Tutorial 3: Credibility Theory

1. (April 2000 – Q9) The annual aggregate claims, X , from a portfolio of insurance policies, are
assumed to have the normal distribution with unknown mean μ and known variance σ 2. Prior
information is such that the mean is assumed to have a normal distribution with known mean η
and known variance τ 2, i.e. X ∨( M =μ ) N ( μ , σ 2 ) and M N ( η , τ 2) .
(a) Independent aggregate claims over the last n years are denoted by x 1, x 2, …, x n.
(i) Derive the posterior distribution of the mean, i.e. M ∨( X=x ).
(ii) Write down the Bayesian point estimate of μ under a quadratic loss function.
(iii) Show that this estimate can be expressed in the form of a credibility estimate, and
derive the form of the credibility factor.
(iv) Determine the limiting form for this estimate as n increases. (8)
(b) The following denote the annual aggregate claims for two companies over five years
Year 1 Year 2 Year 3 Year 4 Year 5
Company A 217 250 249 239 265
Company B 196 239 233 222 244

(i) Determine the Bayes credibility estimate of the risk premium based on the modelling
assumptions of part (a) above, in the two separate cases:
Case 1 Case 2
2
σ η τ
2
σ
2
η 2
τ
Company A 400 270 2 500 Company A 400 270 225
Company B 400 260 2 500 Company B 400 260 225

(ii) Comment on the effect of changing the variance of the hypothetical means, i.e. τ 2.
(iii) Determine the empirical Bayes credibility estimate of the risk premium for each of the
two companies, using the EBCT Model 1. (10)
[Total 18]

Solution 1
a) Suppose X i ∨( M =μ ) N ( μ , σ 2) , independently for i=1 , 2 ,… , n and M N ( η , τ 2) , then

(∏ { ( ) }) { ( )} { (
n 2 n
1 −1 x i−μ 1 −1 μ−η 2
−1 −2
f M ∨ X ( μ∨x ) ∝ f X ∨M ( x∨μ ) f M ( μ )= exp × exp ∝exp ∑
i=1 √2 π σ 2
2 σ √2 π τ 2
2 τ 2 i=1

where

( )(
n

∑ xi
) ( )
−1 2 2 −1 2 2
' i=1 η n 1 η σ +n x τ ( ' ) 2 n 1 σ τ
η= 2
+ 2 2
+ 2 = 2 2
, τ = 2+ 2 = 2 2
σ τ σ τ σ +nτ σ τ σ +nτ

Recognize that M ∨( X=x ) belongs to the family of Normal distributions, i.e.

1
 ( )
2 2 2 2
ησ +n xτ σ τ
M ∨( X=x ) N 2 2
, 2 2
σ +n τ σ +n τ

The posterior mean – point estimate under the quadratic loss function – is
2 2 2 2
η σ +n x τ nτ σ
E [ M ∨ X =x ] = 2
( ) 2
= 2 2
×x+ 2 2
× η=Z x+ (1−Z ) η ,
σ +n τ σ +nτ σ +n τ
2
nτ n
where Z= = is the credibility factor.
σ +n τ n+ ( σ 2 /τ 2 )
2 2

lim Z =1⇒ lim E [ M ∨ ( X=x ) ] =x

n→∞ n→ ∞

b) For Company A: x A=244, and for Company B: x B =226.8

Case 1 Case 2
n 5 5
Z= =0.9690 =0.7377
n+ ( σ /τ )
2 2
5+ ( 400 /2 500 ) 5+ ( 400 /225 )
Credibility premium for Company A:
244.81 250.82
Z x A + ( 1−Z ) ×270=¿
Credibility premium for Company B:
227.83 235.51
Z x B + ( 1−Z ) ×260=¿

The larger variance of the hypothetical means in case 1 suggests that the prior knowledge is vague.
This is reflected in the lower weighting given to the mean of the prior distribution relative to the
sample mean.

