var [ S ]  var  
 j 1
Yj  
 j 1

var[ Y j ]  (q j  2j  q j (1  q j ) 2j ) (20.8)
  j 1
In the special case when {Y j }nj 1 is a sequence of identically distributed, as well as

2
independent, random variables, then for each policy the values of q j ,  j and  j are
identical, say q ,  and  2 . Since F j ( x ) is independent of j , we can refer to it simply as

F ( x ) . Hence, S is compound binomial, with binomial parameters n and q , and individual
claims have distribution function F ( x ) . In this special case, it reduces to the collective risk
model, and it can be seen from (20.7) and (20.8) that:
E [ S ]  nq 
var [ S ]  n q  2  n q (1  q )  2
which correspond to (19.15) and (19.16) respectively in Chapter 19.
Question
The probability of a claim arising on any given policy in a portfolio of 1,000 one-year term
assurance policies is 0.004. Claim amounts have a Gamma(5, 0.002) distribution. Calculate the
mean and variance of the aggregate claim amount.

Solution
We have:
5 5
j   2,500 and:  2j   1,250,000
0.002 0.0022
So the mean and variance of the aggregate claim amount are:
E (S)  nq   1,000  0.004  2,500  10,000
and:
var(S)  nq  2  nq (1  q) 2
 1,000  0.004  1,250,000  0.004  0.996  2,5002 

 
 (£5,468)2
We can use R to simulate the total claim amount payable under the individual risk model.
Suppose we have n life policies, with the probabilities of death for each policy contained in
the vector q and simulated claim amounts for each policy contained in the vector claim.
Then:
S <- q*claim
We can now find moments, the coefficient of skewness, probabilities and quantiles as
before, and also apply reinsurance if appropriate.

3 Parameter variability / uncertainty

This section forms part of the Core Reading, but does not address any specific syllabus objectives.
However, the material here provides useful practice in applying the models we have studied.
3.1 Introduction
So far risk models have been studied assuming that the parameters, that is the moments
and in some cases even the distributions, of the number of claims and of the amount of
individual claims are known with certainty. In general, these parameters would not be
known but would have to be estimated from appropriate sets of data. In this section it will
be seen how the models introduced earlier can be extended to allow for parameter
uncertainty / variability. This will be done by looking at a series of examples. Most, but not
all, of these examples will consider uncertainty in the claim number distribution since this,
rather than the individual claim amount distribution, has received more attention in the
actuarial literature. All the examples will be based on claim numbers having a Poisson
distribution.
3.2 Variability in a heterogeneous portfolio

Consider a portfolio consisting of n independent policies. The aggregate claims from the
i th policy are denoted by the random variable Si , where Si has a compound Poisson
distribution with parameters i , and the CDF of the individual claim amounts distribution is
F ( x ) . Notice that, for simplicity, the CDF of the distribution of individual claim amounts,
F ( x ) , is assumed to be identical for all the policies.
In this example the CDF of individual claim amounts, ie F ( x ) , is assumed to be known but
the values of the Poisson parameters, ie the i s, are not known. In this subsection the i s
are assumed to be (sample values of) independent random variables with the same (known)
distribution. In other words {i }ni 1 is treated as a set of independent and identically
distributed random variables with a known distribution. This means that if a policy is
chosen at random from the portfolio it is assumed that the Poisson parameter for the policy
is not known but that probability statements can be made about it. For example, ‘there is a
50% chance that its Poisson parameter lies between 3 and 5’. It is important to understand
that the Poisson parameter for a policy chosen from the portfolio is a fixed number; the
problem is that this number is not known.
Example
Suppose that the Poisson parameters of policies in a portfolio are not known but are equally
likely to be 0.1 or 0.3.
(i) Find the mean and variance (in terms of m1 and m2 ) of the aggregate claims from a
policy chosen at random from the portfolio.
(ii) Find the mean and variance (in terms of m1 , m2 and n ) of the aggregate claims
from the whole portfolio.

It may be helpful to think of this as a model of part of a motor insurance portfolio. The
policies in the whole portfolio have been subdivided according to their values for rating
factors such as ‘age of driver’, ‘type of car’ and even ‘past claims experience’. The policies
in the part of the portfolio being considered have identical values for these rating factors.
However, there are some factors, such as ‘driving ability’, that cannot easily be measured
and so they cannot be taken explicitly into account. It is supposed that some of the
policyholders in this part of the portfolio are ‘good’ drivers and the remainder are ‘bad’
drivers. The individual claim amount distribution is the same for all drivers but ‘good’
drivers make fewer claims (0.1 pa on average) than ‘bad’ drivers (0.3 pa on average). It is
assumed that it is known, possibly from national data, that a policyholder in this part of the
portfolio is equally likely to be a ‘good’ driver or a ‘bad’ driver but that it cannot be known
whether a particular policyholder is a ‘good’ driver or a ‘bad’ driver.
Solution
Let i , i  1, 2, ..., n be the Poisson parameter of the i th policy in the portfolio. {i }
i 1 is
regarded as a set of independent and identically distributed random variables, each with the
following distribution:
P (i  0.1)  0.5
P (i  0.3)  0.5
From this:
E [ i ]  0.2
var [ i ]  0.01
(i) The moments of Si can be calculated by conditioning on the value of i . Since

