Choosing The Exposure Basis For Claim Frequency

Why the Exposure Basis Matters

A claim frequency is a count divided by an amount of exposure. The denominator is not just a convenient scaling factor: it defines the unit of risk to which the claim rate refers. For an automobile portfolio, a claim rate per policy-year answers a different question from a rate per vehicle-year when policies cover different numbers of vehicles. A sound frequency analysis therefore starts by deciding what one unit of exposure represents.

The exposure basis should fit the insured object, the way claims are counted, and the purpose of the analysis. Once the basis is chosen, the frequency model must use the corresponding exposure consistently. The earlier material on offsets explains how a known exposure amount enters a frequency model; here the focus is what that amount should measure.

Common Exposure Bases

A policy-year is one policy in force for one year. Partial periods contribute fractional policy-years: for example, one policy in force for six months contributes \(0.5\) policy-years, assuming the period is measured on a year basis. Vehicle-years count insured vehicles over time. One vehicle insured for one year contributes one vehicle-year; two vehicles insured for six months each also contribute one vehicle-year in total.

The two measures are related but not identical. A policy covering several vehicles contributes one policy-year for a full year, but several vehicle-years. Other lines may use exposure measures such as payroll, sales, passenger miles, or machine-hours when those quantities are suitable measures of the risk being insured. The appropriate choice depends on the coverage and how the exposure data and claims are defined.

There is no universally best basis. Policy-years may be appropriate when the insured risk and claim count are fundamentally policy-level. Vehicle-years can be more informative when each insured vehicle is a distinct source of claim exposure and policy sizes vary. A denominator should not be selected just because it is available or produces a preferred rate.

Worked Rates on Different Bases

How the Choice Can Change Comparisons

When groups have different average numbers of vehicles per policy, dividing by policy-years and dividing by vehicle-years can produce different comparisons. This is a substantive consequence of changing the unit of exposure, not a rounding issue. Before interpreting a difference, state the denominator and check whether the groups have comparable exposure composition on that basis.

Exposure and a Common Poisson Frequency Rate

A standard frequency formulation assumes that, for a stated exposure basis, the count \(N_j\) in observation \(j\) has a Poisson distribution with mean \(e_j\lambda\). Here \(e_j\) is the known exposure on that basis and \(\lambda\) is a common claim rate per exposure unit. The model assumes a common rate across the observations being pooled, conditional on the model specification, and that the counts follow the stated Poisson model. If the observations have different underlying rates, one common rate is not an appropriate description without further modeling.

Under these assumptions, the maximum-likelihood estimate of the common rate is total claims divided by total exposure. This is an exposure-weighted rate, not generally the simple average of the individual observed rates.

Why Total Claims Divided by Total Exposure Works

For independent observations \(N_j\sim\operatorname{Poisson}(e_j\lambda)\), the likelihood, omitting factors that do not involve \(\lambda\), is proportional to \(\prod_j (e_j\lambda)^{N_j}\exp(-e_j\lambda)\). Taking its logarithm and differentiating gives the score equation below.

Setting the derivative equal to zero gives the estimate. For positive total exposure, the log-likelihood is concave in \(\lambda\), since its second derivative is negative when there are claims. Thus this stationary point is the maximum. If total claims are zero, the likelihood is maximized at the boundary \(\hat{\lambda}=0\).

This derivation also explains why exposure amounts, rather than observations counted equally, determine the pooling weights. Each observation contributes claims and its corresponding amount of exposure. A high-frequency estimate based on very little exposure should not receive the same weight as an estimate based on much more exposure when the model assumes one common rate.

Choosing and Auditing a Basis

A practical choice follows from the intended interpretation and the data definition. Ask which object or quantity is exposed to the covered peril, whether the claims are counted on that same basis, and whether the exposure can be measured consistently across risks and periods. For example, a policy-level coverage may naturally be analyzed per policy-year, whereas vehicle-level claim frequency may be better expressed per vehicle-year if vehicle counts and insured periods are known.

For any selected basis, verify that:

• Exposure includes only the period during which the relevant coverage was in force.
• The unit is consistent across policies, groups, and time periods, including partial periods.
• Claims and exposure refer to the same coverage and risk population.
• A multi-item policy contributes the appropriate number of units when the chosen basis counts items separately.
• The chosen denominator is not confused with premium, which is a price rather than a direct measure of exposure.

A policy-year and a vehicle-year can both be valid for the same portfolio, but their rates should not be interchanged without an explicit conversion and a clear statement of assumptions. In particular, a portfolio-wide average number of vehicles per policy does not guarantee that every rating group has the same mix. A single conversion factor can therefore obscure differences in composition.