Nonlife Actuarial Models

Chapter 7
Bühlmann Credibility
 Learning Objectives

1. Basic framework of Bühlmann credibility

2. Variance decomposition

3. Expected value of the process variance

4. Variance of the hypothetical mean

5. Bühlmann credibility

6. Bühlmann-Straub credibility

2
 7.1 Framework and Notations

• Consider a risk group or block of insurance policies with loss mea-

sure denoted by X, which may be claim frequency, claim severity,
aggregate loss or pure premium.

• Assume that the risk profiles of the group are characterized by a

parameter θ, which determines the distribution of the loss measure
X.

• Denote the conditional mean and variance of X given θ by

E(X | θ) = μX (θ), (7.1)

and
2
Var(X | θ) = σX (θ). (7.2)

3
 • Assume that the insurance company has similar blocks of policies
with different risk profiles.

• Thus, the parameter θ varies with different risk groups.

• We treat θ as the realization of a random variable Θ, the distribution

of which is called the prior distribution.

• When θ varies over the support of Θ, the conditional mean and

variance of X become random variables in Θ, and are denoted by
2
μX (Θ) = E(X | Θ) and σX (Θ) = Var(X | Θ), respectively.

Example 7.1: An insurance company has blocks of workers compensa-

tion policies. The claim frequency is known to be Poisson with parameter
λ, where λ is 20 for low-risk group and 50 for high-risk group. Suppose

4
30% of the risk groups are low risk and 70% are high risk. What are the
conditional mean and variance of the claim frequency?

Solution: The parameter determining the claim frequency X is λ, which

we assume is a realization of the random variable Λ. As X is Poisson, the
conditional mean and conditional variance of X are equal to λ. Thus, we
have the following results

λ Pr(Λ = λ) E(X | λ) Var(X | λ)

20 0.3 20 20
50 0.7 50 50

so that
⎧
⎪
⎨ 20, with probability 0.30,
μX (Λ) = E(X | Λ) = ⎪
⎩ 50, with probability 0.70.

5
Likewise, we have
⎧
⎪
⎨ 20, with probability 0.30,
2
σX (Λ) = Var(X | Λ) = ⎪
⎩ 50, with probability 0.70.

Example 7.2: The claim severity X of a block of health insurance

policies is normally distributed with mean θ and variance 10. If θ takes
values within the interval [100, 200] and follows a uniform distribution,
what are the conditional mean and conditional variance of X?

Solution: The conditional variance of X is 10, irrespective of θ. Hence,

2
we have σX (Θ) = Var(X | Θ) = 10 with probability 1. The conditional
mean of X is Θ, i.e., μX (Θ) = E(X | Θ) = Θ, which is uniformly distrib-

6
uted in [100, 200] with pdf
⎧
⎪
⎨ 0.01, for θ ∈ [100, 200],
fΘ (θ) = ⎪
⎩ 0, otherwise.
• The Bühlmann model assumes that there are n observations of losses,
denoted by {X1 , · · · , Xn }.

• The observations may be losses recorded in n periods and they are

assumed to be iid as X, which depends on the parameter θ.

• The task is to update the prediction of X for the next period, i.e.,
Xn+1 , based on {X1 , · · · , Xn }.

• In the Bühlmann approach the solution depends on the variation be-

tween the conditional means as well as the average of the conditional
variances of the risk groups.

7
 7.2 Variance Components
• The variation of the loss measure X consists of two components: the
variation between risk groups and the variation within risk
groups.

• The first component, variation between risk groups, is due to the

randomness of the risk profiles of each group and is captured by the
parameter Θ.

• The second component, variation within risk group, is measured by

the conditional variance of the risk group.

• We first consider the calculation of the overall mean of the loss mea-
sure X. The unconditional mean (or overall mean) of X is
E(X) = E[E(X | Θ)] = E[μX (Θ)]. (7.3)

8
• For the unconditional variance (or total variance), we have

Var(X) = E[Var(X | Θ)] + Var[E(X | Θ)]. (7.4)

• Var(X | Θ) measures the variance of a given risk group. It is a func-

tion of the random variable Θ and we call this the process variance.

• Thus, E[Var(X | Θ)] is the expected value of the process vari-

ance (EPV).