Si | i has a straightforward compound Poisson distribution, formulae (19.8) and
(19.9) in Chapter 19 can be used to write:
E [ Si ]  E [ E [ Si | i ] ]  E [ i m1 ]  0.2 m1
var [ Si ]  E [ var [ Si | i ] ]  var [ E [Si | i ] ]
 E [ i m2 ]  var[ i m1 ]
 0.2 m2  0.01m12
(ii) The random variables {Si }ni 1 are independent and identically distributed, each with
the distribution of Si given in part (i). Hence, the result in (i) above can be used to
write:
n 
E Si   n E [ Si ]  0.2 n m1
 i 1 
n 
var   Si   n var [ Si ]  0.2 n m2  0.01n m12
 i 1 

Example
Suppose the Poisson parameters for individual policies are drawn from a gamma
distribution with parameters  and  . Find the distribution of the number of claims from a
policy chosen at random from the portfolio.
Solution
Let Ni denote the number of claims from the i th policy in the portfolio and let i be its
Poisson parameter. Then Ni has a Poisson distribution with parameter i but the problem
is that (by assumption) the value of i is not known. What is known is the distribution from
which i has been chosen.
The problem can be summarised as follows:
Given that:
Ni | i ~ Poisson(i ) and i ~ Gamma( ,  )
find the marginal distribution of Ni .
The marginal distribution of Ni is its unconditional distribution. In this example, Ni can only take
whole number values, so it is a discrete random variable. To determine its marginal distribution,
we need to derive a formula for the unconditional probability P(Ni  x ) .
This problem can be solved by removing the conditioning in the usual way.
Recall that if X and Y are discrete random variables, then the unconditional probability P( X  x)
is given by:
P( X  x)   P( X  x ,Y  y)   P( X  x |Y  y )P(Y  y)
y y
In this example X is being replaced by Ni and Y is being replaced by i . Since i is a

continuous random variable, we turn the sum into an integral and formula becomes:
P(Ni  x)   P(Ni  x |i ) f (i ) di

i
For x  0,1, 2, ... :
  x    1
P (Ni  x )  0 exp{  }
x ! ( )
 exp {   } d 
 
  exp {  (  1) }  x  1 d 
( ) x ! 0

Evaluate the integral by comparing the integrand with a gamma density.
We can make the integrand look like the PDF of the Gamma( x   ,   1) distribution by inserting
(  1)x 
a factor of inside the integral. We need to compensate for doing this by inserting a
(x   )
(x   )
factor of outside the integral. This gives:
(  1)x 
  (x   )  (  1)x  x  1  ( 1)

P(Ni  x) 
( )x ! (  1)x  0 (x   )  e d
  (x   ) 

( )x ! (  1)x  0 f ( ) d where   Gamma(x   ,   1)
The integral in the line above is 1 (as we are integrating a PDF over all possible values of the
random variable).
So:
 ( x   )
P (Ni  x ) 
( ) x ! (  1) x 
which shows that the marginal distribution of Ni is negative binomial with parameters 

and .
 1
3.3 Variability in a homogeneous portfolio

Now a different example is considered. Suppose, as before, there is a portfolio of n
policies. The aggregate claims from a single policy have a compound Poisson distribution
with parameters  , and the CDF of the individual claim amounts random variable is F ( x ) .
The Poisson parameters are the same for all policies in the portfolio. If the value of  were
known, the aggregate claims from different policies would be independent of each other. It
is assumed that the value of  is not known, possibly because it changes from year to year,
but that there is some indication of the probability that  will be in any given range of
values. As in the previous example, it is assumed for simplicity that there is no uncertainty
about the moments or distribution of the individual claim amounts, ie about F ( x ) . The
uncertainty about the value of  can be modelled by regarding  as a random variable
(with a known distribution).
Example
Suppose that the Poisson parameter,  , will be equal to 0.1 or to 0.3 with equal probability.
(i) Calculate the mean and variance (in terms of m1 and m2 ) of the aggregate claims
from a policy chosen at random from the portfolio.
(ii) Calculate the mean and variance (in terms of m1 , m2 and n ) of the aggregate
claims from the whole portfolio.

Solution
Using the same notation as before let Si denote the aggregate claims from the i th policy in
the portfolio. The situation can be summarised as follows:
The random variables {Si   }ni 1 are independent and identically distributed, each with a
compound Poisson distribution with parameters  and F ( x ) . The random variable  has
the following distribution:
P (  0.1)  0.5
P (  0.3)  0.5
(i) Conditioning on the value of  :
E [ Si ]  E [ E ( Si |  ) ]  E [  m1 ]  0.2 m1
var [ Si ]  E [ var( Si |  ) ]  var [ E (Si |  ) ]  E [  m2 ]  var [  m1 ]
 0.2 m2  0.01m12
n 
(ii) E  Si   n E [S1 ]  0.2 n m1
 i 1 
(since {Si }ni 1 are identically distributed)
n    n    n 
var    
Si   E  var  Si      var E 
   i 1

Si    

 i 1    i 1    
 E [ n  m2 ]  var [ n  m1 ]
 0.2 n m2  0.01n2 m12
Note that S1 | ,  , Sn | are independent but S1 ,  , Sn are not unconditionally independent

n n
(since they all depend on the value of  ), so var( Si )   var(Si ) .
i 1 i 1
It is useful to compare the answers to the above example with those to the first example in
the Section 3.2. The values of the mean are in all cases the same, as are the variances when
a single policy is considered (part (i)). The difference occurs when variances for more than
one policy are considered (part (ii)), in which case the second example gives the greater
variance. It is important to understand the differences (and the similarities) between the two
examples. A practical situation where the second example could be appropriate would be a
portfolio of policies insuring buildings in a certain area. The number of claims could
depend on, among other factors, the weather during the year; an unusually high number of
storms resulting in a high expected number of claims (ie a high value of  ) and vice versa
for all the policies together.

