Frequency Credibility with Unequal Policy Durations

Why Duration Matters in Frequency Credibility

A claim count is meaningful for frequency credibility only in relation to the time over which the insured risks were exposed to loss. A portfolio with 100 policies in force for a full year has 100 policy-years of exposure. A portfolio with 100 policies in force for half a year has only 50 policy-years, even though both portfolios contain 100 policies. Treating them as having equal exposure can distort both the estimated frequency and the assessment of credibility.

The conversion is direct when the expected claim rate is constant over the covered period: add each policy’s covered fraction of a year. That total is the effective exposure used to estimate frequency and compare the data with a full-credibility standard. This tutorial is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

From Unequal Durations to a Frequency Estimate

If \(N\) claims are observed over \(E\) policy-years, the observed claim frequency is \(N/E\), measured in claims per policy-year. The denominator is total exposure, not the number of distinct policies and not the number of policy terms. A policy covered for nine months contributes \(0.75\) policy-years whether or not a claim occurs.

The proportional-duration calculation assumes that each covered segment has the same underlying claim frequency per unit of time. It also assumes that the exposure and claim count refer to compatible risks and time periods. If claim frequency changes materially by season, policy characteristics, or coverage form, simply adding durations may not represent a homogeneous portfolio; the appropriate response is to adjust or stratify the exposure and frequency assumptions rather than blindly prorate.

Poisson Full Credibility With Effective Exposure

For the standard frequency calculation, assume claims follow a homogeneous Poisson process with rate \(\lambda\) claims per policy-year. In a portfolio with \(E\) policy-years, the total count has distribution \(N\sim\operatorname{Poisson}(\lambda E)\). This model assumes a constant rate over the covered time and independent claim occurrences across the exposure; it does not represent material clustering or an unmodeled change in risk over time.

A full-credibility standard specifies a central confidence level \(p\) and a relative error tolerance \(r\). Under the normal approximation to the Poisson count, let \(z\) be the standard normal quantile satisfying \(\Phi(z)=(1+p)/2\). The requirement is that the estimated frequency \(N/E\) fall within \(r\lambda\) of \(\lambda\) with approximately probability \(p\).

The result follows from the mean and variance of the Poisson count: \(E[N]=\lambda E\) and \(\operatorname{Var}(N)=\lambda E\). Therefore the frequency estimator has mean \(\lambda\) and variance \(\lambda/E\). Its approximate standard deviation is \(\sqrt{\lambda/E}\). A central confidence requirement with relative error \(r\) asks for the error bound \(r\lambda\) to be at least \(z\) standard deviations:

Both sides are nonnegative. Squaring and rearranging gives:

Multiplying the exposure requirement by \(\lambda\) gives \(\lambda E\ge z^2/r^2\), the expected-count form. These are two ways of expressing the same standard. Use the exposure form when policy durations differ, and do not substitute the number of policies unless every policy contributes exactly one policy-year.

This is a normal-approximation standard, not an exact Poisson probability calculation. When expected counts are small, the normal approximation may be poor; use an exact method or the prescribed standard if the problem requires one. The expected frequency \(\lambda\) must be selected on a relevant basis before the comparison. Substituting the observed frequency merely to make the threshold appear consistent with the data can make the assessment circular.

Day-Based Exposure and a Second Assessment

Some records provide exact start, cancellation, or expiration dates rather than rounded months. In that case, calculate each policy’s covered days and divide by the stated annual day-count basis. Be consistent about whether the exposure and claims cover the same period. A policy that is active for 120 days contributes \(120/365\) policy-years under a 365-day convention, not one full policy-year.

Applying the Calculation in Practice

A reliable calculation keeps the duration conversion, frequency estimate, and full-credibility test distinct. In particular, the expected frequency used in the standard is an assumption or selected estimate for the risk being assessed; it is not automatically the observed frequency from the same data. The observed frequency describes what happened, while the standard asks how much exposure would be needed to estimate the assumed frequency to a specified precision.

The most consequential errors are usually denominator errors: counting each policy as a full year despite partial coverage, or using a duration convention inconsistent with the data. A second trap is switching between a required expected claim count and a required exposure without dividing by \(\lambda\). For example, \(z^2/r^2\) is a claim-count standard; it is not a policy-year standard. Exposure depends on expected frequency, so \(E_0=(z^2/r^2)/\lambda\).

Prorating duration is also not a substitute for addressing heterogeneity. If two groups have materially different expected claim rates, their combined frequency can still be calculated as total claims divided by total exposure only when the intended quantity is the exposure-weighted portfolio frequency. For a credibility decision about one homogeneous risk class, mismatched classes should be adjusted or assessed separately.