Applying Bühlmann-Straub With Unequal Exposure

Why Unequal Exposure Matters

A risk’s annual experience is not equally informative in every period if its exposure changes. For example, an annual loss cost based on 2,000 insured units generally carries more information about the underlying rate than one based on 200 units. Bühlmann-Straub credibility accounts for this difference by using exposure both when aggregating historical experience and when determining how much credibility to assign.

The key distinction is between a rate per exposure unit and a total loss amount. The method combines annual rates using exposure weights; it does not give each year equal weight merely because each year is one observation. The resulting credibility estimate is a prospective expected loss cost per exposure unit. Multiplying that rate by the expected future exposure gives the corresponding expected loss amount, before expenses or other premium components.

Definition: In the Bühlmann-Straub model, risk $i$ has an unknown expected loss rate $\theta_i$. For period $j$, $w_{ij}$ is its exposure and $X_{ij}$ is its observed loss cost per exposure unit. Conditional on $\theta_i$, the observations have mean $\theta_i$, variance $\sigma^2/w_{ij}$, and are independent across periods. Across risks, $E[\theta_i]=\mu$ and $\operatorname{Var}(\theta_i)=a$. Here $\mu$ is the collective mean, $\sigma^2$ is the process variance per unit of exposure, and $a$ is the variance of hypothetical means.

The model assumes the observations being combined are on a comparable basis: the same type of loss cost, exposure definition, and cost level. If experience from different periods has not been adjusted for relevant changes, the apparent differences may reflect those changes rather than risk variation. Exposure must also represent information in the model: doubling exposure halves the conditional variance of the observed rate.

Aggregate Experience Before Applying Credibility

For one risk, write its total exposure as $W=\sum_j w_j$. Its exposure-weighted historical rate is $\bar{X}=\sum_j w_jX_j/W$. Equivalently, if the loss amount in period $j$ is $L_j=w_jX_j$, then $\bar{X}=\sum_j L_j/W$. The two forms make clear why a larger-exposure period contributes more: its rate represents more exposure and therefore more of the risk’s total experience.

W=\sum_j w_j,\qquad \bar{X}=\frac{\sum_j w_jX_j}{W} =\frac{\sum_j L_j}{W}

A short justification follows from the model’s conditional variances. Given the risk’s underlying rate, $\operatorname{Var}(X_j\mid\theta)=\sigma^2/w_j$. Thus the inverse conditional variance of $X_j$ is $w_j/\sigma^2$, proportional to exposure. The inverse-variance weighted average of the annual rates is therefore \[ \frac{\sum_j (w_j/\sigma^2)X_j}{\sum_j w_j/\sigma^2} =\frac{\sum_j w_jX_j}{\sum_j w_j} =\bar{X}. \] So the exposure-weighted rate is also the natural precision-weighted summary under the stated model. This argument justifies the aggregation step; the credibility blend then accounts for uncertainty about the risk’s underlying rate.

Worked Example: Combining Three Years With Different Exposures

Method: First calculate the exposure-weighted rate, then use the Bühlmann-Straub credibility factor. A hypothetical fleet has the following adjusted annual pure loss costs. The collective mean is $\mu=\$100$ per exposure unit, $a=100$, and $\sigma^2=200$. Historical exposures are 1, 2, and 3 units, with annual rates of \$80, \$120, and \$100. Find the prospective rate and expected loss amount for 4 future exposure units. Assume no expenses, trend adjustment, or other premium component remains to be added.

Step 1: Total exposure. The historical total is $W=1+2+3=6$ exposure units.

Step 2: Exposure-weighted historical rate. The three period loss amounts are $1(80)=\$80$, $2(120)=\$240$, and $3(100)=\$300$. Total losses are $80+240+300=\$620$, so $\bar{X}=620/6=310/3=\$103.333\ldots$ per exposure unit.

Step 3: Credibility parameters. The Bühlmann-Straub parameter is $K=\sigma^2/a=200/100=2$. Therefore the credibility assigned to the risk’s experience is $Z=W/(W+K)=6/(6+2)=3/4$.

Step 4: Prospective rate. Blend the experience rate with the collective mean: \[ \widehat{\theta}=\frac{3}{4}\left(\frac{310}{3}\right)+\frac{1}{4}(100) =\frac{155}{2}+25=\frac{205}{2}=\$102.50. \] The result is \$102.50 per exposure unit.

Step 5: Future expected loss amount. For 4 exposure units, use the unrounded rate: $4(205/2)=410$. The expected loss amount is \$410.00, rounded to the nearest cent.

