Ordinary Bühlmann As An Equal-Exposure Special Case

Why Equal Exposure Gives Ordinary Bühlmann Credibility

Bühlmann-Straub credibility allows observations to carry different exposure weights: an observation based on more exposure has lower process variance and receives more weight in the experience average. When every period has the same exposure, those weights are equal, so the exposure-weighted average becomes the ordinary arithmetic average. The credibility estimate then has the familiar ordinary Bühlmann form, provided its process-variance parameter is expressed on the same exposure scale.

This relationship is useful both for calculation and for interpretation. It shows that ordinary Bühlmann credibility is not a separate rule that conflicts with Bühlmann-Straub; it is the equal-exposure case of the latter. The key is to distinguish the number of observations from their total exposure. This tutorial develops that equivalence and is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Definition: Let $\Theta$ denote a risk parameter and let $\mu(\Theta)$ be the risk’s conditional mean loss rate. In the Bühlmann-Straub model, observation $X_j$ has exposure weight $w_j>0$, with $E[X_j\mid\Theta]=\mu(\Theta)$ and $\operatorname{Var}(X_j\mid\Theta)=\sigma^2(\Theta)/w_j$. Conditional on $\Theta$, observations are assumed independent. Define the expected process variance per unit exposure as $a=E[\sigma^2(\Theta)]$, the variance of hypothetical means as $v=\operatorname{Var}(\mu(\Theta))$, and the collective mean as $\mu_0=E[\mu(\Theta)]$. Assume $a>0$ and $v>0$.

The total exposure is $W=\sum_{j=1}^n w_j$, and the exposure-weighted experience mean is $\bar X_w=\sum_{j=1}^n w_jX_j/W$. The Bühlmann-Straub credibility estimate of the risk’s mean is a weighted average of that experience mean and the collective mean. Its factor depends on total exposure, not simply on the number of periods.

\widehat{\mu}_{BS} = Z\bar X_w+(1-Z)\mu_0, \qquad Z=\frac{W}{W+K}, \qquad K=\frac{a}{v}.

Here $K$ has units of exposure: $a$ is process variance per unit exposure and $v$ is variance in the risk means. If the weights are equal to a common value $c$, then $W=nc$ and $\bar X_w=\bar X$, the ordinary arithmetic mean. Therefore,

Z=\frac{nc}{nc+a/v}.

For unit exposure in every period, $c=1$, so $Z=n/(n+a/v)$, the ordinary Bühlmann factor. For a common exposure other than one, the ordinary form still applies, but the process variance for one observation must be converted to that observation’s exposure scale.

Why the Factor Has This Form

The credibility factor can be justified by finding the best linear predictor based on the experience mean. Conditional independence and the model variance assumption give the conditional variance of the weighted mean:

\operatorname{Var}(\bar X_w\mid\Theta) = \sum_{j=1}^n \left(\frac{w_j}{W}\right)^2 \frac{\sigma^2(\Theta)}{w_j} = \frac{\sigma^2(\Theta)}{W}.

Taking expectations over $\Theta$, the expected process variance of $\bar X_w$ is $a/W$. Also, $E[\bar X_w\mid\Theta]=\mu(\Theta)$, so the variance in the hypothetical mean is $v$. Consider a linear estimate $\mu_0+Z(\bar X_w-\mu_0)$. Its mean squared error is the sum of the variance contribution from deviations in risk means and the expected process variance contribution:

\operatorname{MSE}(Z) = (1-Z)^2v+Z^2\frac{a}{W}.

Differentiating with respect to $Z$ and setting the derivative equal to zero gives

-2(1-Z)v+2Z\frac{a}{W}=0, \qquad v=Z\left(v+\frac{a}{W}\right), \qquad Z=\frac{v}{v+a/W} =\frac{W}{W+a/v}.

Thus the Bühlmann-Straub factor is the coefficient that minimizes mean squared error among linear estimates of this form. This derivation does not say that the estimate is always the exact Bayesian posterior mean; that stronger conclusion requires additional distributional assumptions. It also depends on conditional independence and the stated inverse-exposure variance structure.

Matching the Ordinary Bühlmann Parameters

Ordinary Bühlmann credibility is commonly stated for conditionally independent observations with a common conditional mean and a common process variance per observation. Its factor is $n/(n+K_B)$, where $K_B$ is the expected process variance per observation divided by the variance in hypothetical means. If each Bühlmann-Straub observation has the same exposure $c$, its conditional variance is $\sigma^2(\Theta)/c$, not $\sigma^2(\Theta)$. Consequently, the ordinary Bühlmann process-variance parameter is $a/c$, and

K_B=\frac{a/c}{v}=\frac{K}{c}, \qquad Z_B=\frac{n}{n+K_B} =\frac{n}{n+K/c} =\frac{nc}{nc+K} =Z.

