Deriving The Bühlmann-Straub Credibility Weight

Why Unequal Exposure Changes Credibility

Experience from a larger exposure base is generally more informative about a risk’s underlying expected loss than experience from a smaller base. The Bühlmann-Straub model formalizes that idea when observations cover different amounts of exposure: each period has its own process variance, and the variance is inversely proportional to that period’s exposure weight. The resulting credibility factor depends on total exposure, not merely on the number of periods observed.

The aim is to estimate a risk’s expected loss level by blending its exposure-weighted observed average with the collective mean. The derivation explains why the blend has that form, what assumptions justify it, and how to interpret important limiting cases.

Definition: Let $\Theta$ be a risk’s underlying expected loss level, with $E[\Theta]=\mu$ and $\operatorname{Var}(\Theta)=\tau^2$. In the Bühlmann-Straub model, observations $X_1,\ldots,X_n$ have positive exposure weights $w_1,\ldots,w_n$ and satisfy $$ E[X_j\mid\Theta]=\Theta,\qquad \operatorname{Var}(X_j\mid\Theta)=\frac{\sigma^2}{w_j}. $$ The observations are conditionally independent given $\Theta$, and the process variance parameter $\sigma^2$ is common across periods. Write $W=\sum_{j=1}^n w_j$ for total exposure and $\bar X_w=\sum_{j=1}^n w_jX_j/W$ for the exposure-weighted average. Assume $\tau^2>0$ and $W>0$.

The assumptions describe a stable underlying risk level across the observed periods: conditional on that level, the expected observation in every period is $\Theta$. Periods with larger $w_j$ have lower conditional variance. This model is intended for observations measured on a comparable basis; trend, coverage changes, or other systematic differences must be handled before treating the observations as exchangeable in this way.

Deriving the Credibility Factor

The Bühlmann-Straub estimate is the best linear predictor of $\Theta$ based on $\bar X_w$, where “best” means it minimizes mean squared error among linear functions of that weighted average. The key is to calculate the variance of the weighted average and its covariance with $\Theta$.

Find the conditional mean.
Since every observation has conditional mean $\Theta$, the weighted average does too: $E[\bar X_w\mid\Theta]=\sum_jw_j\Theta/W=\Theta$.

Find the conditional variance.
Conditional independence makes the variance of the weighted sum the sum of its variances. Substituting $\operatorname{Var}(X_j\mid\Theta)=\sigma^2/w_j$ gives $\operatorname{Var}(\bar X_w\mid\Theta)=\sigma^2/W$.

Find the unconditional variance and covariance.
The law of total variance gives $\operatorname{Var}(\bar X_w)=\tau^2+\sigma^2/W$. Also, $\operatorname{Cov}(\Theta,\bar X_w)=\tau^2$, because the conditional mean of $\bar X_w$ is $\Theta$.

Choose the least-squares linear predictor.
For a predictor $a+b\bar X_w$, minimizing mean squared error gives $b=\operatorname{Cov}(\Theta,\bar X_w)/\operatorname{Var}(\bar X_w)$ and $a=\mu-b\mu$.

The conditional variance in Step 2 follows directly from the model and conditional independence:

\operatorname{Var}(\bar X_w\mid\Theta) =\frac{1}{W^2}\sum_{j=1}^n w_j^2\operatorname{Var}(X_j\mid\Theta) =\frac{1}{W^2}\sum_{j=1}^n w_j^2\frac{\sigma^2}{w_j} =\frac{\sigma^2}{W}.

For the covariance, use $E[\bar X_w\mid\Theta]=\Theta$. The covariance decomposition then gives $\operatorname{Cov}(\Theta,\bar X_w)=\operatorname{Cov}(\Theta,E[\bar X_w\mid\Theta])=\operatorname{Var}(\Theta)=\tau^2$. Consequently the regression coefficient is

Z=\frac{\tau^2}{\tau^2+\sigma^2/W} =\frac{W}{W+K}, \qquad K=\frac{\sigma^2}{\tau^2}.

The estimate can therefore be written in either of two equivalent forms:

\widehat{\Theta}_{BS} =\mu+Z(\bar X_w-\mu) =Z\bar X_w+(1-Z)\mu.

This derivation establishes the best linear predictor under the stated second-moment assumptions. It does not, by itself, establish that the estimate is the full conditional expectation $E[\Theta\mid X_1,\ldots,X_n]$ for every possible distribution. With suitable joint normal assumptions, the conditional expectation is linear and agrees with this predictor; without them, the Bühlmann-Straub result remains the least-squares linear credibility estimate.

Theorem: Under the Bühlmann-Straub assumptions, the credibility factor for the exposure-weighted average is $Z=W/(W+K)$, where $K=\sigma^2/\tau^2$. Thus the estimate gives weight $Z$ to the risk’s weighted experience and weight $1-Z$ to the collective mean.

Worked Calculations

Worked Example: Unequal Exposures and a Risk Premium

A hypothetical portfolio has collective mean loss level $\mu=\$110$, process variance parameter $\sigma^2=400$, and variance of hypothetical means $\tau^2=100$. A risk’s observations are $X_1=\$120$, $X_2=\$90$, and $X_3=\$100$, with weights $w_1=1$, $w_2=2$, and $w_3=3$. Calculate its Bühlmann-Straub estimate.

Method. Calculate total exposure and the exposure-weighted average, then use the derived credibility factor. The model assumes conditional independence and period-specific conditional variance $\sigma^2/w_j$.

