0 INTRODUCTION

Chapter 8 Credibility theory

0 Introduction

In this chapter we will look at credibility theory, which is a technique that can be used to determine premiums or claim frequencies in general insurance. We will look only at the Bayesian approach to credibility.

0 INTRODUCTION

The contents of this chapter are: Credibility The credibility premium formula The credibility factor Bayesian credibility Introduction The Poisson/gamma model Numerical illustrations of the Poisson/gamma model The normal/normal model Further remarks on the normal/normal model Discussion of the Bayesian approach to credibility

1 CREDIBILITY

1
1.1

Credibility
The credibility premium formula

The basic idea underlying the credibility premium formula is intuitively very simple and very appealing. Consider an extremely simple example: Suppose a local authority in a small town has run a eet of ten buses for a number of years. The local authority wishes to insure this eet for the coming year against claims arising from accidents involving these buses. The pure premium for this insurance needs to be calculated, i.e. the expected cost of claims in the coming year. The data for the past ve years for this particular eet of buses show that the average cost of claims per annum (for the ten buses) has been 1,600. Suppose that, in addition to this information, there is data relating to a large number of local authority bus eets from all over the United Kingdom which show that the average cost of claims

1 CREDIBILITY

per annum per bus is 250, so that the average cost of claims per annum for a eet of ten buses is 2,500. However, while this gure of 2,500 is based on many more eets of buses than the gure of 1,600, some of the eets of buses included in this large data set operate under very dierent conditions (which are thought to aect the number and/or size of claims, e.g. in large cities or in rural areas) from the particular eet which is of concern here.

1 CREDIBILITY

There are two extreme choices for the pure premium for the coming year: 1. 1,600 could be chosen on the grounds that this estimate is based on the most appropriate data, whereas the estimate of 2,500 is based on less relevant data, or, 2. 2,500 could be chosen on the grounds that this is based on more data and so is, in some sense, a more reliable gure. The credibility approach to this problem is to take a weighted average of these two extreme answers, i.e. to calculate the pure premium as: Z 1, 600 + (1 Z ) 2, 500 where Z is some number between zero and one. Z is known as the credibility factor.

1 CREDIBILITY

We express this example a little more formally. The problem is to estimate the expected aggregate claims, or, possibly, just the expected number of claims, in the coming year from a risk. These policies are, typically, short term policies and, for convenience, the term of the policies will be taken to be one year, although it could equally well be any other short period. The following information is available: X is an estimate of the expected aggregate claims (number of claims) for the coming year based solely on data from the risk itself is an estimate of the expected aggregate claims (number of claims) for the coming year based on collateral data, i.e. data for risks similar to, but not necessarily identical to, the particular risk under consideration.

1 CREDIBILITY

The credibility premium formula (or credibility estimate of the aggregate claims/number of claims) for this risk is: ZX + (1 Z ) where Z is a number between zero and one and is known as the credibility factor.

1 CREDIBILITY

1.2

The credibility factor

The credibility factor Z is just a weighting factor. Its value reects how much trust is placed in the data from the risk itself (X ) compared with the data from the larger group (). As an estimate of next years expected aggregate claims or number of claims, the higher the value of Z , the more trust is placed in X compared with and vice versa. It can be seen that, in general terms, the credibility factor would be expected to behave as follows: The more data there are from the risk itself, the higher should be the value of the credibility factor. The more relevant the collateral data, the lower should be the value of the credibility factor.

1 CREDIBILITY

One nal point to be made about the credibility factor is that, while its value should reect the amount of data available from the risk itself, its value should not depend on the actual data from the risk itself, i.e. on the value of X . If Z were allowed to depend on X then any estimate of the aggregate claims/number of claims, say , could be written in the form of ZX + (1 Z ) by choosing Z to be equal to ( )/(X ). Because ZX + (1 Z ) = = X+ X X+ X 1 X X X

X = X =

1 CREDIBILITY

10

The problems remain of how to measure the relevance of collateral data and how to calculate the credibility factor Z . There are two approaches to these problems: Bayesian credibility empirical Bayes credibility theory In this chapter we will only study Bayesian credibility.

