Applying Bühlmann Credibility To Risk Premiums

Why Credibility-Weighted Premiums Use a Collective Complement

An experience-rated premium should respond to a risk’s own history, but a short history can be unusually high or low by chance. Bühlmann credibility balances that observed experience against the collective mean: the risk receives more weight on its own record when the record is more informative, and more weight on the collective when it is less informative.

This tutorial focuses on calculating and interpreting that premium. The experience observations and the collective mean must be on the same basis—for example, annual loss costs per exposure—and the credibility factor must correspond to the amount and structure of experience being used.

Definition: Let $X_1,\ldots,X_n$ be a risk’s experience observations, and let $\bar X$ be their arithmetic mean. Let $\mu$ be the collective mean and $Z$ the risk’s credibility factor, with $0\le Z\le1$. The Bühlmann credibility-weighted premium is $$ \widehat P=Z\bar X+(1-Z)\mu. $$ Here $\bar X$ is the risk’s experience component and $\mu$ is the collective complement. The formula is a weighted average: the weights sum to one.

The premium can also be written as an adjustment to the collective: $$ \widehat P=\mu+Z(\bar X-\mu). $$ Thus, the risk’s experience changes the collective premium by the fraction $Z$ of the difference between its average experience and the collective mean. This is a useful interpretation of the calculation, not a separate experience-modification method.

Model Assumptions and the Credibility Factor

In the classical Bühlmann model, a risk has an unobserved risk parameter $\Theta$. Given $\Theta$, its observations have a common conditional mean $m(\Theta)=E[X_j\mid\Theta]$, and observations from different periods are conditionally independent. The observations are identically distributed conditional on $\Theta$, so each has conditional variance $s^2(\Theta)=\operatorname{Var}(X_j\mid\Theta)$. Define the collective mean $\mu=E[m(\Theta)]$, the variance of hypothetical means $a=\operatorname{Var}(m(\Theta))$, and the expected process variance $v=E[s^2(\Theta)]$. For $a>0$, the Bühlmann factor for $n$ equal, independent observations is $$ Z=\frac{n}{n+K}=\frac{n}{n+v/a},\qquad K=\frac{v}{a}. $$

The parameter $a$ measures variation in expected experience among risks; $v$ measures average variation within a risk from period to period. A large $a$ relative to $v$ makes observed experience more informative and increases $Z$. More observations also increase $Z$. If a problem supplies $Z$ directly, use that factor; if it supplies $a$, $v$, and $n$, calculate $Z$ first.

These expressions rely on equal-exposure observations with the stated common conditional distribution. They should not be applied mechanically to unequal exposures or changing observation structures. Nor should the linear Bühlmann premium automatically be described as the exact conditional expected loss for every possible distribution: it is the best linear credibility predictor under the model’s moment structure.

Why the Bühlmann Weight Has This Form

The variance decomposition explains how the experience average’s precision determines the credibility factor. Given $\Theta$, the conditional mean of $\bar X$ is $m(\Theta)$. Conditional independence and equal conditional variances imply $\operatorname{Var}(\bar X\mid\Theta)=s^2(\Theta)/n$. Taking its expectation gives $E[\operatorname{Var}(\bar X\mid\Theta)]=v/n$, not $v/n$ as a general value of the conditional variance itself. By the law of total variance,

\operatorname{Var}(\bar X) =\operatorname{Var}(E[\bar X\mid\Theta]) + E[\operatorname{Var}(\bar X\mid\Theta)] =a+\frac{v}{n}.

To predict the risk’s expected experience $m(\Theta)$ with a linear function of $\bar X$, use a predictor of the form $\mu+c(\bar X-\mu)$. Its least-squares coefficient is $\operatorname{Cov}(m(\Theta),\bar X)/\operatorname{Var}(\bar X)$. Since $E[\bar X\mid\Theta]=m(\Theta)$, conditional covariance with $m(\Theta)$ is zero, and therefore

\operatorname{Cov}(m(\Theta),\bar X) =\operatorname{Cov}(m(\Theta),E[\bar X\mid\Theta]) =\operatorname{Var}(m(\Theta))=a.

