Selecting Exposure Weights in Bühlmann-Straub

Why Exposure Weights Need to Match the Measure

Bühlmann-Straub credibility allows observations to have different volumes of experience. A rate based on 1,000 policy-years is generally more stable than the same kind of rate based on 100 policy-years, so the more exposed observation should normally carry more weight. The important qualification is “the same kind of rate”: the correct weight depends on how the measure was constructed and how its conditional variance changes with volume.

For example, claim frequency is based on policy exposure, while average claim severity is based on claim count. A pure premium is usually based on policy exposure. Treating all three measures as if they had the same weight can produce an inappropriate weighted average and an inconsistent model of process variance. There is no universal weight called “exposure” that automatically applies to every measure.

Definition: In the Bühlmann-Straub model, observation $X_{ij}$ for risk class $i$ and period or segment $j$ has conditional mean $\mu(\Theta_i)$ and conditional variance $\sigma^2(\Theta_i)/w_{ij}$, given the class risk parameter $\Theta_i$. The positive quantity $w_{ij}$ is the exposure weight. Observations are assumed conditionally independent given $\Theta_i$, and the variance is inversely proportional to the weight. Write $w_i=\sum_j w_{ij}$ for the class’s total weight.

The model’s weight is an information measure, not necessarily a count of policies. It should be proportional to the volume that reduces the conditional variance of the particular observation. Multiplying every weight by the same constant is possible if the variance and credibility parameters are rescaled consistently; mixing weights measured on incompatible scales is not.

Key result: Under the Bühlmann-Straub assumptions, the weighted class average is $\bar X_i=\sum_j w_{ij}X_{ij}/w_i$, and its credibility estimate is $Z_i\bar X_i+(1-Z_i)\mu$, where $Z_i=w_i/(w_i+K)$, $\mu$ is the collective mean, and $K$ is the credibility constant expressed in the same weight units. These formulas require weights consistent with the assumed inverse-variance relationship.

Match the Weight to the Ratemaking Measure

A useful way to choose weights is to write down the measure’s numerator and denominator, then ask how its conditional variance changes as the denominator grows. The following choices are standard when the corresponding exposure model is reasonable.

Ratemaking measure	Typical weight	Reason
Claim frequency: claim count divided by policy exposure	Policy exposure, such as earned policy-years	Under a Poisson count model, the variance of the frequency is inversely proportional to policy exposure.
Average claim severity: total losses divided by claim count	Claim count	With independent claims of a common severity distribution, the variance of the average is inversely proportional to the number of claims.
Pure premium or loss cost: total losses divided by policy exposure	Policy exposure	Under a compound Poisson model with a stable loss process per exposure unit, the variance of the loss cost is inversely proportional to exposure.
Loss ratio: losses divided by earned premium	Earned premium, when the loss process scales with premium	Under that model, more premium represents more volume and the variance of the ratio is inversely proportional to premium.

These are model-based choices, not rules independent of assumptions. For instance, premium is an appropriate loss-ratio weight when expected claim volume scales with premium in the relevant portfolio. If premium varies for reasons unrelated to the volume or variability of losses, premium alone may not describe the conditional variance. In that case, the analyst must justify another weight or reconsider whether the simple Bühlmann-Straub structure fits.

Why the Weighted Average Uses These Weights

The weighted average is not merely a convention. Under the model, it is the linear unbiased combination of observations with the smallest conditional variance. To see this, fix a risk class and suppress the index $i$. Suppose $E[X_j\mid\Theta]=\mu(\Theta)$, $\operatorname{Var}(X_j\mid\Theta)=\sigma^2(\Theta)/w_j$, and the observations are conditionally independent. Consider an unbiased linear estimator $\sum_j a_jX_j$, so that $\sum_j a_j=1$.

\operatorname{Var}\left(\sum_j a_jX_j\mid\Theta\right) =\sigma^2(\Theta)\sum_j\frac{a_j^2}{w_j}.

Minimize the sum on the right subject to $\sum_j a_j=1$. Using a Lagrange multiplier $\lambda$, differentiate $\sum_j a_j^2/w_j-\lambda(\sum_j a_j-1)$ with respect to $a_j$. This gives $2a_j/w_j-\lambda=0$, so $a_j=\lambda w_j/2$. Applying the constraint yields $\lambda/2=1/\sum_jw_j$, and therefore:

a_j=\frac{w_j}{\sum_k w_k}, \qquad \operatorname{Var}\left(\sum_j a_jX_j\mid\Theta\right) =\frac{\sigma^2(\Theta)}{\sum_jw_j}.

