Deriving the Bühlmann-Straub Estimator

Why the Bühlmann-Straub Estimate Is a Blend

A class with more exposure usually provides a more stable estimate of its underlying expected loss cost than a class with little exposure. But even a large class has random variation, so its observed experience need not equal its true expected cost. Bühlmann-Straub credibility addresses both facts: it gives more weight to the class’s own experience as its total exposure increases, while retaining the overall mean as a complement.

The focus here is the class estimate and why it has this blended form. The derivation uses the model’s variance components and the total exposure in the class. It is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Definition: In the Bühlmann-Straub model, $X_{ij}$ is the observed loss cost per unit exposure for class $i$ in period $j$, and $w_{ij}>0$ is that observation’s exposure weight. Conditional on the class risk parameter $\theta_i$, observations are independent, with $$ E[X_{ij}\mid\theta_i]=m(\theta_i), \qquad \operatorname{Var}(X_{ij}\mid\theta_i)=\frac{s^2(\theta_i)}{w_{ij}}. $$ The class’s expected loss cost is $m(\theta_i)$. Define the overall mean $\mu=E[m(\theta_i)]$, the expected process variance $v=E[s^2(\theta_i)]$, and the variance of hypothetical means $a=\operatorname{Var}(m(\theta_i))$. Assume these quantities are finite and $a>0$.

The conditional variance decreases as the exposure weight increases: doubling $w_{ij}$ halves the conditional variance under this model. The model also assumes the class parameters are drawn from a common population, so $\mu$, $v$, and $a$ are shared structural quantities. The parameter $\theta_i$ represents the class’s persistent risk characteristics; the observations vary around that class’s expected cost.

The Class Mean and the Credibility Factor

For class $i$, let total exposure be $W_i=\sum_j w_{ij}$. Its exposure-weighted mean is

\bar X_i=\frac{\sum_j w_{ij}X_{ij}}{W_i}.

This is the class’s observed cost per unit exposure. It is not generally the ordinary arithmetic average of the period observations: a period with more exposure contributes more to the result. The Bühlmann-Straub credibility factor is

Z_i=\frac{W_i}{W_i+K}, \qquad K=\frac{v}{a}.

Here $K$ is the complement of credibility parameter: it expresses the amount of exposure, on the model’s scale, needed for the class’s own experience to become influential. The class estimate is

\widehat{m}_i=Z_i\bar X_i+(1-Z_i)\mu.

This form uses the class’s observed mean and the overall mean. If $W_i$ is small relative to $K$, then $Z_i$ is small and the estimate is drawn toward $\mu$. As $W_i$ increases, $Z_i$ increases toward 1, so the estimate approaches the class’s own experience.

Worked Example: Deriving a Class Estimate From Unequal Period Exposures

A hypothetical coverage class has three annual loss-cost observations. The exposure weights are 1, 2, and 3, and the corresponding loss costs are $80, $100, and $120 per exposure unit. The overall mean is $\mu=\$95$, the expected process variance is $v=900$, and the variance of hypothetical means is $a=300$. Find the Bühlmann-Straub estimate. Use dollars per exposure unit and round the final result to the nearest cent.

Method. Calculate the exposure-weighted mean, then calculate $K$ and $Z_i$, and finally blend the class mean with the overall mean. This method avoids treating observations with unequal exposure as equally informative.

Step 1: Total exposure. The total exposure is

$$ W_i=1+2+3=6. $$

Step 2: Exposure-weighted class mean. The weighted loss-cost total is

$$ 1(80)+2(100)+3(120)=80+200+360=640. $$

Therefore,

\bar X_i=\frac{640}{6}=106.6667.

Step 3: Credibility. The complement of credibility parameter and the factor are

K=\frac{v}{a}=\frac{900}{300}=3, \qquad Z_i=\frac{6}{6+3}=\frac{2}{3}.

Step 4: Blend. Substituting the values,

\widehat{m}_i=\frac{2}{3}(106.6667)+\frac{1}{3}(95) =71.1111+31.6667=102.7778.

The estimate is $102.78 per exposure unit, rounded to two decimal places. It lies between $95 and $106.67, as a credibility-weighted average must. The weighted mean used above is $640/6$, not the unweighted average $(80+100+120)/3=\$100$.