3.4 Variability in claim numbers and claim amounts and parameter uncertainty
This section contains two more examples. The first is a rather complicated example
involving uncertainty over claim amounts as well as claim numbers.
Example
An insurance company models windstorm claims under household insurance policies using
the following assumptions.
The number of storms arising each year, K , is assumed to have a Poisson distribution with
parameter  .
The number of claims arising from the i th storm, Ni , i  1, 2, ..., K , is assumed to have a
Poisson distribution with parameter  i .
The parameters  i , i  1, 2, ..., K , are assumed to be independent and identically

distributed random variables, with E ( i )  n and var( i )  s12 .
The amount of the j th claim arising from the i th storm, X ij , j  1, 2, ..., Ni , has a lognormal
distribution with parameters i and  2 , where  2 is assumed to be known. The mean

claim amounts,  i  exp( i   2 / 2) are assumed to be independent and identically
distributed random variables with mean p and variance s22 .
It is also assumed that  i and  i are independent.
(i) Show that E [ X ij ]  p and var [ X ij ]  exp { 2 } ( p2  s22 )  p2 .
(ii) Let Si denote aggregate claims outgo from the i-th storm, so that Si | { i ,  i } is a
compound Poisson random variable. Show that:
E [ Si ]  np
and:
var [ Si ]  ( p2  s22 ) (n2  s12  n exp{ 2 })  n2 p2
(iii) Find expressions for the mean and variance of the annual aggregate claims outgo
from all storms.
Solution
(i) E [ X ij ]  E [ E ( X ij |  i ) ]  E [  i ]  p
var [ X ij ]  E [ var ( X ij |  i ) ]  var [ E ( X ij |  i ) ]
 E [  2i (exp{ 2 }  1) ]  var ( i )
 ( p2  s22 ) (exp{ 2 }  1)  s22
 ( p2  s22 ) exp{ 2 }  p2

(ii) E [ Si ]  E [ E ( Si |  i ,  i ) ]  E [  i  i ]  np since  i and  i are independent.
Now, since Si | { i ,  i } has a compound Poisson distribution:
var [ Si | i ,  i ]  i E [ X ij2 |  i ]  i ( 2i exp{ 2 })
and so:
E [ var (Si |  i ,  i ) ]  n ( p2  s22 ) exp{ 2 }
Also:
var [ E (Si | i ,  i ) ]  var [ i  i ]  E [ 2i  2i ]  n2 p2
 (n2  s12 )( p2  s22 )  n2 p2
Putting these last two results together:
var [ Si ]  (n2  s12 )( p2  s22 )  n2 p2  n( p2  s22 ) exp{ 2 }
(iii) Let R be a random variable denoting the annual aggregate claims outgo from all
storms. Then R can be written:
K
R  Si
i 1
where K has a Poisson distribution and the random variables Si are IID
(independent and identically distributed).
Hence, R has a compound Poisson distribution and so:
E [ R ]   E ( Si )   n p
var [ R ]   E ( Si2 )   ( var[ Si ]  E [ Si ]2 )
  ( p2  s22 )(n2  s12  n exp{ 2 })
Example
Each year an insurance company issues a number of household contents insurance
policies, for each of which the annual premium is £80. The aggregate annual claims from a
single policy have a compound Poisson distribution; the Poisson parameter is 0.4 and
individual claim amounts have a gamma distribution with parameters  and  . The
expense involved in settling a claim is a random variable uniformly distributed between £50
and £ b (>£50). The amount of the expense is independent of the amount of the associated
claim. The random variable S represents the total aggregate claims and expenses in one
year from this portfolio. It may be assumed that S has approximately a normal distribution.

(i) Suppose that:
  1 ;   0.01 ; b  100
Show that the company must sell at least 884 policies in a year to be at least 99%
sure that the premium income will exceed the claims and expenses outgo.
(ii) Now suppose that the values of  ,  and b are not known with certainty but could
be anywhere in the following ranges:
0.95    1.05 ; 0.009    0.011 ; 90  b  110
By considering what, for the insurance company, would be the worst possible
combination of values for  ,  and b , calculate the number of policies the
company must sell to be at least 99% sure that the premium income will exceed the
claims and expenses outgo.
Solution
Let X i be the amount of the i th claim and Yi be the amount of the associated expense.
Let N be the total number of claims from the portfolio and let n be the number of policies
in the portfolio. Then N has a Poisson distribution with parameter 0.4n and S can be
written:
N
S  ( X i  Yi )
i 1
where { X i  Yi }
i 1 is a sequence of independent and identically distributed random
variables, independent of N . From this it can be seen that S has a compound Poisson
distribution with X i  Yi representing the ‘amount of the i th individual claim’. Standard
results can now be used to write down the following formulae for the moments of S :
E [ S ]  0.4n E [ X i  Yi ]
var [ S ]  0.4n E [ ( X i  Yi )2 ]  0.4n ( E [ X i2 ]  2 E [ X i Yi ]  E [ Yi2 ] )
In terms of  ,  and b , the moments of X i and Yi are as follows:
E [ Xi ]   /  E [ Yi ]  (b  50) / 2
E [ X i2 ]   (  1) /  2 E [ Yi2 ]  (b2  50b  2,500) / 3
E [ X i Yi ]  E [ X i ] E [ Yi ]
where the final relationship follows from the independence of X i and Yi .

