Bühlmann Credibility for Claim Frequencies

Why Apply Credibility to Claim Frequency?

A risk’s claim frequency can vary from one observation period to another, even when its underlying tendency to generate claims is stable. A short run of unusually high or low frequency should therefore not automatically determine the estimate for future periods. Bühlmann credibility balances the risk’s observed frequency against the collective average: more consistent, relevant experience receives more weight, while the collective mean supplies a complement.

Here, frequency means the number of claims divided by a stated exposure amount, such as claims per policy-year or per 1,000 vehicle-years. The exposure basis must be specified and used consistently. The standard Bühlmann factor \(Z=n/(n+K)\) applies when the \(n\) observations are comparable frequency measurements with the same exposure basis and model variance structure. It is not a formula for weighting raw claim counts from periods with different exposure.

For a risk with \(n\) observations, let \(\bar X\) be its arithmetic mean frequency. When EPV and VHM are positive, define \(K=\mathrm{EPV}/\mathrm{VHM}\). The Bühlmann credibility-weighted estimate of that risk’s expected frequency is \(Z\bar X+(1-Z)\mu\), where \(Z=n/(n+K)\). The estimate is a frequency, not a prediction that the next period will contain a fractional claim count.

The units matter. If \(X\) is claims per 1,000 policy-years, then \(\mu\), \(\bar X\), EPV, and VHM all refer to observations on that frequency basis. \(K\) is measured in observation-period units, consistent with the definition of \(n\). A calculated \(\widehat m\) can be converted to an expected claim count for a specified future exposure by multiplying by that exposure in the matching units.

Why the Credibility Weight Has This Form

The formula is the best linear estimate of the hypothetical mean under squared-error loss. Consider estimates of the form \(a\bar X+(1-a)\mu\), where \(a\) is a weight between zero and one. The risk-specific target is \(m(\Theta)\). Conditional independence gives \(\operatorname{Var}(\bar X\mid\Theta)=s^2(\Theta)/n\). Averaging over risks gives \(\operatorname{Var}(\bar X-m(\Theta))=\mathrm{EPV}/n\); the difference between the collective mean and the risk-specific mean has variance VHM. The cross-product has expectation zero because the conditional sampling error has mean zero given \(\Theta\).

Thus the mean squared error of the linear estimate is

Differentiate with respect to \(a\) and set the derivative equal to zero:

Rearranging gives \(a(\mathrm{EPV}/n+\mathrm{VHM})=\mathrm{VHM}\), so

The second derivative of the mean squared error is \(2(\mathrm{EPV}/n+\mathrm{VHM})\), which is positive when the variance components are positive. This establishes that the weight is the minimizing linear weight. The result does not say that the weighted estimate is the exact posterior mean for every possible distribution of risk parameters; it is the Bühlmann linear credibility estimate.

Worked Calculations for Frequency Experience

The following examples use hypothetical insurance portfolios. For the standard factor, each observation is one comparable period on a fixed exposure basis. When data are used to estimate variance components, the balanced setup has the same number of periods for each risk.

Conditions, Interpretation, and Exam Traps

The frequency application uses the same credibility structure as other Bühlmann applications, but the observation must be defined carefully. For instance, counts recorded over unequal exposure periods do not have the same variance merely because each is one calendar year. In a conditional Poisson model, a count over exposure \(e\) has conditional mean and variance proportional to \(e\); dividing by \(e\) changes the variance with exposure. Consequently, applying \(n/(n+K)\) to unadjusted counts, or to exposure-normalized rates with materially different exposure without accounting for their differing variances, is not justified by the equal-observation formula. Use comparable exposure units or an appropriate unequal-exposure method.

Other assumptions deserve attention: the risk’s underlying frequency tendency is treated as stable over the observations; periods are conditionally independent; and the collective mean and variance components describe the same portfolio basis as the risk being estimated. A trend, coverage change, or shift in claim definition can make historical frequency observations non-comparable. The credibility calculation does not correct such changes automatically.

The weighted estimate is a convex combination when \(0\leq Z\leq1\), so it lies between the observed mean and collective mean. As \(n\) grows with \(K\) fixed, \(Z\) increases and the risk’s own experience has greater influence. If VHM is zero, there is no modeled variation in hypothetical means across risks, and the ratio defining \(K\) is not usable in its ordinary form; the frequency data provide no between-risk differentiation under that model. If variance-component estimates are unstable or outside plausible bounds, investigate their estimation and modeling assumptions rather than treating the resulting factor as self-validating.

• Do not confuse frequency with count. A count is tied to its exposure; frequency is count divided by stated exposure.
• Do not confuse \(K\) with \(Z\). \(K\) is EPV divided by VHM; \(Z=n/(n+K)\) is the weight on the observed mean.
• Do not omit the VHM correction. The variation among observed risk means includes process variation in those means.
• Do not mix units. The observed mean, collective mean, and variance components must correspond to the same defined observation basis.

The practical purpose is to stabilize a risk’s estimated frequency without discarding its information. In ratemaking, frequency can contribute to a broader estimate of expected claim cost, but this calculation alone does not supply claim severity, expenses, or a final premium. The frequency estimate is one model component, and its interpretation depends on the exposure definition and assumptions stated above. This tutorial is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.