Model-Based Credibility: The Variation Perspective

Why Credibility Depends on Two Kinds of Variation

Observed experience is informative, but it is not a perfect measure of a risk group's underlying expected claim measure. Part of the difference between experience and expectation is random fluctuation within the group. Another part reflects genuine differences in expected outcomes among groups. Model-based credibility describes how these two sources of variation govern the weight assigned to observed experience.

The variation perspective complements the stability perspective developed in earlier material on limited-fluctuation credibility. Instead of asking only whether a group's experience meets a prescribed precision standard, it asks how noisy the observations are relative to how different the groups are. The result is a credibility weight: a way to blend the group's experience with a reference mean. This tutorial uses the standard Bühlmann framework for equal-exposure observations; it does not derive a particular claim-frequency or claim-severity model.

The term “hypothetical mean” refers to the expected claim measure for a particular group, if that group’s underlying characteristics were known. Different groups can have different hypothetical means even when their observations are subject to the same general process. EPV and VHM describe different things: EPV is variability of observations around a group’s expected value, while VHM is variability of those expected values around the overall mean.

The Basic Model and Credibility Formula

For the basic Bühlmann model, suppose a group has \(n\) observations \(X_1,\ldots,X_n\). Conditional on its underlying parameter \(\Theta\), the observations are independent and identically distributed, with conditional mean \(\Theta\). Across groups, assume \(\Theta\) has mean \(m\) and finite variance. Also assume finite process variance, equal exposure units, and that the observations have been made comparable: the same claim measure, exposure basis, and relevant conditions apply. Define the group's experience mean as \(\bar X=(X_1+\cdots+X_n)/n\).

Under these assumptions, the Bühlmann credibility estimate of the group's hypothetical mean is a weighted average of its experience mean and the overall mean:

Here \(Z\) is the credibility assigned to the group's experience, and \(1-Z\) is the weight assigned to the overall mean. When VHM is positive, larger EPV makes \(K\) larger and reduces \(Z\) for a fixed \(n\). Larger VHM makes \(K\) smaller and increases \(Z\): if groups genuinely differ more in their expected values, an observation is more useful for identifying which kind of group was observed. If VHM is zero, all groups have the same expected value under this model, so the estimate is \(m\); the ratio defining \(K\) is not used in that limiting case.

Why the Weight Has This Form

The formula follows from finding the best linear predictor of \(\Theta\) using \(\bar X\). “Best” here means minimizing mean squared error among predictors that are linear in \(\bar X\). The conditional independence assumption gives the variance of \(\bar X\) directly: within a group, averaging \(n\) observations divides the conditional process variance by \(n\).

By the law of total variance, the unconditional variance of \(\bar X\) is the sum of the variance of its conditional mean and its expected conditional variance. Also, since the conditional mean of \(\bar X\) is \(\Theta\), its covariance with \(\Theta\) is \(\operatorname{Var}(\Theta)=\operatorname{VHM}\).

The coefficient of \(\bar X\) in the best linear predictor of \(\Theta\) is its covariance with \(\Theta\), divided by its variance. Substitution gives the credibility factor:

The predictor is centered at \(m\), so its remaining weight, \(1-Z\), applies to \(m\). This argument justifies the linear credibility formula under the stated moment and sampling assumptions. It does not establish that the formula is the full conditional expectation \(E[\Theta\mid X_1,\ldots,X_n]\) for every possible distribution; a full Bayesian calculation generally depends on the distributions assumed for the observations and for \(\Theta\).

Applying the Formula

Interpreting the Result and Knowing Its Limits

The formula makes the effect of additional equal exposure units explicit. For positive EPV and VHM, \(Z=n/(n+K)\) increases as \(n\) increases, but each additional unit has less impact on the weight than the previous one. As \(n\) approaches zero, \(Z\) approaches zero; as \(n\) grows without bound, \(Z\) approaches one. This is a statement about the model's weight, not a guarantee that new observations will be representative or accurate.

The model also clarifies why the same experience count need not imply the same credibility in every setting. A high EPV means substantial random fluctuation within groups, making a group's observed mean less informative. A high VHM means greater underlying differences among groups, making group-specific experience more valuable for distinguishing one group's expected value from the overall mean. Both components must be interpreted on the same claim-measure and exposure basis.

The basic formula is not appropriate without adjustment when observations have unequal exposures, are not comparable, or exhibit dependence not captured by the model. Trends, changes in coverage or claim handling, shifting exposure mix, and data-quality problems can also make the inputs unrepresentative. Nor does a high or low \(Z\) itself establish that the data are sound or that a resulting rate is adequate. Model-based credibility is a structured weighting method conditional on assumptions, not a substitute for checking the experience being weighted.