Pooling Comparable Classes for Credibility

Why Pool Comparable Classes?

A class with little experience may have a volatile observed loss cost. Combining its experience with that of a related class can increase the information available for a credibility estimate. But pooling is not automatically beneficial: if the classes have different underlying expected costs, combining them can make the estimate more stable while making it less appropriate for either class.

The decision therefore has two parts. First, determine whether the classes are comparable on the rating basis being used. Then, if pooling is defensible, combine their experience using exposure weights and apply credibility to the pooled result. A numerical difference between class averages is not by itself proof that their underlying risks differ; random variation can produce differences even when the classes share the same expected cost.

Definition: Pooling comparable classes means treating the classes as sharing a common underlying expected loss cost for the purpose of estimating that cost, and combining their adjusted experience to improve its credibility. The assumption is that any remaining difference between their observed costs is adequately explained by process variation, not by a persistent difference in underlying risk.

“Comparable” refers to the risks as measured on a consistent basis. Before pooling, check that experience uses the same coverage, limits, deductible, loss development and trend treatment, exposure definition, and relevant period. If class experience reflects known differences in risk or coverage, adjust it to a common basis first or do not pool the raw costs. Pooling is a modeling decision, not merely a way to increase the number of observations.

The Pooled Experience and Its Credibility

For a basic Bühlmann credibility calculation, suppose each exposure unit produces an observation $X$. Conditional on a risk parameter $\theta$, observations have common expected value $m(\theta)=E[X\mid\theta]$ and conditional variance $s^2(\theta)=\operatorname{Var}(X\mid\theta)$. Observations are conditionally independent. Across risks, define the hypothetical mean $m=E[m(\theta)]$, the expected process variance $v=E[s^2(\theta)]$, and the variance of hypothetical means $a=\operatorname{Var}(m(\theta))$. The Bühlmann parameter is $K=v/a$, where $a>0$.

If classes are pooled, the model treats the observations from the classes as sharing the same underlying risk parameter. For $N$ equal exposure units, let $\bar X_{\text{pool}}$ be their combined, exposure-weighted mean. Its credibility-weighted estimate is:

\widehat{m}_{\text{pool}} = Z_{\text{pool}}\bar X_{\text{pool}} + (1-Z_{\text{pool}})m, \qquad Z_{\text{pool}}=\frac{N}{N+K}.

For class $j$ with $n_j$ exposure units and observed loss cost $\bar X_j$, the combined mean is $\bar X_{\text{pool}}=\sum_j n_j\bar X_j/\sum_j n_j$. Thus, classes with more exposure contribute more to the pooled experience. The formula assumes equal exposure units under the stated common model; do not use it unchanged when exposure units have different variances or weights.

Why the Pooling Formula Works

Under the common-risk assumption, pooling $N$ observations gives a sample mean with conditional variance $s^2(\theta)/N$. Its expected process variance is therefore $v/N$. The covariance between the hypothetical mean $m(\theta)$ and this sample mean is $a$, while the sample mean’s total variance is $a+v/N$. The least-squares credibility coefficient is consequently:

Z_{\text{pool}} = \frac{\operatorname{Cov}(m(\theta),\bar X_{\text{pool}})} {\operatorname{Var}(\bar X_{\text{pool}})} = \frac{a}{a+v/N} = \frac{N}{N+v/a} = \frac{N}{N+K}.

The derivation explains both the benefit and the condition for pooling. More exposure reduces the process variance of the pooled mean and increases its credibility. But the derivation relies on one shared hypothetical mean for the pooled observations. If classes have distinct underlying expected costs, treating them as one risk can erase meaningful differences; the algebra does not establish that the pooling assumption is true.

Evaluating Whether Classes Are Comparable

Begin with a substantive review of what separates the classes. Are they neighboring bands created from a continuous rating variable, or do they differ in a way likely to affect expected losses? Check that experience has been adjusted for material differences before comparing it. An apparent difference in loss cost may be due to mix, coverage, or period rather than an underlying difference between the classes.