For an empirical Bayesian approach, the linear credibility premium formula is

n
E [ m ( Θ )∨X k ]=Z X k + ( 1−Z ) ^
^ E [ m ( Θ ) ] , where Z=
E [ s 2 (Θ ) ]
^
n+
^ (m (Θ ) )
Var

{∑ }
5 5
1 1
^ ( θk ) =x k = ∑ x kj
m
2
^s ( θk ) = x 2kj −5 x2k
5 j=1 4 j=1

Company A 244 314

Company B 226.8 363.7

2 2
1 ^ [ s 2 ( Θ ) ]= 1 ∑ s^ 2 ( θ k )=338.85
E [ m (Θ ) ] = ∑ m
^ ^ ( θ k )=235.4 , E
2 k=1 2 k=1
2
^ ( m (Θ )) = 1 ∑ ( m
Var ^ [ m ( Θ ) ] )2− 1 E
^ ( θ k )− E ^ [ s 2 ( Θ ) ]=80.15
2−1 k=1 5

5
Z= =0.5418
338.85
5+
80.15

2
Hence, the credibility premium for Company A is 0.5418 x A +0.4582 ^
E [ m ( Θ ) ] =240.06, and the
credibility premium for Company B is 0.5418 x B +0.4582 ^
E [ m (Θ ) ] =230.74 .

PS: I believe that the examiner’s mark allocation is too generous.

2. (April 2004 – Q7(ii)) An insurer decides to use EBCT Model 1, where the credibility premium
combines the mean for the claims of a particular risk with an estimated value of E [ m (Θ ) ], the
overall average of claim amounts. Let x ij denote the aggregate claim amount for risk i=1 , 2 ,3 in
year j=1 ,2 , 3 , 4 , 5. The table below shows summary statistics of the observed data.
5
xi ∑ ( xij−x i )2
j=1
Risk 1 (i=1) 122 2 848
Risk 2 (i=2) 164 1 628
Risk 3 (i=3 ) 106 1 887

Derive the credibility factor, and calculate the credibility premium for Risk 1. [4]

Solution 2
The estimate of E [ m (Θ ) ] is x=( 122+164 +106 ) /3=130.667

The estimate of E [ s2 ( Θ ) ] is ( 2848

4
+
4
+
4 )
1628 1887
/3=530.25

3
1 1
The estimate of Var ( m (Θ )) is ∑
2 i=1
(
2
x i−x ) − ( 530.25 )=791.28
5

5
=0.8818
The credibility factor is 530.25 , such that the credibility premium for Risk 1 is
5+
791.28
0.8818 ×122+ ( 1−0.8818 ) ×130.667=123.02

3. The following table gives the aggregate claim amounts paid out (in millions) by four insurance
companies under a certain type of fire insurance over a period of five years.
Insurer / Year 1 2 3 4 5
1 41 47 54 30 51
2 44 32 29 39 29
3 16 47 12 38 33
4 6 18 30 22 21

Assuming that the EBCT Model 1 is appropriate, use the data to find estimates for E [ m (Θ ) ],
E [ s2 ( Θ ) ] and Var ( m (Θ )) . Hence, estimate the empirical Bayes credibility premium of each of
the insurers for the coming year. [10]

3
4. Let m ( θ )=E [ X ∨( Θ=θ ) ] and s ( θ )=Var ( X ∨( Θ=θ ) ). For each of the Bayesian models
2

discussed in this chapter, i.e. the Binomial | Beta model, the Poisson | Gamma model and the
Normal | Normal model, derive E [ m (Θ ) ], E [ s2 ( Θ ) ] and Var ( m (Θ )) . Hence, derive the credibility
premium formula, where the credibility factor, Z , is defined as below. (13)

n
Z=
n+ E [ s ( Θ ) ] /Var ( m ( Θ ) )
2

Solution 4
The credibility premium formula is defined as
^ n
E [ m (Θ )∨X ] =Z X + ( 1−Z ) E [ m ( Θ ) ] , where Z=
n+ E [ s ( Θ ) ] /Var ( m ( Θ ) )
2

