Deriving the Bühlmann-Straub Credibility Factor

Why the Credibility Factor Depends on Exposure

In the Bühlmann-Straub model, observations can represent different amounts of exposure. A period with more exposure has less process variance than a period with less exposure, so the number of observed periods alone does not measure the strength of the experience. The credibility factor must account for both the total exposure and the variation in outcomes.

The aim here is to derive that factor by minimizing mean squared prediction error for a weighted class mean. This identifies the coefficient assigned to the class’s experience; it does not, by itself, require calculating a particular class estimate.

Model and notation: Let $\Theta$ denote a risk’s underlying parameter. For observations $X_1,\ldots,X_n$, assume that, conditional on $\Theta$, the observations are independent and have common conditional mean $\mu(\Theta)=E[X_j\mid\Theta]$. Each observation has a fixed, positive exposure weight $w_j$, with $\operatorname{Var}(X_j\mid\Theta)=\sigma^2(\Theta)/w_j$. Define the collective mean $\mu=E[\mu(\Theta)]$, the variance of hypothetical means $\nu=\operatorname{Var}(\mu(\Theta))$, and the expected process variance per unit exposure $\phi=E[\sigma^2(\Theta)]$. Assume these moments are finite and that $\nu>0$. Write total exposure as $W=\sum_{j=1}^n w_j$.

The weights are exposure amounts, not arbitrary labels. For example, if $X_j$ is a claim rate and $w_j$ is earned exposure, a larger $w_j$ gives a more stable rate under the model. The weighted class mean is

\overline{X}_w=\frac{\sum_{j=1}^n w_jX_j}{W}.

The credibility factor is the coefficient $Z$ on this weighted mean in a linear blend with the collective mean:

\widehat{\mu}_Z=Z\overline{X}_w+(1-Z)\mu.

To derive $Z$, assess prediction error against the risk’s hypothetical mean $\mu(\Theta)$. The collective mean is the complement in this derivation. The parameters $\phi$ and $\nu$ are treated as known; parameter estimation is a separate issue.

Deriving the Factor by Minimizing Error

First, the weighted mean is unbiased for the hypothetical mean, conditional on $\Theta$:

E[\overline{X}_w\mid\Theta] =\frac{\sum_j w_j E[X_j\mid\Theta]}{W} =\frac{\sum_j w_j\mu(\Theta)}{W} =\mu(\Theta).

Conditional independence gives its conditional variance. Taking the expectation over $\Theta$ then gives the expected process variance of the weighted mean:

\begin{aligned} \operatorname{Var}(\overline{X}_w\mid\Theta) &=\frac{1}{W^2}\sum_j w_j^2\operatorname{Var}(X_j\mid\Theta)\\ &=\frac{1}{W^2}\sum_j w_j^2\frac{\sigma^2(\Theta)}{w_j} =\frac{\sigma^2(\Theta)}{W},\\ E[\operatorname{Var}(\overline{X}_w\mid\Theta)] &=\frac{\phi}{W}. \end{aligned}

The prediction error of the blend can now be separated into its process-error and mean-estimation components. Since $E[\overline{X}_w-\mu(\Theta)\mid\Theta]=0$, the cross-product with $\mu(\Theta)-\mu$ has expectation zero. Thus the mean squared prediction error is

\begin{aligned} E[(\widehat{\mu}_Z-\mu(\Theta))^2] &=E\left[\left\{Z(\overline{X}_w-\mu(\Theta)) +(1-Z)(\mu-\mu(\Theta))\right\}^2\right]\\ &=Z^2\frac{\phi}{W}+(1-Z)^2\nu. \end{aligned}

The first term is error from process variation in the observed weighted mean; the second is error from relying on the collective mean when the risk’s hypothetical mean differs from it. Differentiate with respect to $Z$, set the derivative to zero, and solve:

\begin{aligned} \frac{d}{dZ}\left\{Z^2\frac{\phi}{W}+(1-Z)^2\nu\right\} &=2Z\frac{\phi}{W}-2(1-Z)\nu=0,\\ Z\frac{\phi}{W}&=(1-Z)\nu,\\ Z\left(\frac{\phi}{W}+\nu\right)&=\nu,\\ Z&=\frac{\nu}{\nu+\phi/W} =\frac{W}{W+\phi/\nu}. \end{aligned}