Thus an observation receives a share of the average in proportion to its information weight. If a candidate uses policy exposure to weight severity averages, for example, the resulting average is not generally the minimum-variance average under a model in which severity precision is determined by claim count.

Worked Applications

Worked Example: Choosing Weights for Claim Frequency

A hypothetical insurer observes three fleet segments. Segment A has 24 claims over 400 earned vehicle-years, segment B has 40 claims over 1,000 earned vehicle-years, and segment C has 6 claims over 300 earned vehicle-years. Estimate the pooled claim frequency per 100 vehicle-years using the appropriate weights. Assume conditionally Poisson claim counts with mean $\lambda e$ and variance $\lambda e$, where $e$ is vehicle-year exposure. No interest or mortality assumptions are involved.

Method. For $N\sim\operatorname{Poisson}(\lambda e)$, the frequency $N/e$ has conditional mean $\lambda$ and variance $\lambda/e$. Its variance is inversely proportional to vehicle-year exposure, so use $e$ as the weight.

The segment frequencies per vehicle-year are A: $24/400=0.060$, B: $40/1{,}000=0.040$, and C: $6/300=0.020$. The total number of claims is $24+40+6=70$, and total exposure is $400+1{,}000+300=1{,}700$ vehicle-years. The exposure-weighted frequency is therefore $70/1{,}700=0.04117647$ claims per vehicle-year.

Converting to a rate per 100 vehicle-years gives $0.04117647(100)=4.117647$, or 4.118 claims per 100 vehicle-years to three decimals. Check: weighting the three rates by 400, 1,000, and 300 and dividing by 1,700 gives the same result. The common mistake is to give each segment equal weight: that would produce $(0.060+0.040+0.020)/3=0.040$, or 4.000 per 100 vehicle-years, despite the segments having unequal exposure.

Worked Example: Choosing Weights for Average Severity

A hypothetical repair program records three groups of claims. Group A has 12 claims totaling $36,000; group B has 40 claims totaling $140,000; group C has 8 claims totaling $20,000. Estimate the pooled average severity. Assume claim severities within a group are independent and identically distributed with conditional mean $m$ and variance $s^2$; the sample means have variance $s^2/n$. No interest or mortality assumptions are involved.

Method. Average severity is total loss divided by claim count. Under the stated assumptions, its conditional variance decreases in proportion to the number of claims, not the policy exposure that might have generated those claims. Use claim count as the weight.

The group averages are A: $\$36{,}000/12=\$3{,}000$, B: $\$140{,}000/40=\$3{,}500$, and C: $\$20{,}000/8=\$2{,}500$. Total losses are $\$36{,}000+\$140{,}000+\$20{,}000=\$196{,}000$; total claim count is $12+40+8=60$. The claim-count-weighted average is $\$196{,}000/60=\$3{,}266.6667$, or $3,266.67 to the nearest cent.

Check: $(12(\$3{,}000)+40(\$3{,}500)+8(\$2{,}500))/60=\$196{,}000/60$. The common mistake is to average the three group averages equally, obtaining $(\$3{,}000+\$3{,}500+\$2{,}500)/3=\$3{,}000$. That calculation gives a group with 8 claims the same influence as a group with 40 claims.

Worked Example: Selecting Weights for Loss Ratios

A hypothetical program has two segments. Segment A has $120,000 of earned premium and $72,000 of losses; segment B has $60,000 of earned premium and $48,000 of losses. Estimate the pooled loss ratio. Assume losses follow a compound Poisson model in which the claim count has conditional mean $\lambda P$, with $P$ equal to premium measured in dollars, and claim amounts are independent with conditional mean $m$ and variance $s^2$. Thus the model’s loss process scales with premium. No interest or mortality assumptions are involved.

Method. The segment loss ratios are A: $\$72{,}000/\$120{,}000=0.60$, or 60%, and B: $\$48{,}000/\$60{,}000=0.80$, or 80%. In the stated model, expected losses are $\lambda mP$ and the variance of losses is $\lambda P(s^2+m^2)$. Dividing losses by $P$ makes the ratio variance proportional to $1/P$, so use earned premium as the weight.

Total losses are $\$72{,}000+\$48{,}000=\$120{,}000$, and total earned premium is $\$120{,}000+\$60{,}000=\$180{,}000$. The pooled ratio is $\$120{,}000/\$180{,}000=0.6666667$, or 66.67% to two decimal places. Equivalently, $(\$120{,}000(0.60)+\$60{,}000(0.80))/\$180{,}000=0.6666667$.