For the Binomial | Beta model, we have X i ∨( Θ=θ ) Bern ( θ ), independently for i=1 , 2 ,… , n , and
Θ Beta ( α , β ).
m ( θ )=E [ X i∨( Θ=θ ) ]=θ ⇒ m ( Θ )=Θ Beta ( α , β )
α αβ
E [ m (Θ ) ] = , Var ( m (Θ )) =
α+β 2
( α + β ) ( α + β +1 )

s2 ( θ )=Var ( X i∨( Θ=θ ) )=θ ( 1−θ )

( )
2
α αβ α α ( α + β ) ( α + β+1 )
E [ s ( Θ ) ]=E [ Θ ] −E [ Θ ] =E [ Θ ] −(Var ( Θ ) + ( E [ Θ ]) )=
2 2 2
− + = −
α + β ( α + β ) ( α + β+ 1 ) ( α + β )
2 2
( α + β )2 ( α + β +1 ) (

n n n
Z= = =
E [ s (Θ)] n+α + β
2 2
αβ ( α + β ) ( α + β+ 1 )
n+ n+ ×
Var ( m ( Θ ) ) ( α + β )( α + β +1 ) αβ

∴^
E [ m ( Θ ) ∨X ] =
n
n+ α + β
X+
α+ β α
n+ α + β α + β ( )
For the Poisson | Gamma model, X i ∨( Θ=θ ) Poisson ( θ ), independently for i=1 , 2 ,… , n , and
Θ Gamma ( α , β ).
m ( θ )=E [ X i∨( Θ=θ ) ]=θ ⇒ m ( Θ )=Θ Gamma ( α , β )

α α
E [ m (Θ ) ] = ,Var ( m ( Θ ) )= 2
β β

α
s ( θ )=Var ( X i∨( Θ=θ ) )=θ , E [ s ( Θ ) ]=E [ Θ ] =
2 2
β
n n n
Z= = =
E [ s (Θ)] n+ β
2 2
α β
n+ n+ ×
Var ( m ( Θ ) ) β α

∴^
E [ m ( Θ ) ∨X ] =
n
n+ β
X+
β α
n+ β β ()
4
For the Normal | Normal model, X i ∨( Θ=θ ) N ( θ , σ 2 ), independently for i=1 , 2 ,… , n , and
Θ N ( η , τ 2 ).
m ( θ )=E [ X i∨( Θ=θ ) ]=θ ⇒m ( Θ )=Θ Normal ( η , τ 2 )

E [ m (Θ ) ] =η , Var ( m ( Θ ) )=τ
2

s2 ( θ )=Var ( X i∨( Θ=θ ) )=σ 2 , E [ s2 ( Θ ) ]=σ 2

n n n τ2
Z= = =
E [ s2 ( Θ ) ] σ 2 n τ 2 +σ 2
n+ n+ 2
Var ( m ( Θ ) ) τ
2 2
nτ σ
E [ m ( Θ ) ∨X ] =
∴^ 2 2
X + 2 2 (η)
n τ +σ n τ +σ

5. Suppose X , Y and Z are discrete random variables. Prove that E [ X∨Y ] =E [ E [ X∨ ( Y , Z ) ] ∨Y ]

. [3]

Solution 5

E [ E [ X ∨( Y = y , Z ) ]∨ ( Y = y ) ] =∑ E [ X∨ (Y = y , Z=z ) ] P ( Z=z∨Y = y )=∑

z z
{∑ xP ( X=x ∨Y = y , Z=z ) } P ( Z =
x

6. Using the notation defined in this chapter, show that:

(a) E [ X ] =E [ m ( Θ ) ] (1)
[
(b) E [ X m (Θ ) ] =E ( m ( Θ ) )
2
] (1)
1
E [s (Θ)]
2
(c) Var ( m (Θ )) =Var ( X )− (2)
n