(i) Now put:
  1 ;   0.01 ; b  100
into these formulae to show that:
E [S ]  70n and var[S ]  127.802 n
Hence, S has approximately a normal distribution with mean 70n and standard
deviation 127.80 n . The premium income is 80n and the smallest value of n is
required such that:
P (S  80n)  0.99
Standardising S in the usual way for a normal distribution, this becomes:
 S  70n 80n  70n 

P    0.99
 127.80 n 127.80 n 
The upper 99% point of a standard normal distribution is 2.326 and so the condition
for n is:
80n  70n
 2.326
127.80 n
which gives:
n  883.7
(or n  884 to the next higher integer).
(ii) For the insurance company, the worst possible combination of values for  ,  and
b is the combination which gives the highest possible values for E[S] and var[S].
To see this, let  and  denote the mean and the standard deviation of aggregate
claims and expenses from a single policy. Both  and  will be functions of  , 
and b and:
E [ S ]  n  and var [ S ]  n  2
Following the same steps as in part (i), the condition for n is:
(80   ) n
 2.326

which becomes:
n  [2.326 / (80   )]2
Hence, the highest value of n results from the highest values for  and 
(provided the highest value for  is less than 80). Now note that:
  0.4 E [ X i  Yi ] and  2  0.4 E [ ( X i  Yi )2 ]

From the formulae for the moments of X i and Yi given above,  and  are
maximised when  and b are as large as possible and  is as small as possible,
ie when:
  1.05 ;   0.009 ; b  110
This combination of values gives:
  78.67 and   144.14
so that n must be at least 63,546 for the insurance company to be at least 99% sure
that premium income will exceed claims and expenses outgo.

Chapter 20 Summary
Collective risk model with reinsurance
In the collective risk model, individual claims can be subject to a reinsurance agreement, either
proportional or excess of loss.
Under the collective risk model, the aggregate claim amount S is given by:
S  X1  X 2    X N
where X i is the amount of the i th claim and N is the total number of claims.
If reinsurance is in place, the insurer’s aggregate claims net of reinsurance can be represented
as:
SI  Y1  Y2    YN
where Yi is the amount of the i th claim paid by the insurer and N is defined as above. SI is a
compound random variable and:
E (SI )  E (N)E (Y )
var(SI )  E (N)var(Y )  var(N)[E (Y )]2
MSI (t )  MN [ln MY (t )]
The reinsurer’s aggregate claims can be represented as:
SR  Z1  Z2    ZN
where Z i is the amount of the i th claim paid by the reinsurer and N is defined as above. SR
is a compound random variable and:
E (SR )  E (N)E ( Z )
var(SR )  E (N)var(Z )  var(N)[E (Z )]2
MSR (t )  MN [ln MZ (t )]
Under individual excess of loss reinsurance, some of the claims may fall below the retention
level M . If this is the case, then some of the Z i will be zero. An alternative way of expressing
the reinsurer’s aggregate claims is as:
SR  W1  W2    WNR
where Wi  Z i | Z i  0 and NR is the number of non-zero claims, ie the number of claims in

which the reinsurer is involved.

Under an aggregate excess of loss arrangement with retention limit M , the maximum
payment made by the insurer is M . The insurer’s aggregate claim payment is:
 S if S  M
SI  
M if S  M
The reinsurer’s aggregate claim payment is:
0 if S  M
SR  
 S  M if S  M
Individual risk model

The individual risk model considers the payments made under each risk (eg policy) separately.
The model assumes that:
 the number of risks is fixed
 the risks are independent
 claim amounts from these risks are not necessarily IID
 N j , the number of claims from the j th risk is either 0 or 1.
For a portfolio containing n risks, the aggregate claim amount is given by:
S  Y1  Y2    Yn
where Y j denotes the aggregate claims from risk j . Since each Y j is the sum of a random
number (0 or 1) of random claim amounts, each Y j has a compound binomial distribution.
Suppose that q j is the probability of a claim from the j th risk. If a claim arises from the j th
risk, suppose that the claim amount random variable is X j . Then:
n
E (S )   q j  j
j 1
n
var(S)   q j 2j  q j (1  q j ) 2j 
 
j 1
n
MS  t    q j MX j (t )  (1  q j )
 
j 1
where  j  E ( X j ) and  2j  var( X j ) .
If, for a group of n risks, the probability of a claim is fixed and the claim amounts are IID
random variables, then the individual risk model is equivalent to a collective risk model where
S has a compound binomial distribution with N  Bin(n, q) .

Chapter 20 Practice Questions

20.1 The annual aggregate claims from a risk have a compound Poisson distribution with
parameter 250. Individual claim amounts have a Pareto distribution with parameters   4 and
  900 . The insurer effects proportional reinsurance with a retained proportion of 75%.
Calculate the variances of the total amounts paid by the insurer and by the reinsurer.
20.2 The aggregate claims from a risk have a compound Poisson distribution with parameter  .
Individual claim amounts (in £) have a Pareto distribution with parameters   3 and   1,000 .
The insurer of this risk calculates the premium using a premium loading factor of 0.2 (ie it charges
20% in excess of the risk premium).
The insurer is considering effecting individual excess of loss reinsurance with retention limit £1,000.
The reinsurance premium would be calculated using a premium loading factor of 0.3.
The insurer’s profit is defined to be the premium charged by the insurer less the reinsurance
premium and less the claims paid by the insurer, net of reinsurance.
(i) Show that the insurer’s expected profit before reinsurance is 100  .
(ii) Calculate the insurer’s expected profit after effecting the reinsurance, and hence find the
percentage reduction in the insurer’s expected profit.
(iii) Calculate the percentage reduction in the standard deviation of the insurer’s profit as a
result of effecting the reinsurance.
20.3 Aggregate annual claims from a portfolio of general insurance policies have a compound Poisson
distribution with Poisson parameter 20. Individual claim amounts have a uniform distribution
over the interval (0,200) . Excess of loss reinsurance is arranged so that the expected amount paid
by the insurer on any claim is 50.
Calculate the variance of the aggregate annual claims paid by the insurer.
20.4 A portfolio of policies consists of one-year term assurances on 100 lives aged exactly 30 and 200
lives aged exactly 40. The probability of a claim during the year on any one of the lives is 0.0004
for the 30 year olds and 0.001 for the 40 year olds.
If the sum assured on a life aged x is uniformly distributed between 1,000(x  10) and
1,000(x  10) , calculate the variance of the aggregate claims from this portfolio during the year
(assuming that policies are independent with regard to claims).

