Limiting Cases Of Bühlmann-Straub Credibility

Why the Limiting Cases Matter

The Bühlmann-Straub estimate balances a risk’s observed experience against the collective mean. Its credibility factor determines how much weight goes to each. Limiting cases make that balance easier to interpret: they show what the estimate should approach when a risk has very little or very much exposure, or when one source of variability becomes negligible.

These limits are also useful checks on calculations. With little information, an estimate should generally stay close to the collective mean. With abundant information, it should generally move toward the risk’s own experience. A variance component approaching zero can change that balance even when exposure is fixed.

Definition: In the Bühlmann-Straub model, risk $i$ has an unknown expected loss cost $\mu_i$, with collective mean $m=E[\mu_i]$ and between-risk variance $\tau^2=\operatorname{Var}(\mu_i)$. Conditional on $\mu_i$, an observation $X_{ij}$ has mean $\mu_i$ and variance $\sigma^2/w_{ij}$, where $w_{ij}$ is its exposure weight and $\sigma^2$ is the process variance per unit of exposure. Observations are conditionally independent across periods. For total exposure $W_i=\sum_j w_{ij}>0$, the risk’s weighted mean is $\bar X_i=\sum_j w_{ij}X_{ij}/W_i$. The credibility estimate is $\hat\mu_i=Z_i\bar X_i+(1-Z_i)m$, where $Z_i=W_i/(W_i+K)$ and $K=\sigma^2/\tau^2$, for $\tau^2>0$.

The usual interpretation assumes $\sigma^2>0$ and $\tau^2>0$. The credibility factor $Z_i$ lies between zero and one. The constant $K$ has exposure units: it compares the process variance with the between-risk variance. Exposure weights and the process-variance convention must use compatible units.

The estimate is a linear credibility estimate based on these moment assumptions; normality is not required for the credibility formula. A normal hierarchical model may be used to make a particular example fully specified, but the limiting results below follow from the formula and its variance parameters.

How Exposure Changes the Estimate

Hold the variance components fixed and write $W$ for a risk’s total exposure. Then $Z=W/(W+K)$. As exposure decreases toward zero, the credibility assigned to the risk’s own mean decreases toward zero. As exposure grows without bound, that credibility increases toward one.

\lim_{W\to 0^+} Z=0,\qquad \lim_{W\to 0^+}\hat\mu=m

\lim_{W\to\infty} Z=1,\qquad \lim_{W\to\infty}\hat\mu=\bar X

The small-exposure limit is a statement about the estimate, not about the reliability of the observed mean. In fact, for very small positive exposure, $\bar X$ may be highly variable. The credibility weight on it is small, which pulls the estimate toward $m$. At exactly zero exposure there are no observations for that risk, so $\bar X$ is undefined; the estimate is the collective mean $m$, not a calculation using an observed risk mean.

For fixed positive $K$, the small-exposure approximation follows directly from the factor:

Z=\frac{W}{W+K} =\frac{W/K}{1+W/K} \approx \frac{W}{K}\quad\text{when }W/K\text{ is small}.

Thus the departure from the collective mean is approximately $(W/K)(\bar X-m)$ for sufficiently small $W/K$. This is an approximation, not an exact formula. At the other extreme, $1-Z=K/(W+K)$ tends to zero, so the estimate’s remaining weight on the collective mean vanishes.

Worked Example: A Risk With Very Little Exposure

A hypothetical insurer models annual loss cost under a normal hierarchical model: $\mu_i\sim N(100,100)$, and $X_{ij}\mid\mu_i\sim N(\mu_i,900/w_{ij})$, conditionally independently. Thus $m=100$, $\tau^2=100$, and $\sigma^2=900$. One risk has total exposure $W=0.25$ and observed weighted mean $\bar X=80$. Calculate its credibility estimate. The values are illustrative, and no interest or mortality assumptions are involved.

Method. Calculate $K$, then $Z$, then combine the risk mean and collective mean. This method avoids treating a small exposure as if it justified full credibility.