2 BAYESIAN CREDIBILITY

11

2
2.1

Bayesian credibility
Introduction

The Bayesian approach to credibility involves the following steps: Prior parameter distribution Likelihood function Posterior parameter distribution Loss function Parameter estimate This study of the Bayesian approach to credibility theory is illustrated by considering two models: the Poisson/gamma model which looks at claim frequencies. the normal/normal model which looks at claim amounts.

2 BAYESIAN CREDIBILITY

12

2.2

The Poisson/gamma model

Suppose the claim frequency for a risk, i.e. the expected number of claims in the coming year, needs to be estimated. The random variable X represents the number of claims in the coming year from a risk. The distribution of X depends on the xed, but unknown, value of a parameter, The conditional distribution of X given is P oisson(). The prior distribution of is gamma(, ). x1, x2, . . . , xn are past observed values of X , which will be denoted x. The problem is to estimate given the data x, and the estimate wanted is the Bayes estimate with respect to quadratic loss, i.e. E (|x). Recall from Chapter 2 that the estimate with respect to a quadratic loss function is the mean of the posterior distribution.

2 BAYESIAN CREDIBILITY

13

The posterior distribution of given x is gamma(+ n i=1 xi , + n). and + n i=1 xi E (|x) = +n + = +n +n n = +n n = +n = Z where Z= n +n
n i=1 xi

n i=1 xi

n i=1 xi

n
n i=1 xi

+n

n
n i=1 xi

n + 1 +n + (1 Z )

The observed mean number of claims is

The mean number based on prior beliefs is the mean of the prior gamma distribution, / . Suppose n = 0 Z = 0 E (|x) =

2 BAYESIAN CREDIBILITY

14 n i=1 xi

Z = 1 E (|x) =

The value of Z depends on the amount of data available for the risk, n, and the collateral information, through . As n increases the sampling error of estimate for decreases.
n i=1 xi

as an

reects the variance of the prior distribution for . Thus Z reects the relative reliability of the two alternative estimates of .

2 BAYESIAN CREDIBILITY

15

2.3

Numerical illustrations of the Poisson/gamma model

In this section some simple numerical examples relating to the Poisson/gamma model will be considered which will help to make some further points about this model. gamma(100, 1). (Note that this distribution has mean 100 and standard deviation 10.) The actual number of claims arising each year from this risk are given. Figure 1 on page 14 shows the credibility factor in successive years. 0, 1/2, 2/3, 3/4, 4/5, 5/6, 6/7, 7/8, 8/9, 9/10, 10/11 Figure 2 on page 15 shows the credibility estimate of the number of claims in successive years for the gamma(100, 1) prior distribution for .

2 BAYESIAN CREDIBILITY

16

The following simple observations can be made about Figures 1 and 2. The credibility factor increases with time. The estimated number of claims increasing with time until it reaches the level the actual claim numbers after eight years. Suppose the prior distribution of is gamma(500, 5) rather than gamma(100, 1) Figures 3 and 4 on page 16 show that the credibility factor and the estimated number of claims for the gamma(500, 5) prior have the same general features as for the gamma(100, 1) prior. The most obvious dierence between these two cases is that the credibility factor increases more slowly for the gamma(500, 5) prior than it does for the gamma(100, 1) prior. This feature can be explained in terms of credibility theory.

