Setting Up a Linear Credibility Estimate

Why Blend Individual Experience With a Collective Mean?

An individual risk’s observed experience is informative, but it is also subject to random fluctuation. A risk with only a few observations may have a sample mean far from its underlying expected outcome. The collective mean provides a stable reference point, while the individual’s sample mean allows the estimate to reflect that risk’s own data. Linear credibility combines these two pieces of information.

The setup begins by identifying the quantity to be estimated. In the Bühlmann model, that is the risk’s hypothetical mean: its expected outcome conditional on its risk characteristics. It is not the realized outcome in a particular period, and it need not equal the risk parameter itself.

Definition: Let $\Theta$ represent a risk’s latent characteristics. For an observation $X_j$, define the risk-specific mean $m(\Theta)=E[X_j\mid\Theta]$. The collective mean is $\mu=E[m(\Theta)]$, the average risk-specific mean over the portfolio. For $n$ observations with equal exposure, the individual sample mean is $\bar X=(X_1+\cdots+X_n)/n$. A linear credibility estimate of $m(\Theta)$ is $$ \hat m=Z\bar X+(1-Z)\mu, $$ where $Z$ is the credibility weight on the individual sample mean.

The weights in this expression add to one. Thus the estimate can also be written as $\mu+Z(\bar X-\mu)$: it starts at the collective mean and moves a fraction $Z$ of the way toward the individual’s observed mean. When $Z=0$, the estimate is entirely collective; when $Z=1$, it is entirely based on the individual’s sample mean. A value between zero and one gives a convex blend.

Model Conditions and the Target

The usual Bühlmann setup treats the risk’s observations as conditionally independent and identically distributed given $\Theta$. In particular, each observation has conditional mean $m(\Theta)$. The model allows different risks to have different conditional means and can allow the conditional variance to depend on $\Theta$. The observations must have finite second moments for the mean squared error criterion to be finite.

Equal exposure matters because $\bar X$ gives each observation the same weight. If periods have different exposures, a simple unweighted sample mean may not represent the appropriate individual experience; exposure scaling and the observation unit must be handled consistently. This setup also does not assume that the conditional distribution is normal. The blend is a linear estimator, not necessarily the full conditional expectation of $m(\Theta)$ given the data.

Definition: The expected process variance is $\operatorname{EPV}=E[\operatorname{Var}(X_j\mid\Theta)]$. The variance of hypothetical means is $\operatorname{VHM}=\operatorname{Var}(m(\Theta))$. EPV describes variation within a risk around its own mean; VHM describes variation in those means across risks.

For a given model and sample size, a mean-squared-error criterion can be used to choose the blend’s weight. The target is $m(\Theta)$, so the criterion is the expected squared difference between the estimate and that risk-specific mean. If the estimator is restricted to be a credibility blend, the weight is selected from the allowed range, usually $0\leq Z\leq1$, to minimize that expected squared error. This is a statement of the estimation objective; the calculation of the minimizing weight is a separate step.

Why the Mean Squared Error Splits Into Two Sources

The risk-specific target is itself unknown, so an estimate can miss it for two conceptually different reasons. The individual’s sample mean fluctuates around its own risk-specific mean, and the risk-specific mean can differ from the collective mean. Under the stated model, these two contributions separate cleanly.

\begin{aligned} \hat m-m(\Theta) &=Z\bigl(\bar X-m(\Theta)\bigr) +(1-Z)\bigl(\mu-m(\Theta)\bigr),\\ E\!\left[(\hat m-m(\Theta))^2\right] &=Z^2\frac{\operatorname{EPV}}{n} +(1-Z)^2\operatorname{VHM}. \end{aligned}

To justify the second line, condition first on $\Theta$. Conditional independence and a common conditional mean give $E[\bar X\mid\Theta]=m(\Theta)$ and $\operatorname{Var}(\bar X\mid\Theta)=\operatorname{Var}(X_j\mid\Theta)/n$. Consequently, the conditional sample-mean error has mean zero, so its cross-product with $\mu-m(\Theta)$, which is fixed given $\Theta$, has conditional expectation zero. The expected squared sample-mean error is $\operatorname{EPV}/n$. The squared collective-anchor difference has expectation $E[(\mu-m(\Theta))^2]=\operatorname{VHM}$. Substituting these results gives the displayed criterion.

This decomposition explains the trade-off without requiring a particular distribution. Giving more weight to $\bar X$ reduces reliance on the collective anchor but leaves more process variation in the estimate. Giving less weight to $\bar X$ reduces the influence of noisy individual experience but pulls the estimate toward $\mu$, which can be distant from a particular risk’s mean. The mean squared error criterion accounts for both effects.

