Interpreting Uncertainty in Class Credibility Indications

What a Credibility Indication Does—and Does Not—Tell You

A credibility-weighted class cost is a point estimate: it combines a class’s observed experience with a complement, such as a portfolio mean. That blend answers, “What cost should this class indication use under the selected credibility method?” It does not, by itself, answer, “How uncertain is the class’s underlying cost?” A useful interpretation keeps those questions separate.

For a class with limited experience, the indication is pulled toward the complement. More experience generally gives the class data more weight. But the amount of uncertainty also depends on the assumed variability of individual observations and the extent to which underlying class means vary. Two indications can be different without the evidence establishing that the underlying class costs are meaningfully different.

Definition: In the equal-exposure Bühlmann model, let $\Theta$ be the underlying mean cost for a class, with $E[\Theta]=\mu$ and $\operatorname{Var}(\Theta)=\tau^2$. Given $\Theta$, the $n$ observed costs $X_1,\ldots,X_n$ are conditionally independent, each with $E[X_j\mid\Theta]=\Theta$ and $\operatorname{Var}(X_j\mid\Theta)=\sigma^2$. The class sample mean is $\bar X$, the credibility constant is $K=\sigma^2/\tau^2$, and the Bühlmann estimate is $\hat{\Theta}=\mu+Z(\bar X-\mu)$, where $Z=n/(n+K)$. Here $\mu$ is the complement, $\sigma^2$ is process variance, and $\tau^2$ is variance of hypothetical class means around the complement.

This model describes a hypothetical population of classes. It assumes a common process variance and equal-sized, independent observations within the class. If exposures differ, dependence is present, or class-specific process variances are important, this equal-exposure formulation is not automatically appropriate. The variance parameters are model inputs or estimates; they are not revealed merely by calculating $Z$.

Point Estimate Versus Uncertainty

The point estimate $\hat{\Theta}$ summarizes the modeled class mean after blending. Under only the Bühlmann mean-and-variance assumptions, a standard measure of its accuracy is the mean squared error for estimating $\Theta$, averaged over the modeled class population and its observations:

E[(\Theta-\hat{\Theta})^2] =(1-Z)^2\tau^2+Z^2\frac{\sigma^2}{n} =\tau^2(1-Z).

The first term reflects the error from not fully following the class’s underlying mean; the second reflects sampling noise in its observed mean. The last equality uses $K=\sigma^2/\tau^2$ and $Z=n/(n+K)$. The square root of this mean squared error is an average error scale across the model, not automatically a standard error conditional on one particular class’s observed data.

A distributional assumption is needed to turn a point estimate and a variance into a probability interval. For example, suppose $\Theta\sim N(\mu,\tau^2)$, and $X_j\mid\Theta\sim N(\Theta,\sigma^2)$, independently. Then the posterior distribution of $\Theta$ given the observed costs is normal, with mean $\hat{\Theta}$ and variance $\tau^2(1-Z)$. A 95% posterior interval is $\hat{\Theta}\pm1.96\sqrt{\tau^2(1-Z)}$. This is an interval for the underlying class mean under the stated model, not a guarantee that an individual future cost will fall within it.

Without the normal assumptions, the Bühlmann estimate and its mean squared error can still be useful if the required moments exist, but that variance alone does not specify a 95% interval. A normal interval may be a convenient approximation, but it should be identified as such rather than presented as a consequence of credibility weighting alone.

Why the Mean Squared Error Formula Holds

The decomposition follows from the model’s conditional mean and variance. Write the estimation error as $\Theta-\hat{\Theta}=(1-Z)(\Theta-\mu)-Z(\bar X-\Theta)$. The two terms have zero covariance: conditional on $\Theta$, the sampling error $\bar X-\Theta$ has mean zero. The underlying-mean term has variance $\tau^2$, and the sampling error has unconditional variance $\sigma^2/n$. Therefore:

E[(\Theta-\hat{\Theta})^2] =(1-Z)^2\tau^2+Z^2\frac{\sigma^2}{n}.

