Aggregate Loss Models

Chapter 9

Stat 477 - Loss Models

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 1 / 22

 Objectives

Objectives

Individual risk model

Collective risk model
Computing the aggregate loss models
Approximate methods
Effect of policy modifications
Chapter 9 (sections 9.1 - 9.7; exclude 9.6.1 and examples 9.9 and
9.11)
Assignment: read section 9.7

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 2 / 22

 Individual risk model

Individual risk model

Consider a portfolio of n insurance policies.

Denote the loss, for a fixed period, for each policy i by Xi , for
i = 1, . . . , n.
Assume these lossses are independent and (possibly) identically
distributed.
The aggregate loss, S, is defined by the sum of these losses:

S = X1 + X2 + + Xn .

It is possible that here a policy does not incur a loss so that each Xi
has a mixed distribution with a probability mass at zero.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 3 / 22

 Individual risk model Alternative representation

Alternative representation
Assume that the losses are also identically distributed say as X. Then
we can write X as the product of a Bernoulli I and a positive
(continuous) random variable Y :

X = IY

I = 1 indicates there is a claim, otherwise I = 0 means no claim. Let

Pr(I = 1) = q and hence Pr(X = 0) = 1 q.
In addition, assume E(Y ) = Y and Var(Y ) = Y2 .
Typically it is assumed that I and Y are independent so that

E(X) = E(I)E(Y ) = qY

and
Var(X) = q(1 q)2Y + qY2 .

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 4 / 22

 Individual risk model Mean and variance

Mean and variance in the individual risk model

The mean of the aggregate loss can thus be written as

E(S) = nqY

and its variance is

Var(S) = nq(1 q)2Y + nqY2 .

Example: Consider a portfolio on 1,000 insurance policies where each

policy has a probability of a claim of 0.15. When a claim occurs, the
amount of claim has a Pareto distribution with parameters = 3 and
= 100. Calculate the mean and variance of the aggregate loss.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 5 / 22

 Individual risk model Illustrative examples

Illustrative examples

Example 1: An insurable event has a 10% probability of occurring

and when it occurs, the amount of the loss is exactly 1,000. Market
research has indicated that consumers will pay at most 115 for
insuring this event. How many policies must a company sell in order
to have a 95% chance of making money (ignoring expenses)? Assume
Normal approximation.
Example 2: Exercise 9.20 of textbook.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 6 / 22

 Individual risk model Exact distribution using convolution

Exact distribution of S using convolution

Consider the case of S = X1 + X2 . The density of S can be

computed using convolution as:
Z s Z s
f 2 (s) = fS (s) = f1 (s y)f2 (y)dy = f2 (s y)f1 (y)dy
0 0

To extend to n dimension, do it recursively. Let f (n1) (s) be the

(n 1)-th convolution so that
Z s
n
f (s) = fS (s) = f (n1) (s y)fn (y)dy
Z0 s
= fn (s y)f (n1) (y)dy
0

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 7 / 22

 Individual risk model Mixed random variable

The case of the mixed random variable

Often, Xi has a probability mass at zero so that its density function is

given by 
qi , for x = 0,
fi (x) =
(1 qi )fY (x), for x > 0,
where fY () is a legitimate density function of a positive continuous
random variable. Here, qi is interpreted as the probability that the
loss is zero.
In this case, use the cumulative distribution function:
Z s Z s
2
F (s) = FS (s) = F1 (s y)dF2 (y) = F2 (s y)dF1 (y)
0 0

To extend to n dimension, do it recursively.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 8 / 22

 Individual risk model The case of the Exponential

The case of the Exponential distribution

Consider the case where Xi has the density function expressed as


0.10, for x = 0,
fi (x) =
0.90fY (x), for x > 0,

where Y is an Exponential with mean 1, for i = 1, 2.

Derive expressions for the density and distribution functions of the sum
S = X1 + X2 .

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 9 / 22

 Individual risk model Approximation by CLT

Approximating the individual risk model

For large n, according to the Central Limit Theorem, S can be
approximated with a Normal distribution.
Use the results on slides p. 5 to compute the mean and the variance
of the aggregate loss S in the individual risk model.
Then, probabilities can be computed using Normal as follows:
" #
S E(S) S E(S)
Pr(S s) = Pr p p
Var(S) Var(S)
" #
S E(S)
Pr Z p
Var(S)
!
S E(S)
= p
Var(S)

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 10 / 22

 Individual risk model Illustrative example

Illustrative example 1

An insurer has a portfolio consisting of 25 one-year life insurance policies

grouped as follows:

Insured amount bk Number of policies nk

1 10
2 5
3 10

The probability of dying within one year is qk = 0.01 for each insured, and
the policies are independent.
The insurer sets up an initial capital of $1 to cover its future obligations.
Using Normal approximation, calculate the probability that the insurer will
be able to meet its financial obligation.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 11 / 22

 Collective risk model

Collective risk model

Let Xi be the claim payment made for the ith policyholder and let N
be the random number of claims. The insurers aggregate loss is
N
X
S = X1 + + XN = Xi .
i=1

This is called the Collective Risk Model.