Substitution gives $c=a/(a+v/n)=n/(n+v/a)=Z$. Hence the linear predictor is $\mu+Z(\bar X-\mu)$, equivalent to $Z\bar X+(1-Z)\mu$. This derivation also shows why the weight moves toward the risk’s own average as the experience becomes more precise.

Worked Premium Calculations

Worked Example: Using a Supplied Credibility Factor

A hypothetical insurer is rating a risk using four equal-exposure annual loss-cost observations. The observations are \$1,200, \$1,400, \$1,300, and \$1,700. The collective mean is \$1,300, and the supplied Bühlmann factor is $Z=0.40$. Calculate the premium. For model context, suppose $\Theta$ has probabilities $1/2$ at each of two values, with conditional means \$1,100 and \$1,500, respectively; given $\Theta$, each annual observation is normally distributed with its conditional mean and variance \$240,000, and observations are conditionally independent. No interest or mortality assumption applies: these are annual loss-cost amounts, not discounted benefits or life-contingent payments.

Method. Calculate the arithmetic mean of the four observations, then weight it and the collective mean using $Z$ and $1-Z$.

Step 1: Calculate the risk’s average experience. $$ \bar X=\frac{1{,}200+1{,}400+1{,}300+1{,}700}{4} =\frac{5{,}600}{4}=\$1{,}400. $$

Step 2: Calculate the collective weight. Since $1-Z=1-0.40=0.60$, the collective receives 60% weight.

Step 3: Calculate the premium. $$ \widehat P=0.40(1{,}400)+0.60(1{,}300) =560+780=\$1{,}340. $$

Check. The experience is \$100 above the collective. Applying 40% of that difference gives $1{,}300+0.40(100)=\$1{,}340$, as above. The premium lies between the two inputs. The common mistake to avoid is applying the collective weight to experience, or vice versa.

Worked Example: Calculating the Factor From Model Parameters

For a hypothetical class, let the risk parameter $\Theta$ take two equally likely values. Conditional annual loss costs are normally distributed with mean \$900 when $\Theta$ is in the first state and \$1,500 in the second state, and with variance \$360,000 in either state. Six annual observations for one risk are conditionally independent. Its observed average loss cost is \$1,100. Calculate its Bühlmann premium. No interest or mortality assumption applies.

Method. Find the collective mean, $a$, and $v$; use them to calculate $Z$, then calculate the weighted premium.

Step 1: Calculate the collective mean. $$ \mu=\frac{900+1{,}500}{2}=\$1{,}200. $$

Step 2: Calculate the variance of hypothetical means. Each conditional mean is \$300 from the collective mean, so $$ a=\frac{1}{2}(900-1{,}200)^2+\frac{1}{2}(1{,}500-1{,}200)^2 =\frac{1}{2}(90{,}000)+\frac{1}{2}(90{,}000)=90{,}000. $$ The expected process variance is $v=\$360{,}000$, because the conditional variance is the same in both states.

Step 3: Calculate the credibility factor. $$ K=\frac{v}{a}=\frac{360{,}000}{90{,}000}=4,\qquad Z=\frac{6}{6+4}=0.60. $$

Step 4: Calculate and check the premium. $$ \widehat P=0.60(1{,}100)+0.40(1{,}200) =660+480=\$1{,}140. $$ The observed average is \$100 below the collective; 60% of that difference is \$60, so the premium is \$1,200-\$60=\$1,140. The common mistake to avoid is reversing $v/a$: doing so would incorrectly give $K=0.25$, rather than 4, and an unjustifiably high credibility factor.

Worked Example: Applying a Supplied Factor to a Period Average

A hypothetical commercial property risk has three equal-exposure annual loss-cost observations: \$1,800, \$2,100, and \$1,500. Its collective mean is \$1,600 and its supplied Bühlmann factor is $Z=0.35$. Calculate the premium. Assume the observations arise from a conditionally independent normal model with common conditional variance given $\Theta$, and that the supplied factor is appropriate for these three periods. No interest or mortality assumption applies.