Substitute $K=\sigma^2/\tau^2$, so that $\sigma^2/n=\tau^2K/n$. Then:

E[(\Theta-\hat{\Theta})^2] =\tau^2\left((1-Z)^2+Z^2\frac{K}{n}\right) =\tau^2\left((1-Z)^2+Z(1-Z)\right) =\tau^2(1-Z).

The middle equality uses $ZK/n=1-Z$, which follows directly from $Z=n/(n+K)$. This derivation explains why more credibility reduces this particular average estimation error when the model parameters are fixed. It does not establish that the parameters themselves are known without error.

Worked Examples

Worked Example: Turning an Indication Into an Uncertainty Statement

A hypothetical class has 16 equal-exposure observations with an average cost of $750. The complement is $600. Assume $\Theta\sim N(600,40000)$ and $X_j\mid\Theta\sim N(\Theta,160000)$, independently. Thus the between-class variance is $\tau^2=40000$ dollars squared and the process variance is $\sigma^2=160000$ dollars squared per observation. There is no discounting or mortality assumption: these are costs measured on a common basis. Calculate the credibility indication and a 95% posterior interval for $\Theta$.

Method. Use the normal equal-exposure Bühlmann model. First, $K=\sigma^2/\tau^2=160000/40000=4$. Thus $Z=n/(n+K)=16/(16+4)=0.8$.

The indication is $\hat{\Theta}=\mu+Z(\bar X-\mu)=600+0.8(750-600)=600+120=\$720.00$.

For the underlying class mean, the posterior variance is $\tau^2(1-Z)=40000(0.2)=8000$ dollars squared. The posterior standard deviation is $\sqrt{8000}=\$89.4427$. The 95% interval is $720\pm1.96(89.4427)=720\pm175.308$, or approximately $(\$544.69,\ \$895.31)$, rounded to cents.

Check. The class observation is above the complement, so the indication is above $600, but is below the raw average of $750 because $Z<1$. The interval is for $\Theta$, under the stated normal model. The common mistake is to call $720 certain, or to describe this interval as an interval for every future cost.

Worked Example: Different Class Volumes, Different Indications

Two hypothetical classes have the same complement, $1,000, and each has an observed average cost of $1,200. Class A has 4 observations and Class B has 16. For both, assume independent class means with distribution $N(1000,10000)$ and, conditional on each class mean, independent normal costs with variance $40000$ per observation. There is no discounting or mortality assumption. Calculate the two indications and interpret their uncertainty, assuming the class means are independent.

Method. For both classes, $K=40000/10000=4$. For A, $Z_A=4/(4+4)=0.5$, so $\hat{\Theta}_A=1000+0.5(1200-1000)=\$1,100.00$. For B, $Z_B=16/(16+4)=0.8$, so $\hat{\Theta}_B=1000+0.8(1200-1000)=\$1,160.00$.

Under the normal model, A’s posterior variance is $10000(1-0.5)=5000$, giving a standard deviation of $\$70.7107$. Its 95% interval is $1100\pm1.96(70.7107)$, or $(\$961.41,\ \$1,238.59)$. B’s posterior variance is $10000(1-0.8)=2000$, giving a standard deviation of $\$44.7214$. Its interval is $1160\pm1.96(44.7214)$, or $(\$1,072.35,\ \$1,247.65)$, rounded to cents.

The indicated costs differ by $60, but the intervals overlap. More specifically, under the stated independent normal model, the posterior difference $\Theta_B-\Theta_A$ has mean $60 and variance $2000+5000=7000$. Its standard deviation is $\sqrt{7000}=\$83.6660$, so $P(\Theta_B>\Theta_A\mid\text{data})=\Phi(60/83.6660)\approx0.7633$. The result leans toward B having the higher underlying mean, but is not certainty. The common mistake is to treat the higher-volume class’s higher indication as proof of a real cost difference.

Worked Example: How an Assumed Process Variance Changes Interpretation

A hypothetical class has 10 observations with average cost $650 and complement $500. Compare two model assumptions, each with $\Theta\sim N(500,10000)$. In Model L, $X_j\mid\Theta\sim N(\Theta,10000)$; in Model H, $X_j\mid\Theta\sim N(\Theta,90000)$, independently. There is no discounting or mortality assumption. Calculate each indication and its 95% posterior interval for $\Theta$.