This is a minimum: the second derivative is $2\phi/W+2\nu$, which is positive when $W>0$, $\phi\geq 0$, and $\nu>0$. Define $K=\phi/\nu$. Then $K$ has exposure units, and the factor takes the familiar form:

Result: Under the stated Bühlmann-Straub assumptions, the credibility factor that minimizes mean squared prediction error for the weighted class mean is $$ Z=\frac{W}{W+K},\qquad K=\frac{\phi}{\nu},\qquad W=\sum_j w_j. $$ The factor increases with total exposure $W$. For fixed $W$, it decreases as expected process variance $\phi$ increases and increases as variance of hypothetical means $\nu$ increases.

The same factor minimizes prediction error for a future observation $X_{\mathrm{new}}$, provided that observation is conditionally independent of the experience and has conditional mean $\mu(\Theta)$. Its process variance adds a term that does not depend on $Z$, so it does not change the minimizing coefficient. This result is not a claim that the factor remains optimal under arbitrary dependence, changing class means, or a different model for conditional variance.

Worked Applications

Worked Example: Minimizing Error for a Given Total Exposure

A hypothetical portfolio has $\phi=2{,}400$ squared-rate units per exposure unit and $\nu=600$ squared-rate units. A class has total exposure $W=8$. Derive its credibility factor and verify the minimum prediction error. Use the mean-squared-error method; avoid treating the factor as an arbitrary percentage.

Step 1: Calculate the variance ratio. $K=\phi/\nu=2{,}400/600=4$ exposure units.

Step 2: Calculate the factor. $Z=W/(W+K)=8/(8+4)=2/3$, or $0.6667$ to four decimal places.

Step 3: Write the error function. The weighted-mean process variance is $\phi/W=2{,}400/8=300$. Therefore $M(Z)=300Z^2+600(1-Z)^2$. Its derivative is $M'(Z)=600Z-1{,}200(1-Z)=1{,}800Z-1{,}200$, which is zero at $Z=1{,}200/1{,}800=2/3$. The second derivative is $1{,}800>0$, confirming a minimum.

Step 4: Check the minimum error. At $Z=2/3$, $M(2/3)=300(4/9)+600(1/9)=1{,}200/9+600/9=200$ squared-rate units. The result agrees with the minimizing factor. A common mistake is to use $K$ as a variance in the error function; $K$ is the ratio $\phi/\nu$, while the process-variance term for this class is $\phi/W=300$.

Worked Example: Unequal Period Exposures

A hypothetical risk has three observed annual rates with exposure weights $w_1=1$, $w_2=3$, and $w_3=4$. Let $\phi=1{,}600$ and $\nu=400$, in compatible units. Find the credibility factor and the process-variance contribution to prediction error. Use total exposure rather than the count of years.

Step 1: Add the exposures. $W=1+3+4=8$.

Step 2: Form the weighted mean. $\overline{X}_w=(X_1+3X_2+4X_3)/8$. Its coefficients sum to $(1+3+4)/8=1$, as required for a mean.

Step 3: Calculate the factor and process variance. $K=1{,}600/400=4$, so $Z=8/(8+4)=2/3=0.6667$ to four decimal places. The expected process variance of the weighted mean is $\phi/W=1{,}600/8=200$.

Step 4: Verify the variance calculation from the weights. The expected conditional variance is $1{,}600\{(1/8)^2/1+(3/8)^2/3+(4/8)^2/4\}$. The terms inside braces are $1/64+3/64+4/64=8/64=1/8$, giving $1{,}600/8=200$.

The method uses the exposure-weighted mean and total exposure. A common error is to substitute $n=3$ into a formula intended for equal unit exposures: that would give $3/(3+4)=0.4286$, not $0.6667$. The exposure weights, not the number of annual observations, determine the factor here.