1. Total exposure. $W=1+2+3=6$.

2. Weighted average. The weighted loss total is $1(120)+2(90)+3(100)=120+180+300=600$. Therefore $\bar X_w=600/6=\$100$.

3. Credibility parameter and factor. $K=400/100=4$, so $Z=6/(6+4)=0.6$.

4. Estimate. $\widehat{\Theta}_{BS}=0.6(100)+0.4(110)=60+44=\$104.00$.

Check. The estimate lies between the experience average of $\$100$ and the collective mean of $\$110$, as it must because $0\leq Z\leq1$. The common mistake avoided here is giving each period equal weight when calculating the average; the correct average weights each observation by its exposure.

Worked Example: How Added Exposure Changes the Estimate

A hypothetical risk has $\mu=\$100$, $\sigma^2=256$, and $\tau^2=64$. Its first two period observations are $\$80$ and $\$120$, with respective weights 1 and 3. A new period then contributes an observation of $\$90$ with weight 4. Calculate the estimate before and after adding the new period.

Method. Recalculate both the weighted average and total exposure when the new period is added. Here $K=256/64=4$.

Before the new period. Total exposure is $W=1+3=4$. The weighted average is $[1(80)+3(120)]/4=(80+360)/4=\$110$. Thus $Z=4/(4+4)=0.5$, and the estimate is $0.5(110)+0.5(100)=\$105.00$.

After the new period. Total exposure is $W=1+3+4=8$. The weighted average is $[1(80)+3(120)+4(90)]/8=(80+360+360)/8=\$100$. Thus $Z=8/(8+4)=2/3$, and the estimate is $(2/3)(100)+(1/3)(100)=\$100.00$.

Check. The added period moves the observed average to the collective mean, so the blended estimate is also $\$100.00$, even though the credibility factor rises from $0.5$ to $2/3$. Greater credibility does not guarantee that the estimate moves in a particular direction; the new observation’s value matters as well as its weight.

Worked Example: Zero Process Variance

For a hypothetical risk, let $\mu=\$75$, $\tau^2=25$, and $\sigma^2=0$. The risk has one observation $X_1=\$83$ with weight $w_1=0.5$. Determine its credibility estimate and explain the limiting behavior.

Method. Apply the model directly, paying attention to the fact that its process variance is zero.

1. Total exposure and weighted average. $W=0.5$, and with one observation $\bar X_w=0.5(83)/0.5=\$83$.

2. Credibility parameter. $K=\sigma^2/\tau^2=0/25=0$. Hence $Z=W/(W+K)=0.5/(0.5+0)=1$.

3. Estimate. $\widehat{\Theta}_{BS}=1(83)+0(75)=\$83.00$.

Check. With zero conditional process variance, an observation equals $\Theta$ exactly under the model, so the estimate uses the observation fully for every positive exposure. The common mistake avoided is claiming that credibility always approaches zero as exposure approaches zero; that limit requires positive process variance.

Interpreting Exposure and the Boundary Cases

When $\sigma^2>0$ and $\tau^2>0$, $K>0$. In this case the credibility factor increases with total exposure: $Z=W/(W+K)$ rises as $W$ rises. For positive $W$, it is less than one, and as total exposure becomes very large it approaches one. As positive total exposure approaches zero, it approaches zero. These statements express the balance between process noise and variation among risks: more exposure reduces the process variance of the weighted average from $\sigma^2/W$.

If instead $\sigma^2=0$ while $\tau^2>0$, then $K=0$ and $Z=1$ for every $W>0$. In that case the estimate uses the observed average fully, including as positive $W$ approaches zero; the zero-exposure limit is not zero. At $W=0$, there is no observed weighted average, so the estimate based on no experience is the collective mean $\mu$. If $\tau^2=0$, there is no variation in underlying risk means under the model, and the natural estimate is the collective mean; the displayed ratio defining $K$ is not applicable.

The units of exposure and the variance parameter must also be consistent. The model states that a period with weight $w_j$ has conditional variance $\sigma^2/w_j$. If exposure weights are rescaled, the corresponding process variance parameter must be expressed on that same scale. Treating $K$ as a dimensionless number independent of the exposure convention can produce a miscalculated credibility factor.

Common Exam Trap: Reversing the Credibility Ratio

A frequent error is to use $K/(W+K)$ as the weight on experience. For the first worked example, that incorrect calculation gives $4/(6+4)=0.4$, leading to $0.4(100)+0.6(110)=\$106$. This reverses the roles of process uncertainty and exposure: it gives less weight to experience when total exposure rises.

The correct experience weight is $Z=W/(W+K)=6/10=0.6$, so the estimate is $0.6(100)+0.4(110)=\$104$. Check the interpretation as well as the arithmetic: when process variance is positive, increasing exposure must increase the weight on the risk’s own experience. This tutorial is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

For each problem, use the Bühlmann-Straub assumptions and round money to the cent where needed.

A risk has $\mu=\$96$, $\sigma^2=324$, and $\tau^2=81$. It has observations $\$70$ and $\$100$ with weights 1 and 2. Find its estimate. (Answer: $\$93.43$)
Suppose $\sigma^2=200$, $\tau^2=50$, and total exposure is 10. What is the credibility factor? (Answer: $5/7$, approximately $0.7143$)
A risk has $\mu=\$105$, $\sigma^2=180$, and $\tau^2=45$. Its observations are $\$130$ and $\$90$ with weights 1 and 4. Find the exposure-weighted average and credibility estimate. (Answer: weighted average $\$98$; estimate $\$101.11$)
A model has $\sigma^2=0$ and $\tau^2=16$. A risk’s weighted average is $\$67$, based on total exposure 2, and the collective mean is $\$60$. Find the estimate and credibility factor. (Answer: estimate $\$67.00$; $Z=1$)
For a model with $K=3$ and positive process variance, what total exposure gives $Z=0.8$? (Answer: $W=12$)

What You'll Learn