• On the other hand, E(X | Θ) is the mean of a given risk group. We

call this conditional mean the hypothetical mean.

• Thus, Var[E(X | Θ)] is the variance of the hypothetical means

(VHM), as it measures the variations in the means of the risk
groups.

9
• Verbally, equation (7.4) can be written as

Total variance = Expected value of process variance

+ Variance of hypothetical means, (7.5)

or
Total variance = EPV + VHM. (7.6)

• It can also be stated alternatively as

Total variance = Mean of conditional variance

+Variance of conditional mean. (7.7)

• Symbolically, we use the following notations

2
E[Var(X | Θ)] = E[σX (Θ)] = μPV , (7.8)

10
 and
2
Var[E(X | Θ)] = Var[μX (Θ)] = σHM , (7.9)
so that equation (7.4) can be written as

2
Var(X) = μPV + σHM . (7.10)

Example 7.3: For Examples 7.1 and 7.2, calculate the unconditional
mean, the expected value of the process variance, the variance of the
hypothetical means and the total variance.

Solution: For Example 7.1, the unconditional mean is

E(X) = Pr(Λ = 20)E(X | Λ = 20) + Pr(Λ = 50)E(X | Λ = 50)

= (0.3)(20) + (0.7)(50)
= 41.

11
The expected value of the process variance, EPV, is
E[Var(X | Λ)] = Pr(Λ = 20)Var(X | Λ = 20) + Pr(Λ = 50)Var(X | Λ = 50)
= (0.3)(20) + (0.7)(50)
= 41.
As the mean of the hypothetical means (i.e., the unconditional mean) is
41, the variance of the hypothetical means, VHM, is
Var[E(X | Λ)] = (0.3)(20 − 41)2 + (0.7)(50 − 41)2 = 189.
Thus, the total variance of X is
Var(X) = E[Var(X | Λ)] + Var[E(X | Λ)] = 41 + 189 = 230.
For Example 7.2, as Θ is uniformly distributed in [100, 200], the uncondi-
tional mean of X is
E(X) = E[E(X | Θ)] = E(Θ) = 150.

12
As X has a constant variance of 10, the expected value of the process
variance is
E[Var(X | Θ)] = E(10) = 10.
The variance of the hypothetical means is

(200 − 100)2
Var[E(X | Θ)] = Var(Θ) = = 833.33,
12
and the total variance of X is

Var(X) = 10 + 833.33 = 843.33.

2

Example 7.5: An insurance company sells workers compensation poli-

cies, each of which belongs to one of three possible risk groups. The risk
groups have claim frequencies N that are Poisson distributed with para-
meter λ and claim severity X that are gamma distributed with parameters

13
α and β. Claim frequency and claim severity are independently distrib-
uted given a risk group, and the aggregate loss is S. The data of the risk
groups are given in Table 7.2.
Table 7.2: Data for Example 7.5

Relative Distribution of N: Distribution of X:

Risk group frequency PN (λ) G(α, β)
1 0.2 λ = 20 α = 5, β = 2
2 0.4 λ = 30 α = 4, β = 3
3 0.4 λ = 40 α = 3, β = 2
For each of the following loss measures: (a) claim frequency N, (b) claim
severity X, and (c) aggregate loss S, calculate EPV, VHM and the total
variance.

Solution: (a) Claim frequency We first calculate the conditional

mean and conditional variance of N given the risk group, which is charac-

14
terized by the parameter Λ. As N is Poisson, the mean and variance are
equal to Λ, so that we have the results in Table 7.3.

Table 7.3: Results for Example 7.5 (a)

2
Risk group Probability E(N | Λ) = μN (Λ) Var(N | Λ) = σN (Λ)
1 0.2 20 20
2 0.4 30 30
3 0.4 40 40

Thus, the EPV is

μPV = E[Var(N | Λ)] = (0.2)(20) + (0.4)(30) + (0.4)(40) = 32,

which is also equal to the unconditional mean E[μN (Λ)]. For VHM, we
first calculate

E{[μN (Λ)]2 } = (0.2)(20)2 + (0.4)(30)2 + (0.4)(40)2 = 1,080,

15
so that

2
σHM = Var[μN (Λ)] = E{[μN (Λ)]2 } − {E[μN (Λ)]}2 = 1,080 − (32)2 = 56.

