Choosing Exposure Weights In Bühlmann-Straub Models

Why the Exposure Weight Matters

In experience rating, a period with more exposure generally provides a more precise estimate of a risk’s underlying expected loss cost than a period with less exposure. Bühlmann-Straub credibility represents that difference through an exposure weight. Choosing the weight is not merely a data-formatting step: it determines how much each period contributes to the observed experience and to the final premium.

The central question is: what measured quantity makes the conditional variance of the period’s experience inversely proportional to its weight? Depending on the data and model, that quantity may be insured units, elapsed exposure time, or expected claim volume. These are not interchangeable automatically. The weight must be tied to a defensible variance assumption and expressed in consistent units.

Definition: In the Bühlmann-Straub model, let $\Theta_i$ be the underlying expected loss cost for risk $i$, and let $Y_{ij}$ be its observed loss cost in period $j$, based on exposure weight $w_{ij}>0$. The model assumes $E[Y_{ij}\mid\Theta_i]=\Theta_i$ and $\operatorname{Var}(Y_{ij}\mid\Theta_i)=\sigma^2/w_{ij}$. Conditional on $\Theta_i$, period observations are assumed independent. The weights must measure relative information on a consistent exposure scale.

Write $\mu=E[\Theta_i]$, $a=\operatorname{Var}(\Theta_i)$, and $K=\sigma^2/a$, where $\sigma^2$ is the expected process variance for one unit of exposure and $a$ is the variance of underlying risk means. The total weight for a risk is $W_i=\sum_j w_{ij}$. Its weighted experience mean and Bühlmann-Straub credibility premium are

\overline{Y}_{i,w}=\frac{\sum_j w_{ij}Y_{ij}}{W_i}, \qquad Z_i=\frac{W_i}{W_i+K}, \qquad \widehat{P}_i=Z_i\overline{Y}_{i,w}+(1-Z_i)\mu.

The weight affects the result in two ways. It determines the weighted average of period observations, and its total determines credibility. If all periods have the same exposure, the weighted average reduces to an ordinary average. If periods differ, an ordinary average generally gives too much influence to less informative observations.

Choosing Weights From the Exposure Structure

A suitable weight follows from how the observation was constructed and how its conditional variance changes with exposure. The following cases are common, but each requires that the underlying assumptions be reasonable.

Different numbers of comparable insured units: If each unit contributes an independent observation with the same conditional variance, a period containing $n$ units has an average with conditional variance proportional to $1/n$. The number of units can therefore serve as the weight.
Different durations: If exposure accrues at a consistent rate over time and the claim process is stable over the periods being combined, elapsed exposure time can be used. A half-year of otherwise comparable exposure receives half the weight of a full year.
Different expected claim volumes: For a claim-based measure, expected claim count may be an appropriate precision weight when the model implies that the variance of the period’s average is inversely proportional to that expected count. The expected volume should be defined consistently with the measure being modeled.

In every case, the weight is a relative measure. If all weights are multiplied by the same positive constant, the weighted mean is unchanged, but $K$ must be expressed on the same rescaled exposure basis. Changing the scale of $w_{ij}$ without changing the scale of $K$ changes credibility incorrectly.

A weight should ordinarily be determined from exposure information rather than from the period’s realized claims. Using observed claim count as the weight can make a high-claim period more influential simply because it had more claims, even when the purpose is to estimate a claim rate from a fixed exposure. It can also give a period with zero claims zero weight, discarding the information that no claims occurred. Whether a different model is appropriate for count data is a separate modeling question; do not silently substitute observed claim counts for a justified exposure weight.

Worked Example: Unequal Numbers of Insured Units

A hypothetical portfolio has three periods of experience for one risk. Each period records average loss cost per insured unit. The periods contain 10, 30, and 60 comparable insured units, with loss costs of $\$130$, $\$90$, and $\$110$, respectively. Assume the conditional variance of each period average is $\sigma^2/w_j$, where $w_j$ is its unit count. The collective mean is $\mu=\$100$, and $K=50$ insured-unit weights. Calculate the Bühlmann-Straub premium.

Method and weights. The observations are per-unit averages, and the units are assumed comparable and conditionally independent. Use the unit counts as weights: $w_1=10$, $w_2=30$, and $w_3=60$. Thus $W=10+30+60=100$.

Weighted experience. The weighted total is $10(130)+30(90)+60(110)=1{,}300+2{,}700+6{,}600=10{,}600$. Therefore,

\overline{Y}_{w}=\frac{10{,}600}{100}=\$106.