Constructing the Estimate in Practice

Name the target.
Specify that the estimate concerns $m(\Theta)$, the risk’s expected outcome, rather than a particular observed outcome.

Summarize individual experience.
For $n$ equal-exposure observations, calculate $\bar X$ in the same units as the collective mean.

Identify the collective anchor and weight.
Use the stated portfolio mean $\mu$ and the specified or selected $Z$; check that the two weights are $Z$ and $1-Z$.

Apply the blend and interpret it.
Compute $Z\bar X+(1-Z)\mu$, then verify that it lies between the two means when $0\leq Z\leq1$.

Worked Example: Blending Four Equal-Exposure Observations

A hypothetical insurer tracks four equal-exposure annual loss-cost observations for one account: $82$, $95$, $101$, and $112$ dollars per exposure unit. The collective mean is $100$ dollars per exposure unit, and the assigned credibility weight is $Z=0.70$. Construct the linear credibility estimate. The values are supplied as observations; no particular probability distribution, interest rate, or mortality assumption is needed for this arithmetic calculation.

Step 1: Calculate individual experience. The total is $82+95+101+112=390$. Therefore, $$ \bar X=\frac{390}{4}=97.50. $$

Step 2: Identify both weights. The individual weight is $0.70$, so the collective weight is $1-0.70=0.30$.

Step 3: Apply the blend. $$ \hat m=0.70(97.50)+0.30(100) =68.25+30.00=98.25. $$ The estimate is $98.25$ dollars per exposure unit. It lies between $97.50$ and $100$, as a convex blend should.

Check and common mistake. Reversing the weights would give $0.30(97.50)+0.70(100)=29.25+70.00=99.25$, which is not the stated estimate: it puts more weight on the collective mean than the assigned $Z=0.70$ allows. The method is to attach $Z$ to individual experience and $1-Z$ to the collective mean.

Using the Error Criterion to Compare Candidate Weights

The mean squared error expression is useful even before calculating a minimizing weight. It allows candidate blends to be compared under a specified model and makes clear which assumptions the comparison uses. Here EPV and VHM are model quantities, not statistics calculated from this one risk’s observations.

Worked Example: Comparing Two Candidate Weights

Consider a hypothetical portfolio in which $\Theta$ equals $60$ or $140$, each with probability $0.5$. Conditional on $\Theta$, observations are independent with $X_j\mid\Theta\sim N(\Theta,400)$, where the second parameter is the variance. There are $n=4$ equal-exposure observations. Compare weights $Z=0.40$ and $Z=0.70$ by their mean squared errors for estimating $m(\Theta)$. No interest or mortality assumptions apply.

Step 1: Find the collective mean. Here $m(\Theta)=\Theta$, so $$ \mu=0.5(60)+0.5(140)=30+70=100. $$

Step 2: Calculate the components. The conditional variance is $400$ at either risk type, hence $\operatorname{EPV}=400$. The hypothetical means are $60$ and $140$, each $40$ away from $100$, so $$ \operatorname{VHM}=0.5(60-100)^2+0.5(140-100)^2 =0.5(1600)+0.5(1600)=1600. $$

Step 3: Evaluate the first candidate. Since $\operatorname{EPV}/n=400/4=100$, $$ \operatorname{MSE}(0.40)=(0.40)^2(100)+(0.60)^2(1600) =16+576=592. $$

Step 4: Evaluate the second candidate. $$ \operatorname{MSE}(0.70)=(0.70)^2(100)+(0.30)^2(1600) =49+144=193. $$ The $0.70$ candidate has the lower MSE of these two candidates: $193<592$. This comparison does not establish that $0.70$ is the minimizing weight among all possible weights; it only compares the two requested values.

Check and common mistake. The sample-size adjustment applies to EPV, giving $400/4=100$, not $400$. Using $400$ in its place would overstate the process contribution by a factor of four. The VHM term is not divided by $n$, because it measures variation in the risk-specific means, not sampling variation in the observations.

When the Setup Applies—and What It Does Not Claim

The linear estimate is most directly interpreted when the observations refer to the same risk, use a consistent measurement basis, and have equal exposure. The conditional independence assumption is needed for the usual $1/n$ reduction in the sample-mean process variance. If observations are dependent, or if exposures differ and are not properly accounted for, that variance expression and the ordinary sample mean may not be appropriate.

A credibility estimate is not automatically the exact posterior mean under every probability model. A Bayesian estimate may use the full prior and likelihood and can be nonlinear in the observations. Bühlmann credibility instead sets up a best linear estimate of the risk-specific mean under squared-error loss. Its linear form is practical because it requires only the relevant first and second moment structure, rather than a fully specified conditional distribution.