Why the Blend Formula Holds

The estimate can be derived as the best linear predictor of the class’s expected cost $m(\theta_i)$ from its observed weighted mean $\bar X_i$. “Best linear” means choosing constants $A$ and $B$ to minimize the expected squared prediction error among predictors of the form $A+B\bar X_i$. This derivation does not require the observations to be normally distributed.

Conditional independence and the variance assumption give the conditional mean and variance of the weighted mean:

E[\bar X_i\mid\theta_i]=m(\theta_i), \qquad \operatorname{Var}(\bar X_i\mid\theta_i)=\frac{s^2(\theta_i)}{W_i}.

For the variance result, the conditional variance of the weighted sum is $\sum_j w_{ij}^2s^2(\theta_i)/w_{ij}=W_i s^2(\theta_i)$. Dividing that sum by $W_i^2$ gives $s^2(\theta_i)/W_i$. Taking expectations over the class parameter gives $E[\operatorname{Var}(\bar X_i\mid\theta_i)]=v/W_i$.

The law of total covariance and the law of total variance now give

\operatorname{Cov}(m(\theta_i),\bar X_i)=a, \qquad \operatorname{Var}(\bar X_i)=a+\frac{v}{W_i}.

For the covariance, the conditional mean of $\bar X_i$ is $m(\theta_i)$, so the covariance reduces to $\operatorname{Var}(m(\theta_i))=a$. For the variance, the total-variance identity adds the expected conditional variance $v/W_i$ to the variance of the conditional mean $a$.

For a best linear predictor $A+B\bar X_i$, minimizing expected squared error yields the usual linear projection coefficients: $B=\operatorname{Cov}(m(\theta_i),\bar X_i)/\operatorname{Var}(\bar X_i)$ and $A=E[m(\theta_i)]-B E[\bar X_i]$. Substitution gives

B=\frac{a}{a+v/W_i} =\frac{W_i}{W_i+v/a} =Z_i, \qquad A=\mu-Z_i\mu=(1-Z_i)\mu.

Consequently,

A+B\bar X_i=(1-Z_i)\mu+Z_i\bar X_i.

This establishes the blended estimate. Its two coefficients sum to one, and the coefficient on class experience is exactly the exposure-dependent credibility factor. Under this model it is the best linear predictor of the class mean. It should not be described as the exact conditional expectation under arbitrary distributions; that stronger claim requires additional distributional assumptions.

Worked Example: Separating the Weighted Mean From the Credibility Calculation

A hypothetical equipment-protection class has annual loss costs of $100, $140, and $160, with respective exposure weights 1, 4, and 5. The overall mean is $130, $v=900$, and $a=300$. Calculate the class estimate to the nearest cent.

Method. First compute the weighted mean and total weight; then apply the model-based blend. This avoids confusing the weight attached to a period with the credibility assigned to the entire class.

The total exposure is $W_i=1+4+5=10$. The weighted total is

$$ 1(100)+4(140)+5(160)=100+560+800=1460. $$

Thus,

\bar X_i=\frac{1460}{10}=146.

The complement parameter and credibility factor are

K=\frac{900}{300}=3, \qquad Z_i=\frac{10}{10+3}=\frac{10}{13}.

The class mean exceeds the overall mean by $146-130=16$. Therefore,

\widehat{m}_i=130+\frac{10}{13}(16) =130+\frac{160}{13} =142.3077.

The estimate is $142.31 per exposure unit, rounded to two decimal places. As a check, it is closer to the class mean of $146 than to the overall mean of $130 because $Z_i=10/13$ is greater than $1/2$.

Interpreting the Components and Avoiding Traps

The weighted class mean summarizes observed experience; $Z_i$ describes how much of that experience is retained in the estimate; and $\mu$ supplies the complement. The factor depends on both total exposure and the relative sizes of the variance components. Greater process variance $v$ means more noise within a class and reduces credibility. Greater variance $a$ between class means means that the classes differ more, making a class’s own experience more informative about its particular mean.

The factor has the expected limiting behavior:

\lim_{W_i\to 0} Z_i=0, \qquad \lim_{W_i\to\infty} Z_i=1.

