Frequency Estimates with Unequal Exposure Periods

Why Unequal Observation Periods Require Care

A policy observed for two years contributes more time at risk than a policy observed for three months. If both policies are counted as one observation when averaging their annualized claim rates, the short observation period receives the same weight as the long one. That may answer a question about the average policy-level rate, but it is not the same calculation as estimating claims per total unit of exposure.

For frequency estimation, exposure measures how long, or how much, a risk was observed. It may be recorded in policy-years, policy-months, or another stated unit. To combine policies, first express all exposure in the same unit. The pooled empirical frequency is then the number of observed claims divided by the total exposure.

Definition: For policy $j$, let $N_j$ be its observed claim count and $e_j$ its exposure in a stated unit, such as policy-years. For $m$ policies with positive exposure, the pooled empirical frequency per exposure unit is $$ \widehat{\lambda}_{\mathrm{pool}} = \frac{\sum_{j=1}^{m}N_j}{\sum_{j=1}^{m}e_j}. $$ It describes the observed number of claims per total exposure unit. It is not, in general, the same as the unweighted average of the policy-level rates $N_j/e_j$.

The numerator counts claims, not policies with at least one claim. A policy can have more than one claim during its observation period. The denominator is the total time at risk, not simply the number of policies or the average policy duration. When exposures are stated in years, the result is claims per policy-year; it is an estimated expected count per year under an appropriate frequency model, not a probability of at least one claim.

Why the Pooled Estimator Uses Total Exposure

A standard model for this calculation assumes that each policy has an independent Poisson claim count with a common constant rate $\lambda$ per exposure unit. Conditional on exposure $e_j$, the model is $N_j\sim\operatorname{Poisson}(\lambda e_j)$. This assumes claim arrivals are adequately represented by a Poisson process over the observation periods, the rate is constant over those periods, and the observed counts and exposures are recorded on a consistent basis.

Theorem: Under independent Poisson counts $N_j\sim\operatorname{Poisson}(\lambda e_j)$, with known positive exposures $e_j$, the maximum-likelihood estimate of the common claim rate is total claims divided by total exposure. If the total claim count is zero, the maximum-likelihood estimate is zero.

To see why, write the likelihood as the product of the Poisson probabilities. Terms that do not depend on $\lambda$ can be collected separately:

L(\lambda) = \prod_{j=1}^{m} \frac{(\lambda e_j)^{N_j}e^{-\lambda e_j}}{N_j!} \ \propto\ \lambda^{\sum_j N_j}e^{-\lambda\sum_j e_j}.

Let $N=\sum_jN_j$ and $E=\sum_je_j$. For $N>0$, the log-likelihood, ignoring constants, is $N\log\lambda-\lambda E$. Its derivative is $N/\lambda-E$. Setting this equal to zero gives $\widehat{\lambda}=N/E$; the second derivative is $-N/\lambda^2<0$, so this stationary point is a maximum. If $N=0$, the likelihood is proportional to $e^{-\lambda E}$, which decreases as $\lambda$ increases, so its maximum over $\lambda\geq0$ is at zero.

Choose the exposure unit.
Convert all observed periods to the same unit before combining them.

Sum claims and exposure separately.
Compute $N=\sum_jN_j$ and $E=\sum_je_j$; do not divide each count by a different unit.

Divide and label the result.
Use $N/E$, and state the resulting frequency per exposure unit.

Worked Examples

Worked Example: Pooling Four Policies With Different Durations

A hypothetical insurer observes four policies for $0.5$, $1$, $2$, and $3$ policy-years. Their respective claim counts are $0$, $1$, $3$, and $4$. Calculate the pooled empirical frequency and compare it with the unweighted mean of the four policy-level annualized rates. Assume for interpretation that independent Poisson counts with a common constant rate per policy-year are appropriate.