Therefore, the total variance of N is

2
Var(N) = μPV + σHM = 32 + 56 = 88.

(b) Claim severity There are three claim-severity distributions, which

are specific to each risk group. Note that the relative frequencies of the
risk groups as well as the claim frequencies in the risk groups jointly
determine the relative occurrence of each claim-severity distribution. The
probabilities of occurrence of the severity distributions, as well as their
conditional means and variances are given in Table 7.4, in which Γ denotes
the vector random variable representing α and β.

16
 Table 7.4: Results for Example 7.5 (b)

Group Col 2 Probability of E(X | Γ) Var(X | Γ)

2
Group probability λ × Col 3 severity X = μX (Γ) = σX (Γ)
1 0.2 20 4 0.125 10 20
2 0.4 30 12 0.375 12 36
3 0.4 40 16 0.500 6 12

Column 4 gives the expected number of claims in each group weighted

by the group probability. Column 5 gives the probability of occurrence of
each type of claim-severity distribution, which is obtained by dividing the
corresponding figure in Column 4 by the sum of Column 4 (e.g., 0.125 =
4/(4 + 12 + 16)). The last two columns give the conditional mean αβ and
conditional variance αβ 2 corresponding to the three different distributions
of claim severity. Similar to the calculation in (a), we have

17
 E(X) = E[E(X | Γ)] = (0.125)(10) + (0.375)(12) + (0.5)(6) = 8.75,

and
μPV = (0.125)(20) + (0.375)(36) + (0.5)(12) = 22.
To calculate VHM, we first compute the raw second moment of the con-
ditional mean of X, which is

E{[μX (Γ)]2 } = (0.125)(10)2 + (0.375)(12)2 + (0.5)(6)2 = 84.50.

Hence,
2
σHM = Var[μX (Γ)] = E{[μX (Γ)]2 }−{E[μX (Γ)]}2 = 84.50−(8.75)2 = 7.9375.

Therefore, the total variance of X is

2
Var(X) = μPV + σHM = 22 + 7.9375 = 29.9375.

18
(c) Aggregate loss The distribution of the aggregate loss S is deter-
mined jointly by Λ and Γ, which we shall denote as Θ. For the conditional
mean of S, we have

E(S | Θ) = E(N | Θ)E(X | Θ) = λαβ.

For the conditional variance of S, we use the result on compound distrib-

ution with Poisson claim frequency stated in equation (A.123), and make
use of the assumption of gamma severity to obtain

2
Var(S | Θ) = λ[σX (Γ) + μ2X (Γ)] = λ(αβ 2 + α2 β 2 ).

The conditional means and conditional variances of S are summarized in

Table 7.5.

19
 Table 7.5: Results for Example 7.5 (c)

Group Parameters E(S | Θ) Var(S | Θ)

Group probability λ, α, β = μS (Θ) = σS2 (Θ)
1 0.2 20, 5, 2 200 2,400
2 0.4 30, 4, 3 360 5,400
3 0.4 40, 3, 2 240 1,920
The unconditional mean of S is

E(S) = E[E(S | Θ)] = (0.2)(200) + (0.4)(360) + (0.4)(240) = 280,

and the EPV is

μPV = (0.2)(2,400) + (0.4)(5,400) + (0.4)(1,920) = 3,408.

Also, the VHM is given by

2
σHM = Var[μS (Θ)]

20
 = E{[μS (Θ)]2 } − {E[μS (Θ)]}2
h i
2 2 2
= (0.2)(200) + (0.4)(360) + (0.4)(240) − (280)2
= 4,480.

Therefore, the total variance of S is

Var(S) = 3,408 + 4,480 = 7,888.

2

• EPV and VHM measure two different aspects of the total variance.

• When a risk group is homogeneous so that the loss claims are similar
within the group, the conditional variance is small. If all risk groups
have similar loss claims within the group, the expected value of the
process variance is small.

21
• On the other hand, if the risk groups have very different risk profiles
across groups, their hypothetical means will differ more and thus the
variance of the hypothetical means will be large.

• Thus, it will be easier to distinguish between risk groups if the hy-

pothetical means differ more and the average of the process variance
is small.