Credibility and premium. Credibility is $Z=100/(100+50)=2/3$. The premium is

\widehat{P}=\frac{2}{3}(106)+\frac{1}{3}(100) =\frac{212+100}{3} =\$104.00.

Check and mistake avoided. The experience component is above the collective mean, so the premium must be between $\$100$ and $\$106$; $\$104$ satisfies that check. An unweighted average would be $(130+90+110)/3=\$110$, incorrectly giving the 10-unit period the same influence as the 60-unit period.

Duration and Expected Claim Volume

Duration is a useful weight when the underlying exposure accrues uniformly and the risk process is sufficiently stable during the periods. For example, if a policy with the same coverage contributes exposure for six months in one period and twelve months in another, the time weights may be $0.5$ and $1.0$ policy-years. This does not mean that calendar duration is always the correct weight: changes in coverage, risk mix, or claim conditions may make a simple time weight inappropriate.

Expected claim volume requires particular care. Suppose the observation is an average loss per claim and the expected number of comparable claims is larger in one period. Under a model in which claims are conditionally independent and have a stable conditional severity distribution, averaging over more expected claims typically reduces the variance of the average. In that setting, expected claim volume can represent relative precision. The exact variance model must still be checked; expected claim volume is not a universal substitute for earned exposure.

A claim count is random, so a weight based on expected claim volume is often a model-based precision measure, not the realized number of claims. If weights are intended to reflect exposure known before outcomes occur, using realized counts can create a data-dependent weighting rule that does not match the stated Bühlmann-Straub assumptions.

Worked Example: Unequal Period Durations

A hypothetical risk has three stable coverage periods with durations of 6, 12, and 18 months. The loss costs are $\$80$, $\$100$, and $\$120$ per policy-year-equivalent exposure. Assume exposure accrues uniformly, and the conditional variance of each period’s estimate is inversely proportional to its duration in years. Let the collective mean be $\mu=\$100$, and let $K=1$ policy-year. Calculate the premium.

Method and weights. Convert each duration to years: $6/12=0.5$, $12/12=1.0$, and $18/12=1.5$. The weights are therefore $0.5,1.0,1.5$, and the total is $W=3.0$ policy-years.

Weighted experience. The weighted total is $0.5(80)+1(100)+1.5(120)=40+100+180=320$. Thus the weighted loss cost is $320/3=106.6667$ dollars per policy-year-equivalent exposure before final rounding.

Credibility and premium. The credibility is $Z=3/(3+1)=0.75$. The experience indication is $\$106.6667$, which is $\$6.6667$ above the collective mean. Hence

\widehat{P}=100+0.75(106.6667-100) =100+5 =\$105.00.

Check and mistake avoided. Credibility must lie between zero and one, so the premium lies between the collective mean $\$100$ and experience mean $\$106.6667$; $\$105$ does. Treating the periods as equally informative would produce an ordinary average of $\$100$, ignoring that the 18-month period represents three times the duration of the 6-month period.

Worked Example: Expected Claim Volume as a Weight

A hypothetical liability account is summarized by average cost per claim for two periods. The expected claim volumes, set from exposure and the rating basis rather than observed counts, are 25 and 75. The period average costs are $\$88$ and $\$112$. Assume the conditional variance of each average is $\sigma^2/w_j$ with weight equal to expected claim volume. Let $\mu=\$100$ and $K=50$ expected-claim-volume units. Calculate the premium.

Method and weighted average. Under the stated variance assumption, use weights 25 and 75. Their total is $W=100$, and the weighted total cost is $25(88)+75(112)=2{,}200+8{,}400=10{,}600$. Thus $\overline{Y}_w=10{,}600/100=\$106$.

Credibility and premium. The credibility is $Z=100/(100+50)=2/3$. The credibility-weighted indication is

\widehat{P} =100+\frac{2}{3}(106-100) =100+4 =\$104.00.

Check and mistake avoided. The premium is between $\$100$ and the weighted experience mean of $\$106$, as required. Do not use the observed number of claims as the weights merely because claim-level data are available; this example explicitly uses expected claim volume and a conditional-variance assumption that supports it.

Why the Total Weight Determines Credibility

The weighted mean gives a compact estimate of the risk’s experience. Under the Bühlmann-Straub assumptions, its conditional variance explains why total exposure, rather than simply the number of periods, controls how strongly that experience should be used.