20.5 The number of claims from a given portfolio has a Poisson distribution with a mean of 1.5 per
month. Individual claim amounts have the following distribution:
Amount 200 300

Probability 0.65 0.35
An aggregate reinsurance contract has been arranged so that the insurer pays no more than 400
per month in total.
Assuming that the individual claim amounts are independent of each other and are also
independent of the number of claims, calculate the expected aggregate monthly claim amounts
for the insurer and the reinsurer.
20.6 A portfolio consists of 500 independent risks. For the i th risk, with probability 1  qi there are no
Exam style claims in one year, and with probability qi there is exactly one claim ( 0  qi  1 ). For all risks, if
there is a claim, it has mean  , variance  2 and moment generating function M(t ) . Let T be
the total amount claimed on the whole portfolio in one year.
(i) Determine the mean and variance of T . [4]
The amount claimed in one year on risk i is approximated by a compound Poisson random
variable with Poisson parameter qi and claims with the same mean  , the same variance  2 ,
and the same moment generating function M(t ) as above. Let T denote the total amount
claimed on the whole portfolio in one year in this approximate model.
(ii) Determine the mean and variance of T , and compare your answers to those in part (i).
[4]
Assume that qi  0.02 for all i , and if a claim occurs, it is of size  with probability one.
(iii) Derive the moment generating function of T , and show that T has a compound binomial
distribution. [2]
(iv) Determine the moment generating function of the approximating T , and show that T
has a compound Poisson distribution. [2]
[Total 12]

20.7 A company is analysing the number of accidents that occur each year on the factory floor. It
Exam style
believes that the number of accidents per year N has a geometric distribution with parameter
0.8, so that:
P(N  n)  0.8  0.2n , n  0,1,2,...
For each accident, the number of employees injured is Y , where Y  X  1 , and X is believed to
have a Poisson distribution with parameter 2.2.
The company has taken out an insurance policy, which provides a benefit of £1,000 to each
injured employee, up to a maximum of three employees per accident, irrespective of the level of
injury. There is no limit on the number of accidents that may be claimed for in a year.
(i) Show that E (S)  0.634 and var(S)  2.125 , where S is the total number of employees
claiming benefit in a year under this policy. [7]
(ii) Hence find the mean and variance of the aggregate amount paid out under this policy in a
year. [1]
[Total 8]
20.8 An insurance company offers accident insurance for employees. A total of 650 policies have been
Exam style
issued split between two categories of employees. The first category contains 400 policies, and
claims occur on each policy according to a Poisson process at a rate of one claim per 20 years, on
average. In this category all claim amounts are £3,000. In the second category, claims occur on
each policy according to a Poisson process at a rate of one claim per 10 years, on average. In this
category, the claim amount is either £2,000 or £3,000 with probabilities 0.4 and 0.6, respectively.
All policies are assumed to be independent. Let S denote the aggregate annual claims from the
portfolio.
(i) Calculate the mean, variance and coefficient of skewness of S . [4]
(ii) Using the normal distribution as an approximation to the distribution of S , calculate Y

such that the probability of S exceeding Y is 10%. [3]
(iii) The insurance company decides to effect reinsurance cover with aggregate retention
£100,000, so that the insurance company then pays out no more than this amount in
claims each year. In the year following the inception of this reinsurance, the numbers of
policies in each of the two groups remains the same but, because of changes in the
employment conditions of which the company was unaware, the probability of a claim in
group 2 falls to zero. Using the normal distribution as an approximation to the
distribution of S , calculate the probability of a claim being made on the reinsurance
treaty. [3]
[Total 10]

The solutions start on the next page so that you can

separate the questions and solutions.

Chapter 20 Solutions
20.1 The mean and variance of the gross claim amounts are:
 900
E(X )    300
 1 3
 2   9002
var( X )    180,000
  1)2 (  2) 32  2
So the mean and variance of the net claims for the direct insurer and the reinsurer are:
E (Y )  0.75  300  225
var(Y )  0.752  180,000  101,250
E (Z )  0.25  300  75
var(Z )  0.252  180,000  11,250
Using the formula for the variance of a compound Poisson random variable, the variances of the
aggregate claim payments made by the insurer and the reinsurer are:
var(SI )   E[Y 2 ]    var(Y )  [E (Y )]2   250[101,250  2252 ]  37,968,750

 
var(SR )  E[ Z 2 ]    var(Z )  [E (Z )]2   250[11,250  752 ]  4,218,750

 
20.2 (i) Expected profit before reinsurance
We have:

E(X )   500
 1
So the expected aggregate claim amount is:
E (S)  500 
The insurer’s premium income is:
1.2E (S)  1.2  500   600 
So the expected profit before reinsurance is:
600   500   100 