Method. Sum the counts and exposures separately for the pooled rate. Then calculate each policy’s rate by dividing its own count by its own exposure. This comparison avoids giving the two calculations the same interpretation merely because both are expressed per year.

The total claim count is $$ N=0+1+3+4=8. $$ The total exposure is $$ E=0.5+1+2+3=6.5\text{ policy-years}. $$ Therefore, the pooled estimate is $$ \widehat{\lambda}_{\mathrm{pool}}=\frac{8}{6.5}=1.230769\ldots $$ The pooled estimate is $1.2308$ claims per policy-year, to four decimal places.

The policy-level annualized rates are $0/0.5=0$, $1/1=1$, $3/2=1.5$, and $4/3=1.3333\ldots$ claims per policy-year. Their unweighted mean is $$ \frac{0+1+1.5+1.333333\ldots}{4} = 0.958333\ldots, $$ or $0.9583$ claims per policy-year to four decimal places.

The pooled estimate is higher because it weights each policy-level rate by that policy’s exposure. The unweighted mean gives each of the four policies equal weight. As a check, multiplying the pooled estimate by the total exposure gives $1.230769\ldots\times6.5=8$, the observed total number of claims.

Worked Example: Converting Policy-Months to Policy-Years

A hypothetical portfolio has three policies observed for $3$, $9$, and $15$ months, with $1$, $0$, and $4$ claims, respectively. Calculate the pooled annual frequency and the unweighted mean of the policy-level annualized rates. Assume independent Poisson counts with a common rate per policy-year.

Method. Convert each period to years before calculating a rate per year. This prevents a frequency per month from being mistaken for a frequency per year.

The total count is $1+0+4=5$. Total exposure is $3+9+15=27$ policy-months, or $$ E=\frac{27}{12}=2.25\text{ policy-years}. $$ Thus, $$ \widehat{\lambda}_{\mathrm{pool}}=\frac{5}{2.25}=2.2222\ldots $$ The pooled estimate is $2.2222$ claims per policy-year, to four decimal places.

The individual exposures are $3/12=0.25$, $9/12=0.75$, and $15/12=1.25$ years. The corresponding annualized rates are $1/0.25=4$, $0/0.75=0$, and $4/1.25=3.2$. Their unweighted mean is $$ \frac{4+0+3.2}{3}=\frac{7.2}{3}=2.4 $$ claims per policy-year. The two estimates differ because one pools all claim counts and time, while the other averages three policy-specific ratios. As a check, $2.2222\ldots\times2.25=5$ claims.

Worked Example: Combining Grouped Policy Records

A hypothetical book of business is summarized in three groups: 20 policies observed for $0.5$ year each produced 8 claims; 10 policies observed for $1.5$ years each produced 25 claims; and 5 policies observed for $2$ years each produced 18 claims. Calculate the pooled rate. Then calculate the mean of the policy-level annualized rates, where each policy receives equal weight. Assume the exposure is the same for policies within each group and the common-rate Poisson model is being used for the pooled estimate.

Method. Use the group sizes to recover each group’s total exposure, then add group claims and exposure. For the equal-policy mean, first calculate each group’s average claim count per policy and divide by that group’s exposure per policy.

The group exposures are $20(0.5)=10$, $10(1.5)=15$, and $5(2)=10$ policy-years. Therefore, $$ N=8+25+18=51,\qquad E=10+15+10=35\text{ policy-years}. $$ The pooled estimate is $$ \widehat{\lambda}_{\mathrm{pool}}=\frac{51}{35}=1.457142\ldots, $$ or $1.4571$ claims per policy-year to four decimal places.

For the first group, the average count per policy is $8/20=0.4$, so its average policy-level rate is $0.4/0.5=0.8$. For the second group, the average count per policy is $25/10=2.5$, giving $2.5/1.5=1.6667\ldots$. For the third group, the average count is $18/5=3.6$, giving $3.6/2=1.8$.

