Constructing the Credibility-Weighted Estimate

From Individual Experience to a Credibility-Weighted Estimate

An individual risk’s observed experience is informative, but it may be based on only a few observations and can vary substantially from period to period. The collective mean provides a broader reference point. The Bühlmann estimate combines these two sources: it gives weight $Z$ to the individual’s observed mean and weight $1-Z$ to the collective mean.

The result is an estimate of the risk’s expected value on the same basis and in the same units as the observations. For example, if observations are annual loss costs per exposure, the estimate is also an annual loss cost per exposure. It is not, by itself, a loaded premium that includes expenses or profit.

Definition: Let $X_1,\ldots,X_n$ be the observations for one risk, and let $\bar X=(X_1+\cdots+X_n)/n$ be that risk’s observed mean. Let $\mu$ be the collective mean and $Z$ the Bühlmann credibility factor. The credibility-weighted estimate is $$ \widehat{m}=Z\bar X+(1-Z)\mu. $$ Here $Z\bar X$ is the individual-experience contribution, and $(1-Z)\mu$ is the collective contribution.

In the standard Bühlmann model, each risk has an underlying risk parameter $\Theta$. Conditional on $\Theta$, observations are identically distributed and independent, with conditional mean $m(\Theta)=E[X\mid\Theta]$ and conditional variance $v(\Theta)=\operatorname{Var}(X\mid\Theta)$. The collective mean is $\mu=E[m(\Theta)]$. The process variance, also called EPV, is $\operatorname{EPV}=E[v(\Theta)]$; the variance of hypothetical means, or VHM, is $\operatorname{VHM}=\operatorname{Var}(m(\Theta))$. When VHM is positive, $K=\operatorname{EPV}/\operatorname{VHM}$, and for $n$ equal-weight observations $Z=n/(n+K)$. These definitions explain where the weight comes from; in many calculations, $Z$ or $K$ is supplied.

The blend is a weighted average: $Z$ and $1-Z$ are nonnegative and sum to one when $0\leq Z\leq1$. Consequently, the estimate lies between $\bar X$ and $\mu$, unless those values are equal. A larger weight on the individual mean moves the estimate toward that risk’s experience; a larger weight on the collective mean moves it toward the portfolio benchmark.

Why the Estimate Uses These Weights

The Bühlmann estimate is the best linear credibility estimate of the risk-specific mean $m(\Theta)$ under the model. “Linear” means it has the form $\mu+a(\bar X-\mu)$ for a constant $a$. The coefficient is chosen to minimize mean squared error over the risks and observations in the model.

Conditional independence gives $\operatorname{Var}(\bar X\mid\Theta)=v(\Theta)/n$. Averaging over risks, the law of total variance gives $\operatorname{Var}(\bar X)=\operatorname{VHM}+\operatorname{EPV}/n$. Also, $\operatorname{Cov}(\bar X,m(\Theta))=\operatorname{VHM}$: conditional on $\Theta$, the mean of $\bar X$ is $m(\Theta)$, so the variation shared by the two quantities is the variation in $m(\Theta)$ across risks.

For an estimate $\mu+a(\bar X-\mu)$, its error relative to $m(\Theta)$ is $a(\bar X-\mu)-(m(\Theta)-\mu)$. Using the variance and covariance just stated, its mean squared error is

\begin{aligned} \operatorname{MSE}(a) &=a^2\left(\operatorname{VHM}+\frac{\operatorname{EPV}}{n}\right) +\operatorname{VHM}-2a\operatorname{VHM}\\ &=(1-a)^2\operatorname{VHM}+a^2\frac{\operatorname{EPV}}{n}. \end{aligned}

Differentiating with respect to $a$ and setting the derivative equal to zero yields

-2(1-a)\operatorname{VHM} +2a\frac{\operatorname{EPV}}{n}=0, \qquad a=\frac{\operatorname{VHM}}{\operatorname{VHM}+\operatorname{EPV}/n} =\frac{n}{n+K}=Z.

Thus the coefficient on the observed mean is $Z$, not $1-Z$. This derivation relies on the stated Bühlmann model and equal-weight observations; if observations have unequal exposures, the simple formula using the number $n$ does not automatically apply.

Calculating and Interpreting the Blend

Use a consistent sequence: determine the individual mean on the stated observation basis, calculate or identify $Z$, apply each weight to the correct mean, and add the contributions. Keep the units visible throughout. The collective mean must be for a comparable population and measure; a mean on a different coverage, time, or exposure basis cannot be blended meaningfully without first making the bases consistent.

Worked Example: Blending Two Means When K Is Given

A hypothetical insurer considers a risk with six equal annual observations of loss cost. The risk’s observed mean is $820 per exposure, the collective mean is $700 per exposure, and $K=4$. Calculate the Bühlmann estimate and interpret it.

Method. Calculate $Z$ from $n$ and $K$, then apply $Z$ to the individual mean and $1-Z$ to the collective mean.

There are $n=6$ observations, so

Z=\frac{6}{6+4}=\frac{6}{10}=0.60, \qquad 1-Z=0.40.

The individual contribution is $0.60\times \$820=\$492$ per exposure. The collective contribution is $0.40\times \$700=\$280$ per exposure. Therefore,

\widehat{m}=\$492+\$280=\$772\text{ per exposure}.