(ii) Expected profit after reinsurance
The reinsurance premium is given by 1.3E (SR ) , where:
E (SR )  E (Z )E (N)  E ( Z )
Now:

3  1,0003
E (Z )   (x  1,000)
(1,000  x)4
dx
1,000
Setting u  x  1,000 :
 3 
3  1,0003  1,000  3  2,0003
E (Z )   u du    u du
4
0 (2,000  u)  2,000  0 (2,000  u)4
Recognising this integral as the mean of the Pareto(3, 2000) distribution, we see that:
3
 1  2,000
E (Z )      125
2 31
and:
E (SR )  125
So the reinsurance premium is 1.3  125  162.5 .
Alternatively we could evaluate this integral using the substitution t  1000  x or using integration
by parts.
The insurer’s expected aggregate claim payment is:
E (SI )  E (S)  E (SR )  500   125  375
So the insurer’s expected profit after reinsurance is:
600   162.5  375  62.5
This is the insurer’s premium income, minus the premium paid by the insurer to the reinsurer,
minus the insurer’s expected aggregate claim payment.
The percentage reduction in the expected profit (which was 100  without reinsurance) is 37.5%.
(iii) Percentage reduction in standard deviation
In the absence of reinsurance, the insurer’s profit is equal to its premium income minus the
aggregate claim amount. Since the premium income is a fixed amount and only the cost of claims
is random, the variance of the profit is:
var(S)   E ( X 2 )

We have:
  2
E(X )   500 and var( X )   750,000
 1 (  1)2 (  2)
So:
E ( X 2 )  var( X )  [E ( X )]2  750,000  5002  1,000,000
and:
var(S)  1,000,000 
Hence the standard deviation of the profit is 1,000  .
With reinsurance, the insurer’s profit is equal to premiums charged less the reinsurance premium
less the net claims paid. Since the premiums are fixed amounts, the variance of the insurer’s
profit is:
var(SI )   E (Y 2 )
where:
1,000 
3  1,0003 3  1,0003
 
2 2
E (Y )  x dx  1,0002 dx
0 (1,000  x)4 1,000 (1,000  x)4
1,000 
x2 1
 3  1,0003  4
dx  3  1,0005  4
dx
0 (1,000  x) 1,000 (1,000  x)
The second integral is:
 
1  (1,000  x)3  1
 (1,000  x)4 dx  
 3
 
1,000 3  2,000
3
1,000
For the first integral, we can set u  1,000  x to give:
1,000 2,000 2,000

x2 (u  1,000)2  1 1,000 1,000,000  1
 (1,000  x)4 dx   u4
du     2 
 u u 3u 3  
1,000 24,000
0 1,000
Alternatively, we could integrate by parts (twice).
So:
3  1,0003 3  1,0005
E (Y 2 )    250,000 and var(SI )  250,000 
24,000 3  2,0003
Hence the standard deviation of the insurer’s profit is now 500  , which is a reduction of 50%.

The standard deviation is reduced by a greater percentage than the mean. This is very often the
case for excess of loss reinsurance.
20.3 We have X  U(0,200) and:
 X if X  M 0 if X  M
Y  Z 
M if X  M  X  M if X  M
The expected amount paid by the insurer on any claim is:

M 200
1 1
E (Y )   x
200
dx   M
200
dx  50
0 M
Solving this:
M
 x2   Mx 
200
     50
 400  0  200  M
M2 M2
 M  50
400 200
 M 2  400M  20,000  0
(400)  (400)2  4  1  20,000

 M  58.579 or 341.42
21
Since claims are a maximum of 200, M must be 58.579.
The variance of the aggregate annual claims paid by the insurer is:
var(S)  E (Y 2 )  20E (Y 2 )
where:
M 200
M
1
200
1  x3   M2 x 
 
2 2 2
E (Y )  x dx  M dx     
200 200  600   200  M
0 M 0
M3 M3
  M2 
600 200
58.5793 58.5793
  58.5792 
600 200
 2,761.42
Hence:
var(S)  20  2,761.42  55,228

20.4 For each age group, the individual claim amounts have a uniform distribution. So the mean and
variance of the individual claim distributions are:
E ( X )  12 (b  a)  1,000 x
2
and: 1 (b  a)2  20,000
var( X )  12 12
Using the individual risk model, the variance of the aggregate claim amount is:
 
n
var(S)   qi i2  qi (1  qi )i2
i 1
 2

 100 (0.0004)  20,000  (0.0004)(0.9996)  30,0002 
 12 
 200 (0.001)  20,000  (0.001)(0.999)  40,0002 

2
 12 
 37.32m  326.35m  363.67m
Alternatively, we could model the aggregate claim amount from each group as a compound
binomial random variable. For example, N  Bin(100,0.0004) for the 100 lives aged exactly 30.
We could then use the formula for var(S) from the collective risk model.
20.5 Under this reinsurance arrangement, we have:
S if S  400 0 if S  400
SI   SR  
400 if S  400  S  400 if S  400
where S is the total monthly claim amount.
Since individual claim amounts must be either 200 or 300, the possible values of SI are 0, 200,
300, and 400 and:
E (SI )  0  P(SI  0)  200  P(SI  200)  300  P(SI  300)  400  P(SI  400)
The insurer’s aggregate claim amount is 0 if there are no claims. So:
e 1.5  1.50
P(SI  0)  P(N  0)   e 1.5
0!
The insurer’s aggregate claim amount is 200 if there is one claim and the amount of the claim
is 200. So:
P(SI  200)  P(N  1, X1  200)
Since N and X1 are independent, we have:
e 1.5  1.51
P(SI  200)  P(N  1) P( X1  200)   0.65  0.975 e 1.5
1!