There are 20, 10, and 5 policies at those respective rates. The equal-policy mean is $$ \frac{20(0.8)+10(1.666666\ldots)+5(1.8)}{20+10+5} = \frac{16+16.666666\ldots+9}{35} = 1.190476\ldots, $$ or $1.1905$ claims per policy-year to four decimal places. To check the pooled calculation, weighting the group rates by group exposure gives $\{10(0.8)+15(1.666666\ldots)+10(1.8)\}/35=51/35$, as required.

What the Two Averages Mean

For policy-level rates $r_j=N_j/e_j$, the pooled estimate can also be written as an exposure-weighted average: $$ \widehat{\lambda}_{\mathrm{pool}} = \frac{\sum_j e_jr_j}{\sum_j e_j}. $$ A policy observed for more time contributes more weight because it contributes more exposure. In contrast, the unweighted average $\sum_jr_j/m$ gives every policy the same weight regardless of how long it was observed.

Under the common-rate Poisson model, both quantities can estimate the same underlying $\lambda$, but the pooled estimator is the maximum-likelihood estimator and uses the total exposure directly. If the intended summary is instead the average of observed annualized rates across policies, the unweighted average answers that different descriptive question. The appropriate calculation depends on the target, not just on which result is easier to compute.

If policy rates are not common, the pooled calculation still has a clear descriptive meaning: claims divided by total exposure. It does not automatically estimate the equal-policy average rate. Avoid treating the two estimands as interchangeable.

Common Mistakes and Checks

A frequent error is to calculate each policy’s rate and then take the simple mean when the question asks for claims per total exposure. In the first example, that incorrect substitution would report $0.9583$, rather than the pooled $1.2308$. The correction is to sum the eight claims and divide by the $6.5$ policy-years. The unweighted answer is valid only when the requested target is the equal-policy average rate.

Mixing time units. Do not combine policy-months and policy-years without conversion. Label the final rate with its unit.
Counting affected policies instead of claims. A policy with three claims contributes three to the count, not one.
Dividing by the number of policies. That gives claims per policy over the observed periods, not claims per exposure unit.
Interpreting frequency as probability. A rate of $1.2$ claims per policy-year is not a probability of $1.2$; it is an expected count per unit of exposure under the model.
Ignoring the observation basis. Claims and exposures must refer to compatible risks and time periods. The Poisson interpretation also relies on the stated model assumptions.

As a final arithmetic check, multiply a pooled rate by total exposure: the result should equal the observed total claim count, apart from rounding. A sensible frequency estimate also requires positive total exposure; a zero-exposure record contributes neither claims nor time at risk.

Check Your Understanding

For each problem, use the stated exposure unit and distinguish a pooled estimate from an equal-policy average when requested.

Three policies have exposures of $0.5$, $1$, and $1.5$ years and claim counts of $2$, $1$, and $0$. Find the pooled frequency per policy-year. (Answer: $1.0000$ claims per policy-year)
Three policies have exposures of $6$, $18$, and $30$ months and claim counts of $1$, $2$, and $0$. Find the pooled annual frequency. (Answer: $0.6667$ claims per policy-year)
Four policies have exposures of $0.25$, $0.5$, $1$, and $2$ years and counts of $1$, $0$, $2$, and $1$. Find both the pooled frequency and the unweighted mean of the annualized policy rates. (Answer: pooled $1.0667$; unweighted mean $1.6250$ claims per policy-year)
Ten policies observed for $0.5$ year each have 3 claims, and five policies observed for 2 years each have 14 claims. Find the pooled rate and the equal-policy mean of the group-average annualized rates. (Answer: pooled $1.1333$; equal-policy mean $0.8667$ claims per policy-year)
Three policies have exposures of $120$, $240$, and $360$ days and counts of $1$, $2$, and $3$. Use a 365-day year to find the pooled annual frequency. (Answer: $3.0417$ claims per policy-year)

What You'll Learn