The estimate is $772 per exposure, between the observed mean of $820 and the collective mean of $700. The weights total 1.00, and the result is closer to the individual mean because that mean receives the larger weight. Common mistake avoided: applying the collective weight to the individual mean would reverse the intended blend.

Worked Example: Finding the Individual Mean Before Applying Z

For a hypothetical risk, four equal-period loss costs are $540, $660, $780, and $1,020 per exposure. The collective mean is $600 per exposure and $K=6$. Find the credibility-weighted estimate.

Method. First calculate the arithmetic mean of the four observations. Then use the number of observations, not the number of dollars or the sum of losses, in the credibility formula.

The observation total is $\$540+\$660+\$780+\$1{,}020=\$3{,}000$. Thus the individual mean is $\bar X=\$3{,}000/4=\$750$ per exposure. The credibility factor and its complement are

Z=\frac{4}{4+6}=0.40, \qquad 1-Z=0.60.

The individual contribution is $0.40\times \$750=\$300$ per exposure. The collective contribution is $0.60\times \$600=\$360$ per exposure. Adding them gives

\widehat{m}=\$300+\$360=\$660\text{ per exposure}.

The result lies between $600 and $750, as a weighted average should. Common mistake avoided: using the total of $3,000 as the individual experience would mix a total with per-exposure means and produce an invalid estimate.

Worked Example: Working From a Longer Set of Observations

A hypothetical specialty coverage risk has ten equal annual observations, in dollars per exposure: $500, $650, $700, $800, $900, $550, $750, $850, $600, and $700. Suppose $K=6$ and the comparable collective mean is $640 per exposure. Calculate the estimate and explain which experience source has more influence.

Method. Sum and average the ten observations, calculate the credibility factor, and compute both weighted contributions.

Adding the values gives $\$500+\$650+\$700+\$800+\$900+\$550+\$750+\$850+\$600+\$700=\$7{,}000$. Therefore, $\bar X=\$7{,}000/10=\$700$ per exposure. The credibility factor is

Z=\frac{10}{10+6}=\frac{10}{16}=0.625, \qquad 1-Z=0.375.

The individual contribution is $0.625\times \$700=\$437.50$ per exposure. The collective contribution is $0.375\times \$640=\$240$ per exposure. Thus

\widehat{m}=\$437.50+\$240=\$677.50\text{ per exposure}.

The estimate is $22.50 below the individual mean and $37.50 above the collective mean. The individual mean has more influence because its weight, 0.625, exceeds the collective weight, 0.375. Common mistake avoided: counting the ten observations to calculate $Z$ does not change the definition of $\bar X$; the individual input remains their mean, not their total.

Checks That Catch Common Errors

A short set of checks can identify most calculation mistakes before an answer is reported:

Check the weights. Confirm that $Z+(1-Z)=1$, and that both weights are between zero and one.
Check the direction. The individual mean receives $Z$; the collective mean receives $1-Z$.
Check the inputs. Use the individual arithmetic mean, not the sum, when the model is based on equal-weight observations.
Check the result. A weighted average must fall between the two input means when both weights are between zero and one.
Check comparability. The two means must use the same units, period basis, and type of observation.

For example, in the second worked example, reversing the weights gives $0.40(\$600)+0.60(\$750)=\$690$, not the correct $660. That calculation still falls between the two input means, so the range check alone does not catch it. Verify which mean is assigned to each weight as well as checking the final range.

Key takeaway: Once $Z$ and the two comparable means are established, the estimate is $Z\bar X+(1-Z)\mu$. Keep the risk’s observed mean attached to $Z$, keep the collective mean attached to $1-Z$, and confirm that the result has the same units as both inputs.

Check Your Understanding

For each hypothetical risk, calculate the individual mean if needed, find or use the stated credibility factor, and report the estimate in the stated units.

A risk has three equal annual loss-cost observations with mean $900 per exposure. Its collective mean is $720 per exposure and $K=3$. Find $Z$ and the estimate. (Answer: $Z=0.500$; $810 per exposure.)
Four equal-period observations are $400, $600, $800, and $1,200 per exposure. The collective mean is $650 per exposure and $K=4$. Calculate the individual mean, $Z$, and the estimate. (Answer: Individual mean = $750 per exposure; $Z=0.500$; estimate = $700 per exposure.)
A risk has eight equal observations with an individual mean of $1,050 per unit. Its collective mean is $900 per unit and $K=2$. Find the credibility estimate and identify which mean receives more weight. (Answer: $Z=0.800$; estimate = $1,020 per unit; the individual mean receives more weight.)
Five observations, in dollars per exposure, are $300, $500, $700, $900, and $1,100. The collective mean is $620 per exposure and $K=5$. Calculate the mean, both weights, and the final estimate. (Answer: Individual mean = $700 per exposure; $Z=0.500$, $1-Z=0.500$; estimate = $660 per exposure.)
A hypothetical risk has an observed mean of $1,400 per contract and the comparable collective mean is $1,100 per contract. If the individual credibility factor is $Z=0.35$, calculate the estimate and verify it is between the two inputs. (Answer: $1,205 per contract; it is between $1,100 and $1,400.)

What You'll Learn