Similarly:
e 1.5  1.51
P(SI  300)  P(N  1, X1  300)   0.35  0.525 e 1.5
1!
Finally, the insurer’s aggregate claim amount is 400 if the total claim amount is 400 or more. This
probability can be calculated by subtraction as follows:
P(SI  400)  1  P(SI  0)  P(SI  200)  P(SI  300)
 1  e 1.5  0.975 e 1.5  0.525 e 1.5
 1  2.5 e 1.5
So:
E (SI )  0  e 1.5  200  0.975 e 1.5  300  0.525 e 1.5  400(1  2.5 e 1.5 )  255.52
We can now calculate E (SR ) using the result:
E (SR )  E (S)  E (SI )
We have:
E (S)  E ( X )  1.5 200  0.65  300  0.35  1.5  235  352.50
Hence:
E (SR )  352.50  255.52  96.98
20.6 This is part of Subject 106, April 2003, Question 9.
(i) Mean and variance of T
Let T  Y1  Y2    Y500 , where Yi is the total claim on the i th policy. Then:
E[T ]  E[Y1 ]  E[Y2 ]    E[Y500 ]
var[T ]  var[Y1 ]  var[Y2 ]    var[Y500 ]
Since E[Yi ]  qi  and var[Yi ]  qi 2  qi (1  qi ) 2 , we have:
500
E[T ]    qi [2]
i 1
500 500
var[T ]   2  qi   2  qi (1  qi ) [2]
i 1 i 1
since the risks are independent.

(ii) Mean and variance of T
Let C be the amount claimed in one year on a single risk. Then, according to the approximation:
C  X1  X 2    X N
where N  Poi(qi ) , E[ X ]   and var[ X ]   2 .
Also:
T  C1  C2    C500
where C i is the total amount claimed on the i th risk.
Using the formulae for the mean and variance of compound Poisson random variable:
E[C i ]   qi var[C i ]  qi ( 2   2 ) [2]
Since T is the sum of claims for the whole portfolio, we have:
500 500
E[T ]    qi var[T ]  ( 2   2 )  qi [2]
i 1 i 1
The mean is the same but the variance is larger than that obtained in part (i).
(iii) MGF
By definition we have:
MT (t )  E[etT ]  E[et (Y1 Y2   Y500 ) ]  MY1 (t )MY2 (t ) MY500 (t ) [½]
From the information given in the question, Yi is either 0 with probability 0.98 or  with
probability 0.02. We can therefore work out the moment generating function of Yi :
MY (t )  E[etY ]  et0  0.98  et   0.02  0.98  0.02et  [½]
Substituting this into the expression for the moment generating function for T , we get:
MT (t )  (0.98  0.02et  )500 [½]
This is of the form of the moment generating function for a compound binomial distribution with
parameters 500 and 0.02, and claim size distribution that is constant. [½]
(iv) Compound Poisson
By definition we have:

MT (t )  E[etT ]  E[et (C1 C2   C500 ) ]  MC1 (t )MC2 (t ) MC500 (t )

From the information given in the question, since C i has a compound Poisson distribution it has
moment generating function:
MCi (t )  exp qi (MX (t)  1)  exp  0.02(MX (t )  1) [½]
The random variable X takes the value  with probability 1, so:
M X (t )  et 
and:
MC i (t )  exp 0.02(et   1) [½]

 
Substituting this into the expression for the moment generating function for T , we get:
MT (t )  exp  0.02(et   1) exp 0.02(et   1)  exp  0.02(et   1)
     
 exp 10(et   1) [½]

 
This is of the form of the moment generating function for a compound Poisson distribution with
parameter 10, and claim size distribution that is constant. [½]
20.7 (i) Total number of claimants
The aggregate amount paid out by the company is 1,000S , where:
S  Z1    ZN is the total number of employees claiming benefit in a year
Y Y  3
Z
3 Y  3
Y  X 1
and:
X  Poisson(2.2)
Now:
P(Y  1)  P( X  0)  e 2.2 [½]
P(Y  2)  P( X  1)  2.2e 2.2 [½]
P(Y  3)  1  e 2.2  2.2e 2.2  1  3.2e 2.2 [½]

So:
E (Z )  1P(Y  1)  2P(Y  2)  3P(Y  3)
 
 e 2.2  4.4e 2.2  3 1  3.2e 2.2  3  4.2e 2.2  2.53463 [1]
and:
E (Z 2 )  12 P(Y  1)  22 P(Y  2)  32 P(Y  3)
 
 e 2.2  8.8e 2.2  9 1  3.2e 2.2  9  19e 2.2  6.89474 [1]
Hence the variance of Z is:
 
var(Z )  E Z 2   E (Z )   6.89474  2.534632  0.47041
2
[½]
To find E (N) and var(N) , we use the fact that N has a Type 2 negative binomial distribution with
parameters p  0.8 , q  0.2 and k  1 . Using the formulae for the moments given on page 9 the
Tables, we have:
kq 0.2
E (N)    0.25 [½]
p 0.8
kq 0.2
and: var(N)    0.3125 [½]
p2 0.82
Alternatively, we could derive the moment generating function of N , and then use MGF formulae
to derive the mean and variance of N .
The mean and variance of S are:
E (S)  E (Z )E (N)  2.53463  0.25  0.63366 [1]
and:
var(S)   E (Z ) var(N)  var(Z )E (N)  2.534632  0.3125  0.47041  0.25  2.12521