First, $K=\sigma^2/\tau^2=900/100=9$. Therefore,

Z=\frac{0.25}{0.25+9}=\frac{0.25}{9.25} =\frac{1}{37}=0.027027\ldots

The risk receives about $2.7027\%$ credibility. The estimate is

\hat\mu_i =\frac{1}{37}(80)+\frac{36}{37}(100) =\frac{80+3600}{37} =\frac{3680}{37} =99.459459\ldots

Rounded only at the final step, the estimate is $99.46. It lies close to the collective mean of $100, as expected when exposure is small. As a check, its distance below the collective mean is about $0.54, much smaller than the $20 difference between the observed mean and the collective mean.

Worked Example: Large Exposure and a High Observed Mean

A hypothetical commercial coverage portfolio uses a normal hierarchical model with $\mu_i\sim N(100,64)$ and $X_{ij}\mid\mu_i\sim N(\mu_i,576/w_{ij})$, conditionally independently. Thus $m=100$, $\tau^2=64$, and $\sigma^2=576$. A risk has total exposure $W=90$ and weighted mean $\bar X=120$. Find the estimate, and compare its distance from the observed mean with the distance that would remain if credibility were complete.

Method. Use the exact credibility factor, then evaluate the estimate and its distance from the risk mean. This avoids assuming that “large” exposure automatically means $Z=1$.

The structural constant is $K=576/64=9$. The credibility factor is

Z=\frac{90}{90+9}=\frac{90}{99}=\frac{10}{11}.

Therefore,

\hat\mu_i =\frac{10}{11}(120)+\frac{1}{11}(100) =\frac{1200+100}{11} =\frac{1300}{11} =118.181818\ldots

The estimate is $118.18, rounded to the nearest cent. Its distance below the observed mean is $120-118.181818\ldots=1.818182\ldots$, or about $1.82. Complete credibility would give $120, so a residual difference remains because $Z=10/11$, not exactly one. The estimate lies between $100 and $120, which checks the weighted-average calculation.

How the Variance Components Change the Limits

At fixed positive exposure, $Z=W/(W+\sigma^2/\tau^2)$ also shows the effect of each variance component. Lower process variance makes the risk’s observed experience more informative and increases $Z$. Lower between-risk variance makes risks more alike under the model and decreases $Z$. The corresponding boundary results are:

\lim_{\sigma^2\to 0^+}Z=1 \quad\text{if }\tau^2>0,\ W>0; \qquad \lim_{\tau^2\to 0^+}Z=0 \quad\text{if }\sigma^2>0,\ W<\infty.

In the first limit, observations have vanishing process noise, so the risk’s observed mean identifies its risk mean in the limit. If $\sigma^2=0$ exactly and $W>0$, the conditional observations equal $\mu_i$ almost surely, and the estimate is $\bar X_i=\mu_i$. In the second limit, the risk means have no between-risk variation in the limit; the collective mean is then the appropriate estimate for every risk.

These boundary statements hold with the other quantities fixed as specified. If both variance components approach zero together, the ratio $\sigma^2/\tau^2$ need not approach either zero or infinity. The credibility factor then depends on how quickly the two components approach zero relative to each other. If both are exactly zero, the formula for $K$ is undefined; the model is degenerate and the estimate must be interpreted from the assumptions, not by substituting $0/0$.

Identify the changing quantity.
Decide whether the limit concerns exposure $W$, process variance $\sigma^2$, or between-risk variance $\tau^2$. State which other quantities are held fixed.

Take the limit of the credibility factor.
Use $Z=W/(W+K)$, with $K=\sigma^2/\tau^2$ when $\tau^2>0$. Do not substitute into $K$ when its denominator is zero.

Translate the factor into an estimate.
If $Z\to0$, the estimate approaches $m$; if $Z\to1$, it approaches $\bar X$, provided the risk has positive exposure.

Worked Example: Comparing the Two Variance Boundaries

Consider two separate hypothetical portfolios, each under a normal hierarchical model with conditional independence and the Bühlmann-Straub variance structure. In both, the risk has $W=4$, $\bar X=130$, and the collective mean is $m=100$. In the first portfolio, $\tau^2=64$ and process variance $\sigma^2$ decreases toward zero. In the second, $\sigma^2=36$ and between-risk variance $\tau^2$ decreases toward zero. Calculate the limiting estimate in each case. These are loss-cost models; interest and mortality assumptions are not applicable.