A quantitative diagnostic can compare the observed difference with the process variation expected if the classes have a common mean. For two independent class averages, each based on $n_j$ equal exposure units and with common known process variance $\sigma^2$ per unit, the standard error of their difference under a common-mean assumption is:

\operatorname{SE}(\bar X_1-\bar X_2) = \sqrt{\frac{\sigma^2}{n_1}+\frac{\sigma^2}{n_2}}.

A difference many standard errors from zero is a warning that the shared-mean assumption may not be reasonable. This comparison is a diagnostic, not a universal pooling rule: the variance must be appropriate for the adjusted data, and a decision should also consider the risk characteristics and quality of the experience. A small observed difference does not prove the classes identical; it may simply reflect limited information.

Establish a common basis.
Confirm that exposures and loss costs are measured consistently, and adjust known rating differences before comparing classes.

Assess risk comparability.
Consider whether the class definitions plausibly imply the same underlying expected cost after adjustment.

Compare experience with process variation.
Ask whether observed differences could reasonably arise from sampling fluctuation under a common-mean model.

Pool only when justified.
Combine adjusted experience by exposure, calculate pooled credibility, and apply the estimate only to classes the shared-risk assumption represents.

Worked Example: Pooling Two Neighboring Classes

A hypothetical insurer is reviewing two neighboring vehicle classes on the same coverage and adjusted period. Class A has 100 exposure units and a loss cost of $480 per unit; Class B has 300 units and a loss cost of $520. Assume the classes are sufficiently comparable to share one underlying mean. The collective mean is $500, the expected process variance per unit is $40,000, and the variance of hypothetical means is $400. Calculate the pooled credibility estimate and check whether the observed difference is large relative to process variation.

Method and assumption. Use the equal-unit Bühlmann model. The shared-risk assumption is given for this calculation; the numerical diagnostic can inform that judgment but cannot replace it.

Step 1: Calculate pooled experience. Class A’s aggregate loss is $100\times \$480=\$48{,}000$. Class B’s aggregate loss is $300\times \$520=\$156{,}000$. Total exposure is $100+300=400$, so:

\bar X_{\text{pool}} = \frac{\$48{,}000+\$156{,}000}{400} = \$510.

Step 2: Calculate the Bühlmann parameter and credibility. Here $K=v/a=40{,}000/400=100$. Therefore $Z_{\text{pool}}=400/(400+100)=0.8$.

Step 3: Calculate the estimate.

\widehat{m}_{\text{pool}} = 0.8(\$510)+0.2(\$500) = \$408+\$100 = \$508.

Step 4: Check the observed difference. The difference between the class means is $\$520-\$480=\$40$. Its standard error under a common mean is:

\sqrt{\frac{40{,}000}{100}+\frac{40{,}000}{300}} = \sqrt{400+133.3333} = \sqrt{533.3333} \approx \$23.09.

The difference is $\$40/\$23.09\approx1.73$ standard errors. That is not, by itself, strong evidence of a persistent class difference, although the risk review still matters. The pooled indication is $508 per exposure unit. This method avoids treating the two observed means as equally informative when their exposure volumes differ.

Worked Example: Avoiding a Pool That Hides a Risk Difference

Two hypothetical equipment classes each have 100 exposure units. Their adjusted loss costs are $460 and $640. Assume a proposed common-mean model has process variance $14,400 per unit. The collective mean is $550 and $K=100$. Assess the evidence for pooling and compare pooled and separate credibility indications.

Step 1: Compare the means with process variation. The observed difference is $\$640-\$460=\$180$. Under the common-mean assumption:

\operatorname{SE}(\bar X_1-\bar X_2) = \sqrt{\frac{14{,}400}{100}+\frac{14{,}400}{100}} = \sqrt{288} \approx \$16.97.

The observed difference is $\$180/\$16.97\approx10.61$ standard errors. This is a substantial warning against treating the classes as one homogeneous group. It does not identify the cause; the analyst should check for unadjusted differences or data issues as well as genuine risk differences.

Step 2: See what pooling would do. The pooled mean would be $(\$460+\$640)/2=\$550$, since exposures are equal. Total exposure is 200, so pooled credibility is $200/(200+100)=2/3$. The pooled estimate would be:

\frac{2}{3}(\$550)+\frac{1}{3}(\$550)=\$550.