This equality makes the two parameterizations agree exactly. If each period has unit exposure, then $c=1$, $K_B=K$, and the ordinary formula can be used directly with $n$. If each period has exposure $c\ne1$, using $n/(n+K)$ while leaving $K=a/v$ unchanged mixes incompatible units and generally gives the wrong factor.

Worked Example: Four Unit-Exposure Periods

A hypothetical risk has four annual observations, each with unit exposure. The observations are $84$, $96$, $108$, and $132$ dollars per exposure unit. The collective mean is $\mu_0=\$95$, the expected process variance is $a=400$ dollars squared per unit exposure, and the variance in hypothetical means is $v=100$ dollars squared. Calculate the Bühlmann-Straub estimate and show its ordinary Bühlmann equivalent.

Method and assumptions. Use the equal-exposure Bühlmann-Straub formula. The observations are conditionally independent, have a common conditional mean for the risk, and have conditional variance $\sigma^2(\Theta)/w_j$. Here each $w_j=1$, so the total exposure is $W=4$, and the experience average is the arithmetic mean.

First calculate the experience average:

\bar X=\frac{84+96+108+132}{4} =\frac{420}{4} =\$105.

Next calculate the Bühlmann-Straub parameter and factor:

K=\frac{a}{v}=\frac{400}{100}=4, \qquad Z=\frac{W}{W+K}=\frac{4}{4+4}=\frac{1}{2}.

The credibility estimate is

\widehat{\mu}_{BS} = \frac{1}{2}(105)+\frac{1}{2}(95) =\$100.

Ordinary Bühlmann check. Since exposure per observation is one, the process variance per observation is $a/c=400$. Thus $K_B=400/100=4$, and the ordinary factor is $4/(4+4)=1/2$, the same factor. The estimate checks: it is halfway between the experience average of $\$105$ and the collective mean of $\$95$.

Worked Example: Equal Exposures of Two Units

A hypothetical portfolio has three period loss-rate observations of $\$88$, $\$100$, and $\$112$ per exposure unit. Each observation has exposure $c=2$. The collective mean is $\$94$, expected process variance is $a=240$ dollars squared per unit exposure, and variance in hypothetical means is $v=60$ dollars squared. Find the credibility estimate. Then identify the error in using the ordinary factor with $K=a/v$ unchanged.

Method and assumptions. Apply Bühlmann-Straub with equal weights, so its weighted experience mean is the arithmetic mean. The observations are conditionally independent and each has conditional variance $\sigma^2(\Theta)/2$. Total exposure, not the count of periods, determines the factor.

The average of the three observations is

\bar X=\frac{88+100+112}{3} =\frac{300}{3} =\$100.

The variance ratio is $K=a/v=240/60=4$ exposure units. Total exposure is $W=3(2)=6$, giving

Z=\frac{6}{6+4}=0.6.

Therefore,

\widehat{\mu}_{BS} =0.6(100)+0.4(94) =60+37.60 =\$97.60.

Ordinary Bühlmann check. The process variance per observation is $a/c=240/2=120$. Thus $K_B=120/60=2$, and $Z_B=3/(3+2)=0.6$, matching the Bühlmann-Straub factor.

Common mistake avoided. Using $3/(3+4)=3/7$ treats $K=4$ exposure units as though it were already the ordinary $K_B$ per observation. That gives the incorrect estimate $94+(3/7)(100-94)=\$96.57$ to the nearest cent. The correct conversion is $K_B=K/c=4/2=2$, which gives $\$97.60$.

Known Period Effects Must Be Removed First

The equal-exposure reduction assumes the observations estimate the same risk mean. If period means are known to differ by specified offsets, the unadjusted arithmetic average need not estimate that common mean. When the offsets are additive and known, subtract each offset from its observation before applying the equal-exposure formula. This is an adjustment for a modeled period effect, not a change to the exposure weights.

Worked Example: Equal Exposure With Known Period Offsets

A hypothetical risk has two observations, each with unit exposure. Their conditional means are $\mu(\Theta)-\$10$ in the first period and $\mu(\Theta)+\$10$ in the second. The observed values are $\$80$ and $\$120$. The collective mean of $\mu(\Theta)$ is $\$90$, expected process variance is $a=100$ dollars squared per unit exposure, and variance in hypothetical means is $v=25$ dollars squared. Estimate $\mu(\Theta)$.

Method and assumptions. The period offsets are known and additive. Subtract each period’s offset from its observation to obtain adjusted observations with common conditional mean $\mu(\Theta)$. Both exposures are explicitly one unit, so total exposure is two units.

The first offset is $-\$10$, so its adjusted value is $80-(-10)=\$90$. The second offset is $+\$10$, so its adjusted value is $120-10=\$110$. Their adjusted mean is

\bar X_{\text{adj}}=\frac{90+110}{2}=\$100.