Method. Evaluate each limit with the other variance component fixed. This avoids confusing which variance component increases credibility.

First portfolio: process variance tends to zero. For successive values $\sigma^2=16,4,1$, the corresponding $K$ values are $16/64=0.25$, $4/64=0.0625$, and $1/64=0.015625$. The credibility factors are

\frac{4}{4.25}=0.941176\ldots,\qquad \frac{4}{4.0625}=0.984615\ldots,\qquad \frac{4}{4.015625}=0.996109\ldots.

As $\sigma^2\to0^+$, $K\to0$ and $Z\to1$. Thus $\hat\mu_i\to\bar X=130$. At the exact boundary $\sigma^2=0$, the observations equal the risk mean under the model, so the estimate is $130.

Second portfolio: between-risk variance tends to zero. For $\tau^2=9,1$, the $K$ values are $36/9=4$ and $36/1=36$. The corresponding factors are

\frac{4}{4+4}=0.5,\qquad \frac{4}{4+36}=0.1.

As $\tau^2\to0^+$, $K\to\infty$ and $Z\to0$. The estimate therefore approaches $m=100$. The two limits are consistent with the model: removing process noise makes the individual experience informative, while removing between-risk variation removes the modeled distinction among risk means.

Common Exam Trap and Interpretation

A frequent error is to reverse the roles of the two variance components. For example, with $W=4$, $\sigma^2=36$, and $\tau^2=9$, the correct $K$ is $36/9=4$, giving $Z=4/(4+4)=0.5$. Reversing the ratio gives $K=9/36=0.25$ and the incorrect factor $4/(4+0.25)=0.9412$. The correction is to use process variance in the numerator of $K$ and between-risk variance in its denominator. Process noise reduces credibility; between-risk variation increases it.

Another trap is to report an individual weighted mean at zero exposure. With $W=0$, the sum of exposure weights is zero, so $\bar X_i$ is not defined. The no-experience estimate is $m$, rather than a weighted average involving a nonexistent risk mean. Also distinguish a limit from an exact equality: for finite positive $W$, $K>0$ gives $Z<1$, even if $W$ is large.

These boundary checks matter in practical ratemaking because they test whether a credibility calculation behaves coherently before it is applied to a risk or class. They are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus. For any calculation, confirm the variance definitions, keep exposure units consistent, and state whether the result is an exact boundary value or a limit.

Key takeaway: With positive variance components, little exposure gives an estimate near the collective mean and abundant exposure gives an estimate near the risk’s own experience. At fixed positive exposure, vanishing process variance sends credibility to one; vanishing between-risk variance sends it to zero. Handle zero exposure and degenerate variance cases separately rather than forcing them into a formula with an undefined term.

Check Your Understanding

For each problem, use the stated Bühlmann-Straub quantities and give the requested estimate or limit.

A hypothetical portfolio has $m=100$, $K=12$, total exposure $W=3$, and risk mean $\bar X=76$. Find the credibility estimate. (Answer: $95.20)
A risk has $m=105$, $K=10$, $W=40$, and $\bar X=115$. Find $Z$ and the credibility estimate. (Answer: $Z=0.8000$; estimate $113.00)
At fixed $W=5$, $\tau^2=25$, $m=120$, and $\bar X=132$, what estimate does the model approach as $\sigma^2\to0^+$? Assume positive exposure throughout. (Answer: $132)
At fixed finite $W=8$, $\sigma^2=32$, $m=125$, and $\bar X=150$, what estimate does the model approach as $\tau^2\to0^+$? (Answer: $125)
A risk has $m=110$, $K=6$, $W=0.6$, and $\bar X=90$. Calculate its credibility estimate and explain why it is close to the collective mean. (Answer: $108.18; $Z=0.0909$, so only about 9.09% weight is assigned to the risk mean.)

What You'll Learn