Pooling assigns $550 to both classes, obscuring their sharply different experience. By contrast, separate class estimates using the stated $K$ and collective mean are $0.5(\$460)+0.5(\$550)=\$505$ and $0.5(\$640)+0.5(\$550)=\$595$, respectively. These separate indications retain some distinction while still blending with the collective mean. The method avoids assuming that a larger pooled sample is automatically a better estimate for each class.

Worked Example: Pooling After Adjusting for Rating Factors

A hypothetical insurer compares two related building classes. Their observed loss costs are $660 and $840 per exposure unit, based on 100 and 200 units, respectively. On a common base class, the applicable factors are 1.10 and 1.40. Assume these factors adequately account for the known difference, and assume the adjusted classes can share a mean. The collective base loss cost is $580 and $K=100$. Calculate the pooled indication and express it at each class’s factor.

Step 1: Normalize each cost to the common base. Divide each observed cost by its factor:

\frac{\$660}{1.10}=\$600, \qquad \frac{\$840}{1.40}=\$600.

Step 2: Pool adjusted experience by exposure. The total adjusted aggregate is $100(\$600)+200(\$600)=\$180{,}000$, across 300 units. Thus $\bar X_{\text{pool}}=\$180{,}000/300=\$600$.

Step 3: Apply credibility on the base scale. The pooled credibility is $300/(300+100)=0.75$, so the base indication is $0.75(\$600)+0.25(\$580)=\$595$.

Step 4: Translate back to each class. Class 1’s indication is $1.10(\$595)=\$654.50$. Class 2’s is $1.40(\$595)=\$833.00$. The classes share the same indicated base cost, not the same final class cost. This method avoids pooling the raw $660 and $840 values as if they were on an identical rating basis.

Common Mistake: Averaging Class Means Equally

When exposures differ, the unweighted arithmetic average of class loss costs is generally not the pooled loss cost. In the first example, the incorrect calculation $(\$480+\$520)/2=\$500$ gives equal weight to Class A’s 100 units and Class B’s 300 units. The correct pooled experience is $(100(\$480)+300(\$520))/400=\$510$. Credibility applied to the incorrect $500 would produce $500, rather than the correct pooled estimate of $508. Use total losses divided by total exposure, after any needed adjustments.

A second, more consequential mistake is pooling classes merely because their labels are adjacent. Adjacency does not establish equal expected costs. If underlying risks differ, a pooled indication can conceal the difference and give a misleading class rate. Conversely, noisy observed results alone do not establish that risks differ. The decision requires both a risk-based comparison and an assessment of variation in the experience.

Key takeaway: Pooling increases the volume of experience only under a defensible shared-risk assumption. Compare classes on a consistent basis, weight their experience by exposure, and use process variation to judge whether observed differences are plausible under a common mean. If pooling is not justified, retain separate class indications rather than allowing a larger sample to obscure distinct risk levels.

Check Your Understanding

Work each problem on the stated basis. Answers are rounded to cents where a dollar amount is requested.

Two comparable classes have 80 and 120 exposure units, with observed costs of $720 and $660. The collective mean is $700 and $K=100$. Find the pooled loss cost, pooled credibility, and credibility estimate. (Answer: pooled cost $684.00; credibility $2/3$; estimate $689.33.)
Two classes each have 100 units. Their adjusted means are $510 and $570, and the process standard deviation per unit is $100. Under a common-mean assumption, calculate the standard error of the difference and the difference in standard-error units. (Answer: standard error $14.14; difference $4.24$ standard errors, a warning against pooling without further justification.)
Two related classes have loss costs of $540 and $720, with base factors 0.90 and 1.20. If these factors adequately adjust the classes, what are their normalized base costs? (Answer: $600.00 for each class.)
A defensibly pooled group has 300 exposure units, pooled observed cost $620, collective mean $600, and $K=100$. Calculate its pooled credibility estimate. (Answer: $Z=0.75$; estimate $615.00.)
Two comparable classes have 50 and 150 units, with costs of $400 and $520. Calculate the pooled experience cost. Explain why the unweighted mean is not the answer. (Answer: pooled cost $490.00; the $460.00 unweighted mean incorrectly gives each class equal weight despite unequal exposure.)

What You'll Learn