2
[1]
(ii) Mean and variance of the aggregate amount
So the mean and variance of 1,000S are:
E (1,000 S)  1,000E (S)  634 [½]
and:
var(1,000 S)  1,0002 var(S)  2,125,000 [½]

Alternatively, we could define S  Z1  Z2    ZN , where:
1,000Y if Y  3
Z
3,000 if Y  3
This would give us the aggregate claim amount directly.
20.8 (i) Mean, variance and coefficient of skewness
We have N1  Poisson(400  20
1 )  Poisson(20) , X  £3,000 , N  Poisson(250  1 )  Poisson(25)
1 2 10
and:
£2,000 with probability 0.4

X2  
£3,000 with probability 0.6
Working in £000s, we find that:
E ( X2 )   2  0.4    3  0.6   2.6
    
E X22  22  0.4  32  0.6  7 
E  X23    23  0.4    33  0.6   19.4 [1]
Let Si denote the annual aggregate claims from category i . Using the assumption that the
policies are independent and the result that, for a compound Poisson random variable T , the k th
central moment of T is given by E ( X k ) , we obtain:
E (S)  E (S1 )  E (S2 )  (20  3)  (25  2.6)  125  £125,000 [1]
var(S)  var(S1 )  var(S2 )
 1E ( X12 )  2E ( X22 )
  20  9    25  7 
 355
 £2355,000,000 [1]
skew(S)  skew(S1 )  skew(S2 )
 1E ( X13 )  2E ( X23 )
  20  27    25  19.4 
 1,025
 £3 (1,025  109 )

So the coefficient of skewness is:
skew(S) 1,025
  0.15324 [1]
 var(S) 3/2
3553/2
(ii) Calculate Y using a normal approximation
Assuming that S  N(125, 355) , we have:
 Y  125 
P(S  Y )  0.1  P  N(0,1)    0.1
 355 
Y  125
  1.2816 from page 162 Tables
355
 Y  149.147
So S exceeds £149,000 with a probability of approximately 0.1. [3]
(iii) Probability that reinsurer is involved
The expected value and variance of S are now the same as those of S1 . Working in £000s and
assuming that S  N(60, 180) , we obtain:
 100  60 
P(S  100)  P  N  0,1    1  (2.98)  1  0.9986  0.0014 [3]
 180 

Risk Models 1 & 2

Uploaded by

Copyright:

Available Formats

Risk Models 1 & 2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Risk Models 1 & 2

Uploaded by

Copyright:

Available Formats

CS2-19: Risk models 1 Page 1

1.2.1 Construct models appropriate for short-term insurance contracts in terms of

The Actuarial Education Company © IFE: 2019 Examinations

© IFE: 2019 Examinations The Actuarial Education Company

1 General features of a product

1.1 Insurable interest

 a risk must be of a financial and reasonably quantifiable nature.

1.2 Insurable risk

 Individual risk events should be independent of each other.

 There should be an ultimate limit on the liability undertaken by the insurer.

Other characteristics that most general insurance products share are:

 A claim occurring does not bring the policy to an end.

The Actuarial Education Company © IFE: 2019 Examinations

© IFE: 2019 Examinations The Actuarial Education Company

2 Models for short-term insurance contracts

2.1 The basic model

A short-term insurance contract can be defined as having the following attributes:

 The insurance company receives from the policyholder(s) a premium.

We will see shortly what these terms mean.

where the summation is taken to be zero if N is zero.

The Actuarial Education Company © IFE: 2019 Examinations

This decomposition of S allows consideration of claim numbers and claim amounts

To define a compound distribution, we need to know:

2.2 Discussion of the simplifications in the basic model

© IFE: 2019 Examinations The Actuarial Education Company

An important element in models for long-term insurance is a rate of interest since, as

We will ignore interest in the models used in this chapter.

2.3 Notation and assumptions

 the random variables { X i }N

 the random variable N is independent of { X i }N

The Actuarial Education Company © IFE: 2019 Examinations

In words these assumptions mean that:

G( x ) and F ( x ) shall denote the distribution functions of S and X i , respectively, so that:

The k th moment, ( k  1,2,3... ) of X i about zero will be denoted mk .

Using this notation:

E ( X i )  m1 and var( X i )  m2  m12

© IFE: 2019 Examinations The Actuarial Education Company

3 The collective risk model

3.1 The collective risk model

S is said to have a compound distribution.

3.2 Distribution functions and convolutions

{S  x and N  0} (ie no claims)

or {S  x and N  1} (ie one claim of amount  x )

or {S  x and N  2} (ie two claims which total  x )

or {S  x and N  r } (ie r claims which total  x )

and so on. These events are mutually exclusive and exhaustive.

The Actuarial Education Company © IFE: 2019 Examinations

Determine the distribution function of the aggregate annual claim amount, S .

S can take the values 0, 1, 2, 3 or 4.

S will only equal 0 if N  0 and this has probability 0.6. So:

P(S  0)  P(S  0)  P(N  0)  0.6

S will equal 1 if there is one claim for amount 1. So:

 P(N  1)P( X  1) by independence

P(S  1)  P(S  0)  P(S  1)  0.6  0.15  0.75

P(S  2)  P(S  1)  P(S  2)  0.75  0.3  0.5  0.1  0.52  0.925

P(S  3)  P(S  2)  P(S  3)  0.925  0.1  2  0.52  0.975

This is written mathematically as fZ  fX * fY .

© IFE: 2019 Examinations The Actuarial Education Company

fZ (z)   fX  x  fY (z  x) dx for continuous random variables