Find the conditional variance of the weighted mean.
Given $\Theta_i$, the period observations are independent and have variances $\sigma^2/w_{ij}$. Therefore the variance of their weighted sum is $\sum_j w_{ij}^2(\sigma^2/w_{ij})=\sigma^2 W_i$. Dividing by $W_i^2$ gives $\operatorname{Var}(\overline{Y}_{i,w}\mid\Theta_i)=\sigma^2/W_i$.

Compare experience variation with risk variation.
The weighted mean has mean $\Theta_i$, process variance $\sigma^2/W_i$, and variation between risks $a$. Thus its total variance is $a+\sigma^2/W_i$, while its covariance with $\Theta_i$ is $a$.

Obtain the linear credibility coefficient.
The best linear predictor coefficient on $\overline{Y}_{i,w}$ is its covariance with $\Theta_i$ divided by its variance: $a/(a+\sigma^2/W_i)=W_i/(W_i+\sigma^2/a)=W_i/(W_i+K)$. The complement is assigned to $\mu$, yielding the Bühlmann-Straub premium.

This justification depends on the stated variance structure, conditional independence, and a common process variance per unit of weight. If observations are correlated across periods, or if one period has systematically different conditional variance per stated unit, then $\sigma^2/W_i$ need not hold. In that case, the ordinary Bühlmann-Straub credibility formula may not be appropriate without adapting the model.

Common Weighting Traps

A frequent exam error is to compute a weighted average correctly and then use a credibility factor based on the wrong total weight or an incompatible $K$. Another is to use the arithmetic mean even though periods represent different exposure. The following check illustrates both the arithmetic and the interpretation.

Suppose normalized experience measures are $\$180$ with weight 40 and $\$140$ with weight 60, with $\mu=\$150$ and $K=50$ in the same exposure units. The weighted mean is $(40(180)+60(140))/100=(7{,}200+8{,}400)/100=\$156$, and $Z=100/(100+50)=2/3$. The premium is $150+(2/3)(156-150)=\$154.00$. A result of $\$154.44$ does not follow from these inputs.

Likewise, for expected claim volumes of 30 and 70, experience costs of $\$90$ and $\$110$, $\mu=\$100$, and $K=40$ in the same units, the weighted mean is $(30(90)+70(110))/100=\$104$. Credibility is $100/(100+40)=5/7$, so the premium is $100+(5/7)(104-100)=\$102.8571\ldots$, or $\$102.86$ to the nearest cent. The premium is not $\$105.71$.

Before finalizing a result, check that the exposure weights reflect the modeled observation, that all weights and $K$ use the same unit, and that the premium falls between the collective mean and weighted experience mean. These checks catch many errors without requiring a second full calculation.

Key takeaway: Choose a Bühlmann-Straub weight from a defensible model of relative precision. Use exposure units, duration, or expected claim volume only when the conditional variance assumption supports that choice, and keep the weights and $K$ on the same scale.

Check Your Understanding

For each problem, identify the appropriate weights before calculating. Unless stated otherwise, use the Bühlmann-Straub formula and round money to the nearest cent.

A risk has two periods with 10 and 30 comparable insured units and observed average costs of $\$70$ and $\$100$. Let $\mu=\$90$ and $K=20$ insured-unit weights. Calculate the weighted experience mean and premium. (Answer: weighted mean $\$92.50$; premium $\$91.67$)
A policy’s stable exposure spans 3 and 9 months, with loss costs of $\$160$ and $\$120$. Use duration in years as the weight, $\mu=\$125$, and $K=3$ policy-years. Find the weighted experience mean and premium. (Answer: weighted mean $\$130.00$; premium $\$126.25$)
Two periods have expected claim volumes of 40 and 60 and average costs of $\$95$ and $\$105$. Assume the stated expected-volume weights are justified by the variance model. If $\mu=\$100$ and $K=25$, calculate the premium. (Answer: $\$100.80$)
A hypothetical risk has two annual periods with average costs $\$80$ and $\$140$. The first period has one policy-year of exposure and the second has two. Let $\mu=\$100$ and $K=1$ policy-year. Calculate the weighted experience mean and premium, and explain why the ordinary average is not the correct experience mean. (Answer: weighted mean $\$120.00$; premium $\$115.00$; the periods have unequal exposure weights.)
Two periods have expected claim volumes of 30 and 70, costs of $\$90$ and $\$110$, $\mu=\$100$, and $K=40$ expected-claim-volume units. Find the weighted mean and premium, then state whether using realized claim counts instead of the given expected volumes is justified by the information provided. (Answer: weighted mean $\$104.00$; premium $\$102.86$; no, the stated information justifies expected-volume weights, not realized-count weights.)

What You'll Learn