7. (September 2004 – Q6) An insurance company has to estimate the risk premium for the coming
year for a certain risk.
(a) Describe how the credibility approach to calculating the risk premium differs from the
conventional (frequentist) approach. (1)
(b) State an advantage and disadvantage of using the pure Bayesian approach versus the
empirical Bayes credibility theory (EBCT) approach. (2)
(c) State the differences between the assumptions in EBCT Model 1 and EBCT Model 2, and
state why the latter is more likely to be useful in practice. (3)
[Total 6]

Solution 7
a) The conventional approach only uses data from the risk itself. The credibility approach combines
this with information from other sources using a credibility premium formula.

b) Pure Bayesian credibility: An advantage of the pure Bayesian approach is that it is not an
approximation, but the disadvantage is that we need the exact distributions of X ∨Θ and Θ .

5
EBCT: An advantage of the EBCT approach is that it can be used even when the exact distributions
are not known. A disadvantage of the EBCT approach is that it may not take account of the tail of the
distributions or the possibility of extreme events.

c) Model 1: The X ij ∨Θ i are independent and identically distributed. The joint distributions ( Θi , X ij )
and ( Θk , X km ) for i≠ k are independent and identically distributed.
Model 2: The Y ij ∨Θi are independent, but not necessarily identically distributed. The joint
distributions ( Θi ,Y ij ) and ( Θk , Y km ) for i≠ k are independent. The Θ 1 ,Θ 2 , … ,Θ N are independent
and identically distributed.
The difference is that the Y ij ∨Θi are not identically distributed, and the EBCT Model 2 allows for
different exposure to risk.

8. Suppose that on the basis of 1 observation from X ∨( Θ=θ ) Unif ( 0 , θ ), you must use a linear
decision function, d ( x )=a+ bx , to estimate the unknown value of the parameter θ . The loss
2
function is the squared error loss, i.e. L ( d ( x ) , θ )= ( d ( x ) −θ ) , and the prior distribution for θ is
4
f ( θ )=3/θ for θ>1.
(a) Derive the Bayes risk for these linear decision functions in terms of a and b , then determine
the optimal values for these parameters by minimizing the Bayes risk. (6)
(b) Use the Bayesian approach to derive the true optimal decision function. (3)
(c) Use R to plot the decision functions derived above, as functions of x , on the same graph.
Comment on the difference between them (especially noting the potential hazard of using
the linear decision function). (3)
[Total 12]

Solution 8
a) First, one must derive an expression for the risk function in terms of a and b :

R ( d ,θ )=E [ L ( d ( X ) , Θ )∨( Θ=θ ) ]=E [ ( a+bX−θ )2∨( Θ=θ ) ]=a 2+ b2 E [ X 2∨( Θ=θ ) ]+ θ2 +2 abE [ X∨ (Θ=θ ) ] −2aθ
2
θ θ
X ∨( Θ=θ ) Unif ( 0 , θ ) ⇒ E [ X∨( Θ=θ ) ] = , E [ X ∨( Θ=θ ) ]=
2
2 3
2
2 b 2 2 2
∴ R ( d ,θ )=a + θ +θ + abθ−2 aθ−b θ
3
The Bayes risk is
2
b
E [ R ( d ,Θ ) ] =a + E [ Θ ] + E [ Θ ] + abE [ Θ ] −2 aE [ Θ ] −b E [ Θ ]
2 2 2 2
3

[ ] [ ]
∞ ∞ ∞ ∞
3 −3θ−2 3 3 −3θ−1
E [ Θ ] =∫ θ × 4 dθ= = , E [ Θ ]=∫ θ × 4 dθ=
2 2
=3
1 θ 2 1 2 1 θ 1 1

3
∴ E [ R ( d ,Θ ) ] =a +b +3+ ab−3 a−3 b
2 2
2

6
 ∂ E [R ( d , Θ)] 3 ∂ E [ R ( d ,Θ ) ] 3
=2 a+ b−3 , =2 b+ a−3
∂a 2 ∂a 2
¿ ¿
Setting the partial derivatives equal to zero minimizes the Bayes risk, and yields: α =β =6 /7 ; hence
the optimal linear decision function is d ( x )=6 ( 1+ x ) /7.