If N , X1 , X2 , ... are independent and the individual claims Xi are
i.i.d., then S has a compound distribution.
N : frequency of claims; X: the severity of claims.
Central question is finding the probability distribution of S.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 12 / 22

 Collective risk model Properties

Properties of the collective risk model

X is called the individual claim and assume has moments denoted by

k = E(X k ).
Mean of S: E(S) = E(X)E(N ) = 1 E(N )
Variance of S: Var(S) = E(N )Var(X) + Var(N )21
MGF of S: MS (t) = MN [log MX (t)]
PGF of S: PS (t) = PN [PX (t)]
CDF of S: Pr(S s) =
P
n=0 Pr(S s|N = n)Pr(N = n)

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 13 / 22

 Collective risk model Exponential/Geometric

Example - Exponential/Geometric

Suppose Xi Exp(1) and N Geometric(p).

p 1
MGF of N : MN (t) = ; MGF of X: MX (t) =
1 qet 1t
Mean of S: E(S) = q/p
Variance of S: Var(S) = q/p + q/p2 = (q/p)(1 + 1/p)
p p p
MGF of S: MS (t) = = =p+q
1 qMX (t) 1 q/(1 t) pt
Observe that the mgf of S can be written as a weighted average: p
mgf of r.v. 0 + q mgf of Exp(p) r.v.
Thus, FS (s) = p + q(1 eps ) = 1 qeps for s > 0.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 14 / 22

 Collective risk model Poisson number of claims

Poisson number of claims

If N Poisson(), so that is the average number of claims, then

the resulting distribution of S is called a Compound Poisson.
MGF of N : MN (t) = exp et 1



It can then be shown that:

Mean of S: E(S) = E(X) = 1
Variance of S: Var(S) = E(X 2 ) = 2
MGF of S: MS (t) = exp [ (MX (t) 1)]

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 15 / 22

 Collective risk model Illustrative example

Illustrative example 2

Suppose S has a compound Poisson distribution with = 0.8 and

individual claim amount distribution

x Pr(X = x)
1 0.250
2 0.375
3 0.375

Compute fS (s) = Pr(S = s) for s = 0, 1, . . . , 6.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 16 / 22

 Collective risk model Panjers recursion formula

Panjers recursion formula

Let aggregate claims S have a compound distribution with

integer-valued non-negative claims with pdf p(x) for x = 0, 1, 2, . . ..
Let the probability of n claims satisfies the recursion relation
pk = (a + nb )pk1 for k = 1, 2, . . . for some real a and b. Recall this is
the (a, b, 1) class of distribution.
Then, for the probability of total claims s, we have
s  
1 X bh
fS (s) = a+ p(h)fS (s h).
1 ap(0) s
h=1

Pr(N = 0), if p(0) = 0
For s = 0, we have fS (0) = .
MN [log p(0)] , if p(0) > 0

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 17 / 22

 Collective risk model The case of Compound Poisson

Panjers recursion for Compound Poisson

This is the case where the number of claims is Poisson.

Starting value:

Pr(N = 0), if p(0) = 0
fS (0) = .
MN [log p(0)] , if p(0) > 0

With Poisson():
s
1X
fS (s) = hp(h)fS (s h).
s
h=1

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 18 / 22

 Collective risk model Illustrative example

Back to illustrative example 2

Same as previous example, but now we solve using Panjers recursion.
Effectively, the recursion formula boils down to
1
fS (s) = [0.2fS (s 1) + 0.6fS (s 2) + 0.9fS (s 3)].
s
[Why??] Initial value is: fS (0) = Pr(N = 0) = e0.8 = 0.44933.
Successive values satisfy:

fS (1) = 0.2fS (0) = 0.2e0.8 = 0.089866

1
fS (2) = [0.2fS (1) + 0.6fS (0)] = 0.32e0.8 = 0.14379
2
1
fS (3) = [0.2fS (2) + 0.6fS (1) + 0.9fS (0)] = 0.3613e0.8 = 0.16236
3
and fS (4) = 0.0499, fS (5) = 0.04736, and fS (6) = 0.03092. [calculated
by hand - verify!]
Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 19 / 22
 Collective risk model Approximating the compound distribution

Use CLT to approximate the compound distribution

If the average number of claims is large enough, we may use the

Normal approximation to estimate the distribution of S.
All you need are the mean and variance of the aggregate loss [slides
p. 13]

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 20 / 22

 Collective risk model Closed under convolution

Compound Poisson is closed under convolution

Theorem 9.7 of book.
If S1 , S2 , . . . , Sm are independent compound Poisson random
variables with Poisson parameters i and claim distributions (CDF)
Fi , for i = 1, 2, . . . , m, then

S = S1 + S2 + + Sm

also has a compound Poisson distribution with parameter

= 1 + 2 + + m and individual claim (severity) distribution
m
X i
F (x) = Fi (x).

i=1

You add the number of claim parameters, and the individual claim
distribution is a weighted sum.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 21 / 22

 Collective risk model Illustrative example

Illustrative example 3

You are given S = S1 + S2 , where S1 and S2 are independent and have

compound Poisson distributions with 1 = 3 and 2 = 2 and individual
claim amount distributions:

x p1 (x) p2 (x)
1 0.25 0.10
2 0.75 0.40
3 0.00 0.40
4 0.00 0.10

Determine the mean and variance of the individual claim amount for S.

Chapter 9 (Stat 477) Aggregate Loss Models Brian Hartman - BYU 22 / 22