• We define k as the ratio of EPV to VHM, i.e.,

μPV EPV
k= 2
= . (7.11)
σHM VHM

• A small EPV or large VHM will gives rise to a small k. The risk
groups will be more distinguishable in the mean when k is smaller,
in which case we may put more weight on the data in updating our
revised prediction for future losses.

22
Example 7.6: Frequency of claim per year, N, is distributed as a Bi-
nomial random variable BN (10, θ), and claim severity, X, is distributed
as an exponential random variable with mean cθ, where c is a known
constant. Given θ, claim frequency and claim severity are independently
distributed. Derive an expression of k for the aggregate loss per year, S,
in terms of c and the moments of Θ, and show that it does not depend on
c. If Θ is 0.3 or 0.7 with equal probabilities, calculate k.

Solution: We first calculate the conditional mean of S as a function of θ.

Due to the independence assumption of N and X, the hypothetical mean
of S is

E(S | Θ) = E(N | Θ)E(X | Θ) = (10Θ)(cΘ) = 10cΘ2 .

The process variance is

2 2
Var(S | Θ) = μN (Θ)σX (Θ) + σN (Θ)μ2X (Θ)

23
 = (10Θ)(cΘ)2 + [10Θ(1 − Θ)](cΘ)2
= 10c2 Θ3 + 10c2 Θ3 (1 − Θ)
= 10c2 Θ3 (2 − Θ).
Hence, the unconditional mean of S is
E(S) = E[E(S | Θ)] = E(10cΘ2 ) = 10cE(Θ2 )
and the variance of the hypothetical means is
2
σHM = Var[E(S | Θ)]
= Var(10cΘ2 )
= 100c2 Var(Θ2 )
= 100c2 {E(Θ4 ) − [E(Θ2 )]2 }.
The expected value of the process variance is
μPV = E[Var(S | Θ)]

24
 = E[10c2 Θ3 (2 − Θ)]
= 10c2 [2E(Θ3 ) − E(Θ4 )].

Combining the above results we conclude that

μPV 10c2 [2E(Θ3 ) − E(Θ4 )] 2E(Θ3 ) − E(Θ4 )
k= 2 = 2 4 2 2
= 4 2 2
.
σHM 100c {E(Θ ) − [E(Θ )] } 10{E(Θ ) − [E(Θ )] }
Thus, k does not depend on c. To compute its value for the given distri-
bution of Θ, we present the calculations as follows:

θ Pr(Θ = θ) θ2 θ3 θ4
0.3 0.5 0.09 0.027 0.0081
0.7 0.5 0.49 0.343 0.2401

Thus, the required moments of Θ are

E(Θ) = (0.5)(0.3) + (0.5)(0.7) = 0.5,

25
 E(Θ2 ) = (0.5)(0.09) + (0.5)(0.49) = 0.29,

E(Θ3 ) = (0.5)(0.027) + (0.5)(0.343) = 0.185

and

E(Θ4 ) = (0.5)(0.0081) + (0.5)(0.2401) = 0.1241,

so that
2(0.185) − 0.1241
k= = 0.6148.
10 [0.1241 − (0.29)2 ] 2

In this example, note that both EPV and VHM depend on c. However,
as the effects of c on these components are the same, the ratio of EPV to

26
VHM is invariant to c. Also, though X and N are independent given θ,
they are correlated unconditionally due to their common dependence on
Θ.

27
 7.3 Bühlmann Credibility

• Bühlmann’s approach of updating the predicted loss measure is

based on a linear predictor using past observations.

• It is also called the greatest accuracy approach or the least

squares approach.

• For the classical credibility approach, the updated prediction U is

given by (see equation (6.1))

U = ZD + (1 − Z)M. (7.12)

• The Bühlmann credibility method has a similar basic equation, in

which D is the sample mean of the data and M is the overall prior
mean E(X).

28
• The Bühlmann credibility factor Z depends on the sample size n
and the EPV to VHM ratio k. In particular, Z varies with n and k
as follows:

1. Z increases with the sample size n of the data.

2. Z increases with the distinctiveness of the risk groups. As argued

above, the risk groups are more distinguishable when k is small.
Thus, Z increases as k decreases.

• We now state formally the assumptions of the Bühlmann model

and derive the updating formula as the least mean-squared-error
(MSE) linear predictor.

