Exposure-Adjusted Empirical Frequency

Why Exposure Adjustment Matters

A count of claims is difficult to interpret without knowing how long, or how much, the insured risks were observed. Four claims from two policy-years represent a different observed frequency from four claims across eight policy-years. When records have unequal exposure, the relevant empirical rate uses both the claim counts and the exposures.

This tutorial develops the calculation for claim counts recorded over policy exposure, such as policy-years. It distinguishes the rate for the combined portfolio from a simple average of policy-level rates, and explains how zero-exposure records should be handled. The methods are aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.

Definition: For record $j$, let $N_j$ be the number of claims observed during the record's exposure $e_j$, measured in a specified exposure unit. The empirical claim frequency per exposure unit is the total number of observed claims divided by total observed exposure: $$ \widehat{\lambda} = \frac{\sum_{j=1}^{m}N_j}{\sum_{j=1}^{m}e_j}, \qquad \sum_{j=1}^{m}e_j>0. $$ If exposure is measured in policy-years, $\widehat{\lambda}$ is claims per policy-year. It is a ratio of observed totals, not the fraction of policies with at least one claim.

The records must use a consistent exposure unit and a consistent definition of a claim. Each count should correspond to the exposure reported for that record. For example, the calculation is not meaningful if exposure is recorded in policy-years for some records and policy-months for others without conversion.

The Pooled Rate and Policy-Level Rates

When every record has positive exposure, a policy-level rate can be calculated as $N_j/e_j$. The portfolio rate is an exposure-weighted average of these rates: a record with more exposure contributes more weight. The weights are exposure shares, $e_j/\sum_k e_k$, and sum to one.

\widehat{\lambda} = \frac{\sum_{j=1}^{m}N_j}{\sum_{j=1}^{m}e_j} = \frac{\sum_{j=1}^{m}e_j(N_j/e_j)}{\sum_{j=1}^{m}e_j} = \sum_{j=1}^{m} \left(\frac{e_j}{\sum_{k=1}^{m}e_k}\right)\frac{N_j}{e_j}, \qquad e_j>0\ \text{for every }j.

The condition $e_j>0$ for every record matters: a policy-level rate $N_j/e_j$ is undefined when $e_j=0$. If zero-exposure records are present, their claim counts must also be zero for the records to be consistent with observation over exposure. Calculate the ratio of totals using all records, or calculate the weighted-average expression only over positive-exposure records. Do not form a policy-level rate for a zero-exposure record.

Result: If some records have zero exposure, assume their claim counts are zero and let $J_+=\{j:e_j>0\}$. Provided total exposure is positive, $$ \frac{\sum_{j=1}^{m}N_j}{\sum_{j=1}^{m}e_j} = \frac{\sum_{j\in J_+}e_j(N_j/e_j)}{\sum_{j\in J_+}e_j}. $$ The ratio of totals can include all records; the weighted-rate form uses only records with positive exposure.

The justification is direct. For each positive-exposure record, $e_j(N_j/e_j)=N_j$. Summing those identities gives the total claims from positive-exposure records. Zero-exposure records with zero claims add zero to the numerator and denominator, so including them in the totals does not change the ratio. Dividing by total exposure then gives exactly the exposure-weighted average over positive-exposure records.

Worked Calculations

Worked Example: Pooling Unequal Policy Exposures

A hypothetical insurer summarizes three records for a coverage year. Record A has exposure $0.5$ policy-year and 1 claim; record B has exposure $1.5$ policy-years and 0 claims; record C has exposure $2.0$ policy-years and 3 claims. Calculate the empirical frequency per policy-year and verify it using policy-level rates.

Method. First calculate the ratio of total claims to total exposure. Then, because all three exposures are positive, check the result as an exposure-weighted average.

Total claims are $1+0+3=4$. Total exposure is $0.5+1.5+2.0=4.0$ policy-years. Therefore, $$ \widehat{\lambda}=\frac{4}{4.0}=1.00 $$ claim per policy-year.

The record-level rates are $1/0.5=2.00$, $0/1.5=0$, and $3/2.0=1.50$ claims per policy-year. Their weighted mean is $$ \frac{0.5(2.00)+1.5(0)+2.0(1.50)}{4.0} = \frac{1.0+0+3.0}{4.0} = 1.00. $$

Check. The weighted calculation agrees with the ratio of totals. A common mistake would be to add the policy-level rates and divide by three: $(2+0+1.5)/3=1.1667$. That gives each record equal weight, despite their different exposures, so it is not the portfolio rate per policy-year.

Worked Example: How an Unweighted Average Can Mislead

A hypothetical portfolio has two records. One has exposure $0.25$ policy-year and 0 claims; the other has exposure $4.00$ policy-years and 4 claims. Calculate the pooled frequency and compare it with the unweighted average of the two record-level rates.

Method. Calculate total claims and total exposure first. Since both exposures are positive, also calculate each record's rate and average the rates with exposure weights.

There are $0+4=4$ claims and $0.25+4.00=4.25$ policy-years. The pooled rate is $$ \widehat{\lambda}=\frac{4}{4.25}=\frac{16}{17}=0.941176\ldots $$ claims per policy-year, or $0.9412$ to four decimal places.