Thus, with very little exposure the estimate approaches $\mu$, while with very large exposure it approaches $\bar X_i$. These limits are useful checks, but they do not remove the need to use the correct weights and variance components.

Common exam trap: Do not use $v/a$ as the credibility factor. It is $K$, the complement of credibility parameter; the factor is $Z_i=W_i/(W_i+K)$. Also do not use an unweighted class average when the exposure weights differ.

Worked Example: Correcting an Unweighted-Mean Error

A hypothetical service-contract class has loss costs of $70 and $110 at exposure weights 2 and 3. The overall mean is $100, $v=500$, and $a=250$. A preliminary calculation uses the ordinary mean, $90, and reports an estimate of $92.86. Identify the error and calculate the correct estimate to the nearest cent.

Method. Recalculate the class mean with the stated exposure weights before applying credibility. The common mistake is to average the two period costs equally even though their exposure differs.

The total exposure and weighted loss-cost total are

W_i=2+3=5, \qquad 2(70)+3(110)=140+330=470.

The correct exposure-weighted mean is $\bar X_i=470/5=94$, not $(70+110)/2=90$. The credibility quantities are

K=\frac{500}{250}=2, \qquad Z_i=\frac{5}{5+2}=\frac{5}{7}.

The corrected estimate is

\widehat{m}_i=\frac{5}{7}(94)+\frac{2}{7}(100) =\frac{470+200}{7} =\frac{670}{7} =95.7143.

The correct result is $95.71 per exposure unit, rounded to two decimal places. It lies between the overall mean of $100 and the correctly calculated class mean of $94. The preliminary $92.86 used the wrong class mean, even though its credibility factor was correct.

When to Use the Result

The Bühlmann-Straub class estimate is appropriate when observations have unequal exposure and the model’s conditional mean and variance structure is a reasonable representation of the data. The estimate predicts the class’s expected cost, not the next observation itself. A future observation has additional process variation around that expected cost, even after the class mean has been estimated.

Before using the blend, check that the $X_{ij}$ values are expressed on a comparable per-unit-exposure basis, the weights correspond to the exposure attached to each observation, and $v$ and $a$ refer to the same modeled quantity. If all observations have equal weight, the weighted mean reduces to the ordinary mean, but the factor still depends on the class’s total weight.

Check Your Understanding

For each problem, calculate the requested class estimate using the Bühlmann-Straub blend.

A hypothetical class has loss costs $90 and $130 with weights 1 and 3. Given $\mu=100$, $v=400$, and $a=200$, find $\bar X_i$, $Z_i$, and the estimate. (Answer: $\bar X_i=\$120$, $Z_i=2/3$, estimate $=\$113.33$ per exposure unit.)
A hypothetical fleet class has observations $40, $70, and $100 with weights 2, 2, and 6. Given $\mu=80$, $v=600$, and $a=300$, calculate the estimate. (Answer: $\bar X_i=\$82$, $Z_i=10/12=5/6$, estimate $=\$81.67$ per exposure unit.)
A hypothetical property class has two loss costs, $150 at weight 4 and $90 at weight 2. Given $\mu=120$, $v=800$, and $a=400$, calculate $K$, $Z_i$, and the estimate. (Answer: $K=2$, $Z_i=6/8=0.75$, $\bar X_i=\$130$, estimate $=\$127.50$ per exposure unit.)
A class has total exposure $W_i=8$, weighted total loss cost of $960, overall mean $\mu=\$110$, $v=1{,}200$, and $a=300$. Calculate the estimate and state which mean receives more weight. (Answer: $\bar X_i=\$120$, $K=4$, $Z_i=2/3$, estimate $=\$116.67$ per exposure unit; the class mean receives more weight.)
For a hypothetical class, $W_i=3$, $\bar X_i=\$70$, $\mu=\$100$, $v=900$, and $a=300$. A calculation uses $Z_i=v/a=3$. Identify the error and give the correct estimate. (Answer: $v/a=3$ is $K$, not $Z_i$; $Z_i=3/(3+3)=0.5$, so the estimate is $\$85.00$ per exposure unit.)

What You'll Learn