Worked Example: Exposure Needed to Reach a Target Factor

For a hypothetical class, $\phi=1{,}500$ and $\nu=500$. Its current periods have exposure weights $2$, $3$, and $4$. How much additional exposure is needed for the factor to reach at least $0.80$? Use the factor formula and state whether the target is met at the required total.

Step 1: Calculate $K$ and current exposure. $K=1{,}500/500=3$ exposure units, and $W_{\mathrm{current}}=2+3+4=9$. Thus the current factor is $9/(9+3)=0.75$.

Step 2: Solve the target inequality. Require $W/(W+3)\geq0.80$. Since $W+3>0$, this is equivalent to $W\geq0.80(W+3)$, so $0.20W\geq2.4$, or $W\geq12$.

Step 3: Find the additional exposure and check. The class needs $12-9=3$ additional exposure units. At $W=12$, $Z=12/(12+3)=0.80$, so the target is met exactly. Rounding the current factor upward and treating it as already at the target would be a mistake; the inequality must be solved using the unrounded factor formula.

Interpretation, Limits, and Exam Traps

The ratio $K=\phi/\nu$ is the exposure amount at which $Z=1/2$, because $K/(K+K)=1/2$. When $W=K$, expected process variance $\phi/W$ equals between-class variance $\nu$; the optimal blend gives equal weight to experience and the collective mean. More generally, increasing exposure reduces $\phi/W$, making the observed weighted mean more informative about $\mu(\Theta)$.

Check the model.
Confirm that the weights represent exposure and that conditional means, conditional variances, and independence are consistent with the stated Bühlmann-Straub assumptions.

Calculate the variance components.
Use $\phi$ for expected process variance per unit exposure and $\nu$ for variance among hypothetical means. Keep their roles and units distinct.

Sum the exposure weights.
Calculate $W=\sum_j w_j$, then evaluate $Z=W/(W+\phi/\nu)$. Do not replace $W$ with the number of periods unless every period has unit exposure.

Check the result.
For positive finite $\phi,\nu,W$, the factor lies between zero and one and increases as $W$ increases. Check boundary assumptions before applying this formula when a variance component is zero.

If $\nu=0$, all hypothetical means are identical under the model, so experience cannot improve prediction of the hypothetical mean; the optimal coefficient is $Z=0$. The ratio $K=\phi/\nu$ is then undefined, so use this interpretation rather than substituting into the ratio formula. If $\phi=0$ and $\nu>0$, positive exposure reveals the hypothetical mean without process noise, giving $Z=1$. With no experience, $W=0$, the weighted mean is unavailable and the collective mean is used. If both variance components are zero, there is no prediction uncertainty to trade off, and the minimizing coefficient is not uniquely determined by this criterion.

Key takeaway: The Bühlmann-Straub factor comes from balancing two sources of mean squared error: process variance $\phi/W$ in the experience and between-class variance $\nu$. Minimizing that error gives $Z=W/(W+\phi/\nu)$. The result is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Check Your Understanding

For each problem, use the model and definitions stated above. Answers are included for checking.

A class has $\phi=1{,}800$, $\nu=600$, and total exposure $W=9$. Find $K$ and $Z$. (Answer: $K=3$ exposure units; $Z=0.7500$.)
Three periods have weights $0.5$, $1.5$, and $2$, with $\phi=400$ and $\nu=200$. Find the total exposure and credibility factor. (Answer: $W=4$; $Z=4/(4+2)=0.6667$.)
A hypothetical class has $\phi=2{,}000$ and $\nu=500$. What total exposure is needed for $Z\geq0.75$? (Answer: $K=4$, and $W\geq12$ exposure units.)
For $\phi=900$, $\nu=300$, and weights $1,2,3$, calculate $Z$ and the expected process variance of the weighted mean. (Answer: $W=6$, $K=3$, $Z=0.6667$, and $\phi/W=150$.)
A model has $\phi=1{,}200$, $\nu=400$, and three periods each with weight $2$. A calculation uses $3/(3+3)$ for the factor. Identify the error and give the correct factor. (Answer: The calculation uses the period count rather than total exposure; $W=6$, $K=3$, so $Z=6/(6+3)=0.6667$, not $0.5000$.)

What You'll Learn