Check and common mistake avoided. The rate lies between the collective mean of \$100 and the weighted experience rate of about \$103.33, as a credibility blend should. An unweighted average of the three annual rates is \$100, but it incorrectly gives the one-unit and three-unit years equal influence. The exposure-weighted rate is the appropriate input.

The Bühlmann-Straub Credibility Premium

Under the model above, the standard Bühlmann-Straub credibility estimate for the risk’s expected rate is a blend of its exposure-weighted historical rate and the collective mean. Define $K=\sigma^2/a$, the ratio of process variance per exposure unit to variance among the underlying risk means. For positive total exposure, the credibility factor is $Z=W/(W+K)$.

Theorem: Under the Bühlmann-Straub assumptions, with known collective mean $\mu$, process variance $\sigma^2$, and variance of hypothetical means $a>0$, the credibility estimate of the risk’s expected rate is $\widehat{\theta}=Z\bar{X}+(1-Z)\mu$, where $W=\sum_j w_j$, $\bar{X}=\sum_j w_jX_j/W$, and $Z=W/(W+K)$ with $K=\sigma^2/a$. The factor increases with total exposure. This formula is not appropriate without qualification if the observations are not comparable or the assumed variance structure does not apply.

The formula has a useful interpretation. When total exposure is small relative to $K$, $Z$ is small, so the estimate stays close to the collective mean. As exposure grows, $Z$ approaches 1 and the risk’s own experience has more influence. The experience weight depends on total exposure, not simply the number of periods. A period with twice the exposure contributes twice as much to $\bar{X}$, while all periods together determine $W$ and therefore $Z$.

The result is a prospective pure loss rate. If the future exposure is $w_{\text{new}}$, the indicated expected loss amount is $w_{\text{new}}\widehat{\theta}$. Keep the rate and the amount distinct: the credibility estimate does not itself become a total dollar premium until it is multiplied by the relevant future exposure. If the exposure unit is earned vehicle-years, for instance, the rate is dollars per vehicle-year and future exposure must be expressed in vehicle-years.

Worked Example: A Small Exposure Period and a Large Exposure Period

Method: Weight each annual rate by its exposure, then calculate the credibility factor from total exposure. A hypothetical equipment account has exposure 0.5 units with a rate of \$140 and exposure 1.5 units with a rate of \$90. The collective mean is $\mu=\$105$, $a=50$, and $\sigma^2=100$. Determine the prospective rate and expected loss amount for 1.2 future units. Assume these are comparable pure loss costs and no expense load is included.

Step 1: Aggregate the experience. Total exposure is $W=0.5+1.5=2$. The period losses are $0.5(140)=\$70$ and $1.5(90)=\$135$. Total losses are $70+135=\$205$, so $\bar{X}=205/2=\$102.50$ per exposure unit.

Step 2: Find credibility. $K=\sigma^2/a=100/50=2$, so $Z=2/(2+2)=1/2$.

Step 3: Blend the rates. \[ \widehat{\theta}=\frac{1}{2}(102.50)+\frac{1}{2}(105) =51.25+52.50=\$103.75. \] The prospective expected loss rate is \$103.75 per exposure unit.

Step 4: Convert to a future amount. $1.2(103.75)=\$124.50$, so the expected loss amount is \$124.50, to the nearest cent.

Check and common mistake avoided. The experience rate is below the collective mean, so the credibility estimate must also be below \$105; \$103.75 satisfies this check. Giving the two annual rates equal weight would produce \$115, wrongly treating the half-unit period as equally informative as the 1.5-unit period.

Application Steps and Checks

A reliable calculation keeps aggregation, credibility, and prospective exposure as separate steps. In particular, do not multiply by future exposure before calculating the per-unit credibility rate, and do not confuse the total historical exposure $W$ with a count of years. The following sequence is useful when the problem supplies unequal exposures.

Confirm the units and comparability.
Identify each observation as a rate per stated exposure unit. Check that loss costs and exposures are measured on a consistent basis before combining periods.

Calculate total exposure and experience rate.
Compute $W=\sum_j w_j$ and $\bar{X}=\sum_j w_jX_j/W$. If total losses are given, divide their sum by total exposure.

Calculate credibility and blend.
Find $K=\sigma^2/a$, then $Z=W/(W+K)$, and calculate $Z\bar{X}+(1-Z)\mu$.

Apply future exposure and check.
Multiply the unrounded indicated rate by the stated future exposure. Confirm the rate lies between $\bar{X}$ and $\mu$, and that the total amount has the correct units.