1. {X1 , · · · , Xn } are loss measures that are independently and identi-

cally distributed as the random variable X. The distribution of X
depends on the parameter θ.

29
2. The parameter θ is a realization of a random variable Θ. Given θ,
the conditional mean and variance of X are

E(X | θ) = μX (θ), (7.13)

and
2
Var(X | θ) = σX (θ). (7.14)

3. The unconditional mean of X is E(X) = E[E(X | Θ)] = μX . The

mean of the conditional variance of X is

2
E[Var(X | Θ)] = E[σX (Θ)]
= μPV
= Expected value of process variance
= EPV, (7.15)

30
 and the variance of the conditional mean is

Var[E(X | Θ)] = Var[μX (Θ)]

2
= σHM
= Variance of hypothetical means
= VHM. (7.16)

The unconditional variance (or total variance) of X is

Var(X) = E[Var(X | Θ)] + Var[E(X | Θ)]

2
= μPV + σHM
= EPV + VHM. (7.17)

4. The Bühlmann approach formulates a predictor of Xn+1 based on a

linear function of {X1 , · · · , Xn }, where Xn+1 is assumed to have the

31
 same distribution as X. The predictor minimizes the mean squared
error in predicting Xn+1 over the joint distribution of Θ, Xn+1 and
{X1 , · · · , Xn }. Specifically, the predictor is given by

X̂n+1 = β0 + β1 X1 + · · · + βn Xn , (7.18)

where β0 , β1 , · · · , βn are chosen to minimize the mean squared error,

MSE, defined as
∙³ ´2 ¸
MSE = E Xn+1 − X̂n+1 . (7.19)

• We skip the proof and state the result here that

X̂n+1 = Z X̄ + (1 − Z)μX , (7.34)

where
n
Z= . (7.35)
n+k

32
• Z defined in equation (7.35) is called the Bühlmann credibility
factor or simply the Bühlmann credibility. It depends on the
EPV to VHM ratio k, which is called the Bühlmann credibility
parameter.

• The optimal linear forecast X̂n+1 given in equation (7.34) is also

called the Bühlmann premium.

• Note that k depends only on the parameters of the model, while Z

is a function of k and the size n of the data.

• For predicting claim frequency N, the sample size n is the number

of periods over which the number of claims is aggregated. For pre-
dicting claim severity X, the sample size n is the number of claims.
As aggregate loss S refers to the total loss payout per period, the
sample size is the number of periods of claim experience.

33
Example 7.7: Refer to Example 7.5. Suppose the claim experience last
year was 26 claims with an average claim size of 12. Calculate the updated
prediction of (a) the claim frequency, (b) the average claim size, and (c)
the aggregate loss, for next year.

Solution: (a) Claim frequency From Example 7.5, we have k =

0.5714 and M = E(N) = 32. Now we are given n = 1 and D = 26.
Hence,
1
Z= = 0.6364,
1 + 0.5714
so that the updated prediction of the claim frequency of this group is

U = (0.6364)(26) + (1 − 0.6364)(32) = 28.1816.

(b) Claim severity We have k = 2.7717 and M = E(X) = 8.75, with

34
n = 26 and D = 12. Thus,
26
Z= = 0.9037,
26 + 2.7717
so that the updated prediction of the claim severity of this group is

U = (0.9037)(12) + (1 − 0.9037)(8.75) = 11.6870.

(c) Aggregate loss With k = 0.7607, M = E(S) = 280, n = 1 and

D = (26)(12) = 312, we have
1
Z= = 0.5680,
1 + 0.7607
so that the updated prediction of the aggregate loss of this group is

U = (0.5680)(312) + (1 − 0.5680)(280) = 298.1760.

35
 7.4 Bühlmann-Straub Credibility

• An important limitation of the Bühlmann credibility theory is that

the loss observations Xi are assumed to be identically distributed.
This assumption is violated if the data are over different periods
with different exposures (the definition of exposure will be explained
below).

• The Bühlmann-Straub credibility model extends the Bühlmann

theory to cases where the loss data Xi are not identically distrib-
uted. In particular, the process variance of the loss measure is as-
sumed to depend on the exposure.