9. (September 2002 – Q9(iii)) An insurance company has insured a fleet of cars for the last four
years. Using the data from 10 similar fleets over the last four years, E [ m (Θ ) ], E [ s2 ( Θ ) ] and
Var ( m (Θ )) are estimate to be 62.8 , 106.32, and 5.8 respectively.
(a) Calculate next year’s credibility premium, using the EBCT Model 2, for a fleet of cars with
claims over the last four years given below, if the fleet will be 16 cars next year. (3)

Year 1 Year 2 Year 3 Year 4

Total amount claimed 1 000 1 200 1 500 1 400
Number of cars 15 16 18 15
(b) Explain how and why the credibility factor would be affected if the estimate of Var ( m (Θ ))
increases, and comment on the effect on the credibility premium. (2)
[Total 5]

PS: Part (ii) of the original question derives the credibility factor for the EBCT Model 2 setting. You
are encouraged to have a look at the solution, but it will not be examined in this course.

10. Assuming the assumptions underlying EBCT Model 2 hold and that the required data are
available, update the R-function below to compute the credibility premium (per unit of
exposure) for N risks based on n years of experience. [10]

11. An actuarial student is using EBCT Model 2 to calculate credibility premiums for a group of
insurers. He has analysed the experience for six different insurers, using 10 years of past data
from each insurer. He has obtained the following figures:
6 10

∑ ∑ P ij=1498 , P¿ =18.24 , ^E [ m (Θ ) ]=4

i=1 j=1

E [ s 2 (Θ ) ]=62.8
^ ( m (Θ )) =42.1 , ^
Var

He has just received the following information relating to a 7 th insurer, and he wishes to update
his estimates using the past ten years of claims data for this insurer given in the table below.

Year 1 2 3 4 5
Aggregate claims 100 85 90 102 109
Volume 22 24 26 20 25

Year 6 7 8 9 10
Aggregate claims 106 128 132 150 131

7
 Volume 30 29 35 40 36

Calculate his updated estimates for E [ m (Θ ) ], E [ s2 ( Θ ) ] and Var ( m (Θ )) , and hence find the
credibility premium for the new insurer for the coming year, given that this insurer is expected
to have a volume figure of 38 for the coming year. [20]

Solution 11
6 10

∑ ∑ Pij X ij 6 10
i=1 j=1
6 10
=4 ⇒ ∑ ∑ Pij X ij =4 × 1498=5992
∑ ∑ Pij i=1 j=1

i=1 j=1

10 10 10

∑ Y 7 j=∑ P7 j X 7 j=1133 , ∑ P 7 j=287

j=1 j=1 j=1

So our new estimate of E [ m (Θ ) ] is

7 10

∑ ∑ P ij X ij 5992+1133
^
E [ m (Θ ) ] = i =17 j=110 = =3.9916
1498+287
∑ ∑ Pij
i=1 j=1

Now consider E [ s2 ( Θ ) ]. The formula for the estimator of E [ s2 ( Θ ) ] is:

{ }
N n
1 1
E [ s2 ( Θ ) ]=
^ ∑ ∑ P ( X −X i) 2
N i=1 n−1 j=1 ij ij

Since N=6 (initially) and n=10, we know that:

6 10

∑ ∑ P ij ( X ij− X i )2=6 × 9× 62.8=3 391.2

i=1 j=1

( ( ) ( )
10 10 10 10 10 10 2 2
100 85
∑ P7 j ( X 7 j −X 7 ) =∑ P 7 j X 7 j −2 X 7 ∑ P 7 j X 7 j + X 7 ∑ P7 j=∑ P7 j X 7 j− X 7 ∑ P7 j= 22
2 2 2 2 2
22
+24
24
+…
j=1 j=1 j=1 j=1 j=1 j=1