Method. In Model L, $K=10000/10000=1$, so $Z_L=10/(10+1)=10/11$. The indication is $500+(10/11)(150)=\$636.36$, rounded to cents. The posterior variance is $10000(1-10/11)=10000/11=\$909.0909$ squared, and the standard deviation is $\$30.1511$. The interval is $636.3636\pm1.96(30.1511)$, or approximately $(\$577.27,\ \$695.46)$.

In Model H, $K=90000/10000=9$, so $Z_H=10/(10+9)=10/19$. The indication is $500+(10/19)(150)=\$578.95$. The posterior variance is $10000(1-10/19)=90000/19=\$4,736.8421$ squared, and the standard deviation is $\$68.8247$. The interval is $578.9474\pm1.96(68.8247)$, or approximately $(\$444.05,\ \$713.85)$, rounded to cents.

Both models use the same observed average, but Model H assumes more process noise. It therefore gives the class experience less weight and leaves more uncertainty about the class mean. The common mistake is to interpret the difference between the indications as a data-driven discovery: here it arises from changing a model assumption.

Common Mistakes and Practical Interpretation

A credibility indication is most informative when it is reported with its complement, credibility factor, experience volume, and material variance assumptions. If a normal interval is used, state that it is conditional on the normal model and identify what the interval concerns. These details help a reviewer distinguish a small indication difference caused by low credibility from evidence that class means are separated.

Using process variance as the uncertainty of the class mean. The quantity $\sigma^2$ describes observation variability conditional on $\Theta$. Under the normal model, uncertainty about $\Theta$ after observing the data is $\tau^2(1-Z)$, not $\sigma^2$.
Reporting a variance without its assumptions. The Bühlmann mean squared error is a model-based average. It is not automatically a class-specific confidence interval or a posterior interval.
Comparing indications as if they were exact costs. A difference between point estimates does not alone establish a difference between underlying class means. Volume and variance assumptions affect how much weight the data receive and how uncertain the estimates remain.
Ignoring model-parameter uncertainty. The calculations treat $\mu,\sigma^2,\tau^2$ as given. If they are estimated from limited data, that additional estimation uncertainty is not included in the formulas above.

Key takeaway: A credibility-weighted indication is a model-based point estimate, not a complete description of uncertainty. Interpret it alongside class volume, process and between-class variance assumptions, and—only when justified—the distributional assumptions used for an interval.

Check Your Understanding

Use the equal-exposure model unless a question says otherwise. Treat any normal interval as conditional on an explicitly stated normal model.

A hypothetical class has $n=9$, complement $\mu=\$500$, observed mean $\bar X=\$650$, and $K=3$. Find $Z$ and the credibility indication. (Answer: $Z=0.7500$; indication $=\$612.50$.)
Assume $\Theta\sim N(700,6400)$, $X_j\mid\Theta\sim N(\Theta,19200)$, and $n=12$. Find the posterior variance and standard deviation of $\Theta$ after observing the class mean. (Answer: $K=3$, $Z=0.8000$; posterior variance $=1280$ dollars squared and standard deviation $\approx\$35.78$.)
Two hypothetical classes share complement $800, $K=5$, and observed means $900. Their volumes are 5 and 20. Find both indications and explain the difference. (Answer: $Z_1=0.5000$, indication $=\$850.00$; $Z_2=0.8000$, indication $=\$880.00$. The larger volume gives the same observed deviation more weight.)
Under a normal model, let $\tau^2=2500$, $n=10$, and $\sigma^2=10000$. Find the posterior standard deviation of the class mean. (Answer: $K=4$, $Z=10/14=0.7143$; posterior variance $=2500(4/14)=714.2857$ dollars squared and standard deviation $\approx\$26.73$.)
A class indication is $720, and a report gives only the Bühlmann mean squared error of $8,000 dollars squared. Can one infer a 95% interval from these two values alone? (Answer: No. The mean squared error alone does not specify a probability distribution; a normal or other distributional assumption is needed to form a probability interval.)

What You'll Learn