Unbalanced Experience Across Risk Classes

Why Unbalanced Experience Requires Exposure Weighting

Risk classes rarely have identical experience records. One class may have more observation periods, while another has fewer periods with more exposure in each. In the Bühlmann-Straub model, neither the number of periods nor an unweighted average alone captures the amount of information in a class. The key quantity is its total exposure, which combines period-level weights and determines the class’s credibility.

This matters when experience is summarized across different policy years, territories, or other risk groupings. A period with twice the exposure of another period contributes twice as much to the weighted average. But even when two classes have different period counts, they may have the same total exposure and therefore the same credibility factor under the model.

The Model and the Class Estimate

For a class \(i\), let \(\Theta_i\) denote its underlying expected level. The standard Bühlmann-Straub assumptions are that \(E[X_{ij}\mid\Theta_i]=\Theta_i\) and \(\operatorname{Var}(X_{ij}\mid\Theta_i)=\sigma^2/w_{ij}\). Conditional on \(\Theta_i\), observations from distinct periods are uncorrelated. Across classes, the structural mean is \(\mu=E[\Theta_i]\), and the variance of the underlying class means is \(\tau^2=\operatorname{Var}(\Theta_i)\).

The exposure weight represents the relative amount of information in a period. A larger weight reduces the conditional variance: a period with weight \(4\) has one quarter the conditional variance of a period with weight \(1\), under these assumptions. The weights must be on a consistent scale, and the same variance parameters must apply across the classes being compared.

The estimate blends the class’s own weighted average with the structural mean. A class with more total exposure receives a larger \(Z_i\), so its observed experience has greater influence. The complement \(1-Z_i=K/(W_i+K)\) is the weight on the structural mean.

Why Total Exposure Determines Credibility

The result follows by examining the weighted average for a single class. Given its underlying mean \(\Theta_i\), the weighted average is unbiased for \(\Theta_i\). Conditional independence and the model’s variance assumption give:

Thus the weighted average’s conditional variance depends on the period weights through their sum, not through the number of terms. Unconditionally, the variance includes both variation among class means and process variation:

For a linear estimate of the form \(\mu+c(\bar X_i-\mu)\), the coefficient minimizing mean squared prediction error is the covariance of \(\Theta_i\) with \(\bar X_i\), divided by the variance of \(\bar X_i\). Therefore,

This establishes the credibility factor and the stated estimate. The derivation relies on the model assumptions above; it is not a general rule for arbitrary weighted data. In particular, if the weights do not represent exposure in the specified variance relationship, using their sum in this formula may not be justified.

Equal Total Exposure, Different Period Schedules

Two classes can have different period schedules yet the same weighted average and total exposure. In that case, they have the same estimate under this model. This is a consequence of the model’s variance structure: once the total weight and weighted average are known, the details of how the exposure was divided among periods do not change the formula.

Unbalanced Classes and the Period-Count Trap

When class schedules are unbalanced, the period count can still be useful for describing the data, but it is not a substitute for \(W_i\). A class with only a few heavily weighted periods may have as much credibility as a class with many lightly weighted periods. Conversely, many periods need not imply high credibility if their total exposure is small.

The model also distinguishes a weighted average from an ordinary average. If period weights differ, an ordinary average gives each observation equal influence, even though the model assigns different information to those observations. Using an ordinary average together with exposure-based credibility can therefore produce an estimate inconsistent with the model.

Interpretation and Practical Checks

Total exposure has a direct interpretation in the model. As \(W_i\) increases, the conditional variance \(\sigma^2/W_i\) of the class weighted average decreases. The class experience is then more informative about \(\Theta_i\), and \(Z_i\) rises toward \(1\). As \(W_i\) becomes small relative to \(K\), the estimate places more weight on \(\mu\).

Some useful checks are:

• Exposure check: confirm that each \(W_i\) is the sum of the period weights, not the period count.
• Average check: verify that the weighted products sum to \(W_i\bar X_i\).
• Parameter check: calculate \(K=\sigma^2/\tau^2\), and ensure its exposure scale matches the weights.
• Blend check: for \(0\leq Z_i\leq1\), the estimate must lie between \(\bar X_i\) and \(\mu\).

The credibility formula treats \(\mu,\sigma^2,\tau^2\) as specified inputs; uncertainty from estimating those parameters is not included in this class-level calculation. If \(W_i=0\), there is no class experience to average, and the formula gives \(Z_i=0\), so the estimate is \(\mu\). These conclusions are tied to the Bühlmann-Straub assumptions, especially the inverse relationship between conditional variance and exposure weight.