The policy-level rates are $0/0.25=0$ and $4/4.00=1$. Their unweighted average is $(0+1)/2=0.5$, which is not the pooled rate. The exposure-weighted calculation is $$ \frac{0.25(0)+4.00(1)}{4.25} = \frac{4}{4.25} = 0.941176\ldots. $$

Check. The long-exposure record supplies most of the observed exposure, so its rate should have most of the influence on the combined rate. The unweighted average incorrectly treats $0.25$ and $4.00$ policy-years as equally informative amounts of exposure.

Worked Example: A Zero-Exposure Record

A hypothetical data extract contains three records: one with exposure $1.2$ policy-years and 2 claims, one with exposure $0.8$ policy-year and 0 claims, and one administrative record with exposure $0$ and 0 claims. Find the empirical frequency and show the valid weighted-rate calculation.

Method. Use the ratio of totals, then apply the weighted-rate identity only to records with positive exposure.

The total claim count is $2+0+0=2$, and total exposure is $1.2+0.8+0=2.0$ policy-years. Thus, $$ \widehat{\lambda}=\frac{2}{2.0}=1.00 $$ claim per policy-year.

For the positive-exposure records, the rates are $2/1.2=5/3$ and $0/0.8=0$. Their weighted calculation is $$ \frac{1.2(5/3)+0.8(0)}{1.2+0.8} = \frac{2+0}{2.0} = 1.00. $$

Check. The two calculations agree. It would be invalid to compute the administrative record's rate as $0/0$, or to include a term $0(0/0)$ in a weighted average. If a record has zero exposure but a positive claim count, investigate the data definition or recording error rather than treating that record as an ordinary exposure-based rate.

Interpretation and Conditions

The empirical frequency is an observed portfolio average over the stated exposure period. It is not automatically the underlying future frequency for every policy or risk. Interpreting it as a common rate assumes, at a minimum, that the exposure units and claim definitions are comparable and that the records reasonably represent the risks to which the rate will be applied. If underlying claim frequencies differ systematically across policies, the pooled result is still the aggregate claims-per-exposure ratio, but it need not describe any particular policy's rate.

For example, suppose higher-risk policies tend to have longer observed exposure. The pooled rate then reflects the exposure mix in these records. It should not be assumed to apply unchanged to a different portfolio with a different mix of risks. Exposure adjustment corrects for different amounts of observation; by itself, it does not adjust for differences in risk characteristics, coverage, or claim reporting.

Exposure is not always simply elapsed calendar time. It may be defined according to the coverage being studied, such as vehicle-years or insured property-years. Use the exposure measure appropriate to the intended frequency. If a record covers only part of a year, use the corresponding fraction of an exposure unit when that is the specified convention. Do not count each policy as one full exposure unit merely because each policy appears once in a data extract.

Check the unit.
Confirm that every exposure is measured in the same unit and that every claim count corresponds to that exposure period.

Check zero exposure.
A zero-exposure record cannot have a defined record-level rate. Confirm that its claim count is zero or resolve the inconsistency before interpreting the data.

Pool the records.
Add the claim counts, add the exposures, and divide total claims by total exposure, provided total exposure is positive.

Interpret in context.
State the unit of the result and remember that the pooled rate reflects the observed exposure and risk mix.

Common Exam Trap

The most common error is averaging policy-level rates equally when the requested quantity is the empirical frequency for the combined exposure. That calculation answers a different question: it gives the mean of the listed policy rates, with each record receiving the same weight regardless of exposure.

A second error is confusing claim frequency with the proportion of policies that have at least one claim. A policy with two claims contributes two to $N_j$, not one. The pooled frequency can therefore exceed one claim per policy-year; it is a rate of counts per exposure, not a probability that a policy has a claim.

Key takeaway: For exposure-adjusted empirical frequency, divide total claims by total exposure. When all records have positive exposure, this equals an exposure-weighted—not unweighted—average of record-level rates. For zero-exposure records, never calculate a record-level rate; if their claim counts are zero, calculate the weighted expression over positive-exposure records only.

Check Your Understanding

For each problem, calculate the requested rate using the stated exposure unit. Round only where requested.

A hypothetical book has 2 claims over $2$ policy-years, 1 claim over $4$ policy-years, and 3 claims over $2$ policy-years. Find the pooled frequency per policy-year. (Answer: $6/8=0.75$ claims per policy-year.)
Two records have exposures of $0.5$ and $1.5$ policy-years, with 1 claim on each. Find the pooled rate and the unweighted average of the record-level rates. (Answer: pooled rate $2/2=1.00$; unweighted average $(2+2/3)/2=4/3\approx1.3333$ claims per policy-year.)
A portfolio has $0.4$ policy-year with 0 claims, $2.0$ policy-years with 3 claims, and $1.6$ policy-years with 1 claim. Find the pooled frequency. (Answer: $4/4.0=1.00$ claim per policy-year.)
A data extract includes a record with 0 exposure and 0 claims, another with $0.75$ policy-year and 1 claim, and a third with $1.25$ policy-years and 0 claims. Find the pooled frequency and state which records can have policy-level rates. (Answer: $1/2.0=0.50$ claims per policy-year; only the two positive-exposure records have defined rates.)
Record A has 2 claims over $1$ policy-year; record B has 3 claims over $3$ policy-years. Find the pooled rate and compare it with the unweighted average of their rates. (Answer: pooled rate $5/4=1.25$; unweighted average $(2+1)/2=1.50$ claims per policy-year.)

What You'll Learn