The credibility parameter and factor are

K=\frac{100}{25}=4, \qquad Z=\frac{W}{W+K}=\frac{2}{2+4}=\frac{1}{3}.

Thus the estimate is

\widehat{\mu} =90+\frac{1}{3}(100-90) =90+\frac{10}{3} =\$93.33

rounded to the nearest cent. The result is checked by noting that it lies one-third of the way from the collective mean of $\$90$ toward the adjusted experience mean of $\$100$. The unit exposures matter: if they had a different common size, total exposure and the ordinary per-observation process variance would need to reflect that size.

When the Reduction Does Not Apply

The equivalence is specific to equal exposure weights and a model in which the observations, after any required adjustments, share a common conditional mean. It is not enough that observations are recorded in the same calendar periods or that there are the same number of observations for each risk. Check the model and units before selecting the formula.

Unequal weights: If $w_j$ differ, the exposure-weighted mean is generally not the ordinary arithmetic mean. Use $W=\sum w_j$ and $\bar X_w=\sum w_jX_j/W$, rather than substituting $n$ for $W$.
Unadjusted period means: If period $j$ has a known additive offset, an unadjusted average does not generally target the underlying common risk mean. Adjust first when the model specifies such offsets.
Wrong process-variance scale: With common exposure $c$, ordinary Bühlmann uses process variance $a/c$ per observation. Using $a$ as the per-observation variance is correct only when $c=1$.
Dependence or a different variance pattern: Conditional dependence changes the variance of the experience mean. Likewise, if conditional variance is not inversely proportional to exposure, the stated Bühlmann-Straub factor need not apply.
Zero exposure: A zero-weight period contributes no information to total exposure or the weighted mean; it is not an equal positive-exposure observation.

A practical check is to calculate the factor in both parameterizations. For equal weights $c$, the Bühlmann-Straub calculation uses $W=nc$ and $K=a/v$, while the ordinary calculation uses $n$ and $K_B=a/(cv)$. The two factors must agree:

\frac{nc}{nc+a/v} = \frac{n}{n+a/(cv)}.

If they do not, the likely issue is a mismatch in exposure units, a failure to adjust period effects, or use of an ordinary average when weights are unequal.

Key takeaway: With equal exposure $c$ in every period, Bühlmann-Straub reduces to ordinary Bühlmann credibility because the weighted mean becomes an arithmetic mean. Preserve the variance scale: use total exposure $nc$ in Bühlmann-Straub, or use the per-observation process variance $a/c$ in ordinary Bühlmann.

Check Your Understanding

For each problem, identify the relevant exposure scale before calculating the estimate.

A hypothetical risk has five unit-exposure observations with average loss rate $\$82$ per unit. The collective mean is $\$70$, $a=300$ dollars squared per unit exposure, and $v=100$ dollars squared. Find $Z$ and the credibility estimate. (Answer: $Z=5/8=0.625$; estimate $\$77.50$)
Three hypothetical observations each have exposure $2$, with values $\$88$, $\$100$, and $\$112$. The collective mean is $\$94$, $a=240$ dollars squared per unit exposure, and $v=60$ dollars squared. Find the estimate and the ordinary Bühlmann $K_B$. (Answer: estimate $\$97.60$; $K_B=2$)
A risk has unequal exposure weights $1,2,3$ and observed rates $\$60,\$90,\$100$, respectively. The collective mean is $\$84$, $a=180$ dollars squared per unit exposure, and $v=60$ dollars squared. Calculate the weighted average and Bühlmann-Straub estimate; explain why the equal-exposure ordinary formula is not applicable. (Answer: weighted average $\$90$; $W=6$, $Z=2/3$, estimate $\$88$; the weights are unequal, so the experience average is not an ordinary arithmetic mean.)
Four unit-exposure observations are $\$74,\$102,\$118,\$114$. Their known additive period offsets are $-\$6,-\$2,+\$2,+\$6$, respectively. The collective mean is $\$94$, $a=80$ dollars squared per unit exposure, and $v=20$ dollars squared. Find the adjusted average and credibility estimate. (Answer: adjusted values $\$80,\$104,\$116,\$108$; adjusted average $\$102$; $Z=1/2$; estimate $\$98$)
Six hypothetical observations each have exposure $0.5$, with experience average $\$40$ per unit. The collective mean is $\$52$, $a=150$ dollars squared per unit exposure, and $v=30$ dollars squared. Calculate the Bühlmann-Straub factor and verify it using ordinary Bühlmann parameters. (Answer: $W=3$, $K=5$, $Z=3/8=0.375$; $K_B=10$, $Z_B=6/(6+10)=3/8$; estimate $\$47.50$)

What You'll Learn