The setup also distinguishes estimation from pricing decisions. The estimate concerns a risk’s expected outcome under the model. Any later adjustments for expenses, trend, coverage changes, or regulatory constraints are separate ratemaking considerations and should not be inserted into $\bar X$ or $\mu$ without an explicit modeling rationale.

Worked Example: A Count Model and the Same Credibility Structure

A hypothetical fleet program records claim counts in two equal-exposure periods. The latent claim frequency $\Theta$ is $20$ or $60$, each with probability $0.5$, and $X_j\mid\Theta\sim\operatorname{Poisson}(\Theta)$, independently across periods conditional on $\Theta$. The observed counts are $44$ and $56$. Given $Z=0.50$, construct the estimate and calculate its model-based MSE. No interest or mortality assumptions apply.

Step 1: Calculate the collective mean. For a Poisson observation, $m(\Theta)=E[X_j\mid\Theta]=\Theta$. Thus, $$ \mu=0.5(20)+0.5(60)=10+30=40. $$

Step 2: Calculate the sample mean and estimate. $$ \bar X=\frac{44+56}{2}=50, \qquad \hat m=0.50(50)+0.50(40)=25+20=45. $$ The estimate is $45$ claims per exposure period, in the same units as the observations.

Step 3: Calculate the model components. A Poisson distribution has variance equal to its mean, so the conditional variances are $20$ and $60$. Therefore, $$ \operatorname{EPV}=0.5(20)+0.5(60)=40. $$ The risk-specific means $20$ and $60$ are each $20$ away from the collective mean, giving $$ \operatorname{VHM}=0.5(20-40)^2+0.5(60-40)^2 =0.5(400)+0.5(400)=400. $$

Step 4: Evaluate the estimate’s MSE. With $n=2$, $\operatorname{EPV}/n=40/2=20$. Hence, $$ \operatorname{MSE}(0.50)=(0.50)^2(20)+(0.50)^2(400) =5+100=105. $$ This is the mean squared error under the stated model, not the squared error for the one observed risk, whose latent mean is unknown.

Check and common mistake. The two counts have equal exposure, so their arithmetic mean is appropriate. Treating the total count $100$ as the sample mean would mix the two-period total with the one-period collective mean $40$; the units would not match. Dividing the total by two gives the required period-level mean of $50$.

Common Exam Traps

Estimating the wrong quantity. The target is $m(\Theta)$, not the latest observed $X_j$, the latent label $\Theta$ in every possible model, or the collective mean alone.
Reversing the weights. In $Z\bar X+(1-Z)\mu$, $Z$ multiplies the risk’s sample mean.
Mixing units. A sample mean per exposure period cannot be blended directly with an aggregate total or a collective mean measured on a different basis.
Confusing the two variance components. EPV is within-risk process variation and is reduced by the number of equal-exposure observations; VHM is between-risk variation in hypothetical means.
Calling a candidate weight optimal without comparison or minimization. A supplied weight can be used to calculate an estimate, but that alone does not show that it minimizes mean squared error.

Key takeaway: Start with the risk-specific mean as the target, calculate the individual sample mean on a consistent exposure basis, and blend it with the collective mean using $Z$ and $1-Z$. Under conditional independence and equal exposure, the MSE criterion separates into process variation and variation among risk-specific means.

Check Your Understanding

For each problem, identify the relevant target and keep the individual and collective weights in the stated order.

A risk has four equal-exposure observations $72,84,96,108$, collective mean $100$, and $Z=0.25$. Calculate its linear credibility estimate. (Answer: $97.50$)
Three equal-exposure observations have sample mean $60$. The collective mean is also $60$, and $Z=0.60$. What is the credibility estimate, and why? (Answer: $60$; both means are equal, so every convex blend gives $60$)
For $n=5$, suppose $\operatorname{EPV}=900$, $\operatorname{VHM}=2500$, and a candidate weight is $Z=0.40$. Calculate its MSE. (Answer: $928.80$)
In one sentence, state two model conditions that support using the ordinary sample mean and the $\operatorname{EPV}/n$ process term. (Answer: The observations are conditionally independent given the risk characteristics and have equal exposure with a common conditional mean and variance structure.)
A hypothetical count model has $\Theta=30$ or $70$ with equal probability, and $X_j\mid\Theta\sim\operatorname{Poisson}(\Theta)$ independently. For $n=2$ and $Z=0.50$, calculate the model-based MSE of the credibility blend. (Answer: $106.25$)

What You'll Learn