• We denote the exposure by mi , and the loss per unit of exposure by

Xi . Note that the exposure needs not be the number of insureds,

36
 although it may often be the case. We then assume the following for
the conditional variance of Xi
2
σX (Θ)
Var(Xi | Θ) = , (7.36)
mi
2
for a suitably defined σX (Θ).

• The following are some examples.

2
1. Xi is the average number of claims per insured in year i, σX (Θ) is
the variance of the claim frequency of an insured, and the exposure
mi is the number of insureds covered in year i.

2. Xi is the average aggregate loss per month of the ith block of policies,
2
σX (Θ) is the variance of the aggregate loss of the block in a month,
and the exposure mi is the number of months of insurance claims
for the ith block of policies.

37
 2
3. Xi is the average loss per unit premium in year i, σX (Θ) is the
variance of the claim amount of an insured per year divided by the
premium per insured, and the exposure mi is the amount of premi-
ums received in year i.

• The parameter θ is a realization of a random variable Θ. Given θ,

the conditional mean and variance of Xi are

E(Xi | θ) = μX (θ), (7.38)

and
2
σX (θ)
Var(Xi | θ) = , (7.39)
mi
2
for i ∈ {1, · · · , n}, where σX (θ) is suitably defined as in the examples
above.

38
• The unconditional mean of Xi is E(Xi ) = E[E(Xi | Θ)] = E[μX (Θ)] =
μX . The mean of the conditional variance of Xi is
" #
2
σX (Θ)
E[Var(Xi | Θ)] = E
mi
μPV
= , (7.40)
mi
2
for i ∈ {1, · · · , n}, where μPV = E[σX (Θ)], and the variance of its
conditional mean is

Var[E(Xi | Θ)] = Var[μX (Θ)]

2
= σHM . (7.41)

• Now we define n
X
m= mi , (7.48)
i=1

39
 n
1 X
X̄ = mi Xi (7.51)
m i=1
and
μPV
k= 2 . (7.52)
σHM
• Denoting
m
Z= . (7.54)
m+k
we have
X̂n+1 = Z X̄ + (1 − Z)μX . (7.57)

Example 7.9: The number of accident claims incurred per year for each
insured is distributed as a binomial random variable BN (2, θ), and the
claim incidences are independent across insureds. The probability θ of the
binomial has a beta distribution with parameters α = 1 and β = 10. The
data in Table 7.7 are given for a block of policies.

40
 Table 7.7: Data for Example 7.9

Year Number of insureds Number of claims

1 100 7
2 200 13
3 250 18
4 280 —

Calculate the Bühlmann-Straub credibility prediction of the number of

claims in the fourth year.

Solution: Let mi be the number of insureds in Year i, and Xi be the

number of claims per insured in Year i. Define Xij as the number of claims
for the jth insured in Year i, which is distributed as BN (2, θ). Thus, we
have mi
1 X
E(Xi | Θ) = E(Xij | Θ) = 2Θ,
mi j=1

41
and
2
σHM = Var[E(Xi | Θ)] = Var(2Θ) = 4Var(Θ).
As Θ has a beta distribution with parameters α = 1 and β = 10, we have
αβ 10
Var(Θ) = = = 0.006887.
(α + β)2 (α + β + 1) (11)2 (12)
For the conditional variance of Xi , we have
2Θ(1 − Θ)
Var(Xi | Θ) = .
mi
Thus,
μPV = 2E[Θ(1 − Θ)].
As
α
E(Θ) = = 0.0909,
α+β

42
we have

μPV = 2[E(Θ) − E(Θ2 )]

n ³ ´o
2
= 2 E(Θ) − Var(Θ) + [E(Θ)]
= 2{0.0909 − [0.006887 + (0.0909)2 ]} = 0.1515.

Thus,
μPV 0.1515
k= 2 = = 5.5.
σHM (4)(0.006887)
As m = 100 + 200 + 250 = 550, we have
550
Z= = 0.9901.
550 + 5.5
Now

μX = E[E(Xi | Θ)] = (2)(0.0909) = 0.1818

43
and
7 + 13 + 18
X̄ = = 0.0691.
550
Thus, the predicted number of claims per insured is

(0.9901)(0.0691) + (1 − 0.9901)(0.1818) = 0.0702,

and the predicted number of claims in Year 4 is

(280)(0.0702) = 19.66.
2

44