Check: the pooled ratio lies between 60% and 80%, and is closer to 60%, the ratio with more premium. An equal average would be 70%, which is not the premium-weighted result. The common mistake is to assume premium is always the right weight for any loss ratio: it is justified here by the explicit model linking claim volume to premium; another portfolio may require a different variance model.

How Inconsistent Weights Distort Credibility

The weights affect both the class average and the credibility assigned to it. In a standard Bühlmann-Straub formulation, let $\phi$ be the expected process variance parameter and $\tau^2$ the variance of hypothetical means. Then $K=\phi/\tau^2$, using the same weight scale as $w_i$. If the weights do not reflect the observations’ actual precision, the weighted average can emphasize the wrong experience, and the assumed process variance can be wrong. The resulting $Z_i$ may then give too much or too little credibility.

Worked Example: The Cost of Assigning Equal Weights

A hypothetical class has two observed rates: 120 based on weight 200 and 80 based on weight 800. The collective mean is $\mu=100$, and the model parameters are $\phi=1{,}600$ and $\tau^2=4$, so $K=\phi/\tau^2=400$. Assume the conditional variance of each rate is $\phi/w_j$, with conditional independence; no parametric distribution is assumed. No interest or mortality assumptions are involved.

Correct calculation. The total weight is $200+800=1{,}000$. The weighted class mean is $(200(120)+800(80))/1{,}000=(24{,}000+64{,}000)/1{,}000=88$. The credibility factor is $Z=1{,}000/(1{,}000+400)=1{,}000/1{,}400=0.7142857$. The estimate is $0.7142857(88)+(1-0.7142857)(100)=62.8571+28.5714=91.4286$, or 91.43 to two decimals.

The model’s conditional variances for the two observations are $1{,}600/200=8$ and $1{,}600/800=2$. Now incorrectly assign weight 800 to each observation. The equal-weight class mean becomes $(120+80)/2=100$, total assigned weight is 1,600, and assigned credibility is $1{,}600/(1{,}600+400)=0.80$. The resulting estimate is $0.80(100)+0.20(100)=100$, not 91.43.

This is not merely a different averaging convention. The incorrect model assigns each observation variance $1{,}600/800=2$, although the actual conditional variances under the stated model are 8 and 2. It treats the less precise observation as if it were four times as precise, and also inflates total weight from 1,000 to 1,600. The common mistake is to replace unequal weights with equal weights without revisiting either the variance model or the credibility factor. Check: the correctly weighted estimate lies between 88 and 100, pulled toward the collective mean; the incorrect equal weights erase the observed difference from that mean.

Practical Checks Before Using Weights

Before calculating a Bühlmann-Straub estimate, identify the observation being modeled and the volume that controls its variability. A frequency needs a clearly defined exposure base; a severity needs a claim count; a pure premium needs exposure; and a loss ratio needs a defensible connection between premium and loss-process volume. Use consistent units across periods, such as policy-years rather than a mixture of policy-years and policy-months.

Also distinguish the weight used to combine observed rates from the raw numerator or denominator. For example, when segment loss ratios are combined with premium weights, the calculation is equivalent to total losses divided by total premium. It is not equivalent to averaging loss amounts, nor to averaging ratios without weights. If the same data are expressed in different units, such as dollars versus thousands of dollars, all relevant weights and model parameters must be kept on a coherent scale.

Key takeaway: Select $w_{ij}$ so that conditional variance is proportional to $1/w_{ij}$ for the observation being modeled. Then use those same weights in the class average and in the credibility calculation. A numerically convenient weight is not acceptable unless it represents the observation’s information under the assumed model.

Check Your Understanding

For each situation, identify the suitable weight and calculate the requested pooled quantity.

A hypothetical book reports 36 claims over 900 earned policy-years and 28 claims over 700 earned policy-years. Under a Poisson claim-count model, find the pooled frequency per 100 policy-years. (Answer: 4.000 claims per 100 policy-years.)
Two repair groups have total losses of $54,000 from 18 claims and $126,000 from 42 claims. With independent claims from a common severity distribution, find the pooled average severity. (Answer: $3,000.00; use claim counts.)
A hypothetical program has $120,000 premium and $72,000 losses in one segment, and $60,000 premium and $48,000 losses in another. If the loss process scales with premium, find the pooled loss ratio. (Answer: 66.67%; use earned premium.)
Two property portfolios have losses of $45,000 over 900 exposure units and $50,000 over 1,000 exposure units. Find the pooled pure premium. (Answer: $50.00 per exposure unit.)
Observed rates are 70 with weight 300 and 100 with weight 700. Given collective mean 80 and $K=500$, calculate the credibility estimate. (Answer: $Z=2/3$, weighted mean 91, estimate 87.33.)

What You'll Learn