Method. Average the three period amounts before applying the supplied credibility factor.

Step 1: Calculate the average. $$ \bar X=\frac{1{,}800+2{,}100+1{,}500}{3} =\frac{5{,}400}{3}=\$1{,}800. $$

Step 2: Determine the two weights. The experience weight is $0.35$, and the collective weight is $1-0.35=0.65$.

Step 3: Calculate the premium. $$ \widehat P=0.35(1{,}800)+0.65(1{,}600) =630+1{,}040=\$1{,}670. $$

Check. The risk average exceeds the collective by \$200. The premium increases the collective by $0.35(200)=\$70$, yielding \$1,670. It lies between \$1,600 and \$1,800. The common mistake to avoid is using the most recent observation, \$1,500, in place of the three-period average.

Common Errors and Practical Checks

A calculated credibility premium is easier to audit when each input and weight is explicit. Keep the experience average and collective mean on the same basis and in the same units, and use the experience period count that matches the stated model. Do not treat unequal exposures as equal observations without an appropriate adjustment; credibility methods for unequal exposure require a different formulation.

Reversing the weights. $Z$ multiplies the risk’s experience; $1-Z$ multiplies the collective mean.
Using a total where an average is required. The formula uses $\bar X$, not the sum of $n$ observations.
Misreading the model variance. In general, $\operatorname{Var}(\bar X\mid\Theta)=s^2(\Theta)/n$ depends on $\Theta$. The quantity $v/n$ is its expectation under equal conditional process variances across periods.
Claiming more than the model supports. The Bühlmann premium is a linear credibility predictor. It is not necessarily the exact posterior mean for every distribution.

As a final check, the premium must be between $\bar X$ and $\mu$ when $0\le Z\le1$. If $\bar X>\mu$, the premium must be at least $\mu$ and no greater than $\bar X$; if $\bar X<\mu$, it must be at least $\bar X$ and no greater than $\mu$. A result outside that interval usually signals an incorrect weight, sign, or input.

Key takeaway: Calculate the risk’s average experience, identify the collective mean, and apply the credibility factor to experience and its complement to the collective. For equal independent observations under the Bühlmann model, $Z=n/(n+v/a)$; larger $Z$ moves the premium closer to the risk’s own experience.

Check Your Understanding

For each hypothetical situation, calculate the requested premium or factor. All amounts are annual loss costs, and no interest or mortality assumption applies.

A risk has three equal-exposure observations of \$700, \$900, and \$1,000. The collective mean is \$800 and $Z=0.50$. Find its Bühlmann premium. (Answer: \$833.33)
Under a conditional normal model, two equally likely risk states have means \$1,000 and \$1,400; the conditional variance in either state is \$160,000. For four conditionally independent observations with average experience \$1,300, calculate $a$, $v$, $Z$, and the premium. (Answer: $a=40{,}000$, $v=160{,}000$, $Z=0.50$, premium \$1,250)
A risk has six equal-exposure observations averaging \$2,400. The collective mean is \$2,000, with $a=100{,}000$ and $v=500{,}000$. Calculate $K$, $Z$, and the premium. (Answer: $K=5$, $Z=6/11\approx0.5455$, premium \$2,218.18)
A risk’s average experience is \$1,450, the collective mean is \$1,600, and the applicable credibility factor is 0.30. Calculate the premium and state whether it is closer to experience or to the collective. (Answer: \$1,555; closer to the collective)
For two equally likely risk states, the conditional means are \$800 and \$1,200, and the common conditional variance is \$90,000. A risk has five equal-exposure, conditionally independent observations averaging \$1,100. Calculate the collective mean, $a$, $v$, $Z$, and premium. (Answer: $\mu=\$1{,}000$, $a=40{,}000$, $v=90{,}000$, $Z=20/29\approx0.6897$, premium $\$1{,}068.97$)

What You'll Learn