Worked Example: Applying the Estimate to Future Exposure

Method: Use the full historical exposure to determine credibility, and preserve the exact rate through the final multiplication. A hypothetical specialty contractor has three comparable years with exposures 1, 2, and 3 units and respective rates of \$90, \$110, and \$130. The collective mean is $\mu=\$100$, $a=100$, and $\sigma^2=300$. Calculate the rate and expected loss amount for 2.5 future exposure units. Assume expenses and profit provisions are excluded.

Step 1: Weighted experience rate. Historical exposure is $W=1+2+3=6$. The period losses are $1(90)=\$90$, $2(110)=\$220$, and $3(130)=\$390$, for total losses of $\$700$. Thus $\bar{X}=700/6=350/3=\$116.666\ldots$.

Step 2: Credibility factor. $K=300/100=3$, so $Z=6/(6+3)=2/3$.

Step 3: Prospective rate. \[ \widehat{\theta}=\frac{2}{3}\left(\frac{350}{3}\right)+\frac{1}{3}(100) =\frac{700}{9}+\frac{300}{9} =\frac{1000}{9} =\$111.111\ldots \] per exposure unit.

Step 4: Prospective loss amount. Use the exact rate, not its displayed decimal approximation: $2.5(1000/9)=2500/9=\$277.777\ldots$. Rounded to the nearest cent, the expected loss amount is \$277.78.

Check and common mistake avoided. The estimate is between the collective mean of \$100 and the weighted experience rate of about \$116.67. Multiplying the rounded rate \$111.11 by 2.5 gives \$277.775, which can produce an inconsistent cent-rounding result. Carrying $1000/9$ through the calculation gives the correctly rounded amount.

Common Traps and Limits

Using an unweighted annual average. When exposures differ, $\sum_jX_j/n$ generally is not the appropriate experience rate. Use $\sum_jw_jX_j/W$.
Using the number of years in place of total exposure. The credibility factor uses $W=\sum_jw_j$. A history of three periods can have total exposure 1.5 or 30; those imply different credibility.
Reversing $a$ and $\sigma^2$. The parameter is $K=\sigma^2/a$. More process variation increases $K$ and reduces credibility at a fixed exposure, while more variation among risk means increases credibility.
Mixing rate and amount. The credibility blend is a rate per exposure unit. Multiply it by future exposure only when asked for a total expected loss amount.
Ignoring assumptions. The usual factor presumes the conditional variance decreases in inverse proportion to exposure and that observations are conditionally independent and comparable. A changed exposure definition or unadjusted cost level can undermine that interpretation.

For experience rating, the practical value of Bühlmann-Straub is that it uses a small-exposure period cautiously without discarding it, while allowing a larger exposure period to carry proportionally more weight. That makes it suitable for rating histories whose size changes over time. It remains a model-based estimate, not a guarantee of the next period’s loss cost; the credibility blend is an expected rate under its assumptions.

Key takeaway: With unequal exposures, first form the exposure-weighted rate $\bar{X}$, then use total exposure $W$ in $Z=W/(W+K)$, and finally blend $Z\bar{X}+(1-Z)\mu$. Multiply by future exposure only to obtain an expected loss amount.

Check Your Understanding

For each hypothetical risk, use the stated Bühlmann-Straub assumptions and report rates per exposure unit unless an amount is requested.

A risk has rates \$70 and \$100 at exposures 1 and 4. Let $\mu=\$100$, $a=100$, and $\sigma^2=500$. Find its credibility rate and expected loss amount for 3 future exposure units. (Answer: rate $\$97.00$; amount $\$291.00$)
A risk reports \$120 at exposure 0.5 and \$80 at exposure 1.5. Let $\mu=\$100$, $a=50$, and $\sigma^2=100$. Find the credibility rate. (Answer: $\$95.00$)
A risk has rates \$110, \$90, and \$120 at exposures 2, 1, and 3. Let $\mu=\$105$, $a=100$, and $\sigma^2=300$. Find the credibility rate and amount for 1.8 future exposure units. (Answer: rate $\$985/9$, or about $\$109.4444$; amount $\$197.00$)
A risk’s weighted experience rate is \$130, total exposure is 8, the collective mean is \$115, $a=25$, and $\sigma^2=100$. Find the credibility factor, prospective rate, and expected amount for 2.4 future units. (Answer: $Z=2/3$; rate $\$125.00$; amount $\$300.00$)
A risk has one observed period at rate \$80 with exposure 1, no exposure in a second period, and rate \$110 with exposure 2 in a third period. Let $\mu=\$120$, $a=100$, and $\sigma^2=300$. Find the rate and amount for 2 future units. (Answer: rate $\$110.00$; amount $\$220.00$)

What You'll Learn