So our new estimate of E [ s2 ( Θ ) ] is

1 1
E [ s ( Θ ) ]= × × ( 3391.2+66.3035 ) =54.881
^ 2
7 9

We now want the estimate for Var ( m (Θ )) . The formula for the estimator of Var ( m (Θ )) is:

( )
N n
1 1 ^ [ s2 ( Θ ) ]
^
Var ( m (Θ )) = ¿ ∑ ∑
P Nn−1 i=1 j=1
Pij ( X ij −X )2− E

Substituting in the relevant information for the first 6 risks, we get:

8
 ( )
6 10
1 1
42.1= ∑ ∑
18.24 59 i=1 j=1
Pij ( X ij −X )2 −62.8

6 10 6 10 6 10
∴ ∑ ∑ Pij ( X ij −X ) =∑ ∑ Pij X 2ij− X 2 ∑ ∑ P ij =59 ( 42.1× 18.24+62.8 )=49 011.536
2

i=1 j=1 i=1 j=1 i =1 j=1

6 10

∑ ∑ P ij X 2ij=49 011.536+ 42 ×1 498=72 979.536

i=1 j=1

7 10
∴ ∑ ∑ Pij X 2ij =72979.536 +4 539.0874=77 518.6234
i=1 j=1

7 10

∑ ∑ P ij ( X ij− X )2=77 518.6234−3.99162 × ( 1 498+ 287 )=49 078.449

i=1 j=1

¿ ¿
We also need the updated value of P . Using the formula of P , we know that:

( ) (∑ P2i
)
N N N
1 P 1
P= ¿
∑
Nn−1 i=1
Pi 1− i =
P Nn−1
Pi − ∑
P
i=1 i=1

( )
6

1
∑ P2i 6
⇒ ∑ Pi =1 498 ( 1498−59× 18.24 )=631 916.32
i=1 2
18.24= 1 498−
59 1 498 i=1

7
∴ ∑ P2i =631 916.32+287 2=714 285.32
i=1

( ) (
2

)
7 7
1 Pi 1 714 285.32
P=
¿
69
∑ P i−∑ P
=
69
( 1 498+287 )−
1498+287
=20.07015
i=1 i=1

So our new estimate of Var ( m (Θ )) is:

^
Var ( m (Θ )) =
1 1
20.07015 69 (
( 49 078.449 ) −54.881 =32.70533 )
Therefore the updated parameter estimates are:

^ E [ s 2 ( Θ ) ] =54.881, Var
E [ m (Θ ) ] =3.9916 , ^ ^ ( m (Θ ) )=32.7053

The credibility factor for the 7 th insurer is

287
Z7 = =0.994 187
54.881
287 +
32.7053

So the credibility premium per unit volume is:

9
 1 133
CP=0.994187 × + ( 1−0.994187 ) × 3.9916=3.948
287

And, the credibility premium for the coming year is 3.948 ×38=150.02

10
R-programming Assignment 2

Write a function in R that accepts a matrix X (where the rows of X represent individual risks and the
columns of X represent different years) that calculates the credibility premium for each of the risks.

Even though your function should work on matrices of any dimensions, test your function on the
following data and supply the output/answers

dat <- matrix(c(85,62,122,72,170,73,107,144,78,81,91,

104,99,90,153,88,71,80,88,140,70,68,150,98,175), nrow=5, ncol=5)

Note: you should not “hard code” the results into your function. The function should be a maximum
of 20 lines (excluding comments) with only single instructions on each line. Hint: use the apply
function.

11
R-programming Assignment 2

Perform the following simulation study:

Use the Normal|Normal Bayesian model to generate data for N risks and n years each. Using a few
values of N and n investigate the accuracy of the Empirical Bayes approach and comment on its
validity.

12