2 BAYESIAN CREDIBILITY

17

E (X ) = E [E (X |)] = E () v.s. E (X |X )

2 BAYESIAN CREDIBILITY

18

2.4

The normal/normal model

In this section we will estimate the pure premium, i.e. the expected aggregate claims, for a risk. Let X be a random variable representing the aggregate claims in the coming year for this risk and the following assumptions are made: The distribution of X depends on the xed, but unknown, value of a parameter, The conditional distribution of X given is 2 N (, 1 ). The uncertainty about the value of is modelled in the usual Bayesian way by regarding it as a random variable. 2 The prior distribution of is N (, 2 ). 2 2 The values of ,1 and 2 are known. n past values of X have been observed, which will be denoted x. If the value of were known, the correct pure premium for this risk would be E (X |) =

2 BAYESIAN CREDIBILITY

19

The problem then is to estimate E (X |) given x. E [E (X |)|x] = E (|x) 2 2 x 1 + n2 = 2 + n 2 1 2 2 2 1 n2 = 2 x + 2 2 2 1 + n2 1 + n2 = Zx + (1 Z ) where: Z= n 2 / 2 n + 1 2

This is a credibility estimate of E (|x). There are some further points to be made about the credibility factor Z . It is always between zero and one. It is an increasing function of n, the amount of data available. It is an increasing function of 2, the standard deviation of the prior distribution. These features are all exactly what would be expected for a credibility factor.

2 BAYESIAN CREDIBILITY

20

2.5

Further remarks on the normal/normal model

In this section the model discussed in last section will be considered, without making any dierent assumptions, but in a slightly dierent way. The assumptions are: Let X1, X2, . . . , Xn, Xn+1, . . . be random variables representing the aggregate claims in successive years. The distribution of Xj depends on the xed, but unknown, value of a parameter, The conditional distribution of Xj given is 2 N (, 1 ). Given , the random variables {Xj } are independent. (*) 2 The prior distribution of is N (, 2 ). The value of X1, X2, . . . , Xn have already been observed and the expected aggregate claims in the coming year, i.e. E [Xn+1|x1, x2, . . . , xn] need to be estimated.

2 BAYESIAN CREDIBILITY

21

Some important consequences are: Given , the random variables {Xj } are i.i.d. The random variables {Xj } are (unconditionally) identically distributed. Because

P (Xj y ) =

f ()FX | (y )d

y ( )2 d exp 2 2 2 1 2 f (x|)f ()d

f (x ) =

2 BAYESIAN CREDIBILITY

22

The random variables {Xj } are not (unconditionally) independent. E [X1X2] = = = = E (X1) = = = E (X2) = = = E (X1X2) = E [E (X1X2|)] E [E (X1|)E (X2|)] E ( 2 ) 2 2 + 2 E [E (X1|)] E ( ) E [E (X2|)] E ( ) E (X1)E (X2)

The relationship between X1 and X2 is that their means are chosen from a common distribution. If this mean, is known, then this relationship is broken and there exists conditional independence.

2 BAYESIAN CREDIBILITY

23

2.6

Discussion of the Bayesian approach to credibility

The approach used in the Poisson/gamma and normal/normal models was essentially the same. This approach can be summarized as follows: The problem is stated, e.g. to determine a claim size or claim number distribution. In each case the aim is to estimate some quantity which characterizes this distribution, e.g. the mean number of claims or the mean claim size. Some model assumptions are then made within a Bayesian framework, e.g. it is assumed that the claim number distribution is Poisson and that the unknown parameter of the Poisson distribution has a gamma distribution with specied parameters. A (Bayesian) estimate of the particular quantity is derived. This estimate is then shown to be in the form of a credibility estimate.

2 BAYESIAN CREDIBILITY

24

This approach has been very successful in these two cases. It has made the notion of collateral data very precise (by interpreting it in terms of a prior distribution) and has given formulae for the calculation of the credibility factor. The two drawbacks of this approach are: The rst diculty is whether a Bayesian approach to the problem is acceptable, and, if so, what values to assign to the parameters of the prior distribution. The second diculty is that even if the problem ts into a Bayesian framework, the Bayesian approach may not work in the sense that it may not produce an estimate which can readily be rearranged to be in the form of a credibility estimate.