Combining Exposure Across Multiple Periods

Why Combine Rating Periods?

A policy group may have claim experience recorded separately for several rating periods. To estimate its claim frequency, the relevant information is the total number of claims and the total exposure across those periods. The resulting estimate gives more influence to periods with more exposure, rather than treating every period as equally informative. This tutorial is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

The pooling calculation is appropriate when the periods use the same exposure basis and the claim frequency is assumed to be consistent across them. Under that assumption, combining periods also gives a direct way to assess how much exposure is available for a limited-fluctuation credibility standard.

Definition: For rating period $j$, let $N_j$ be the observed number of claims and $e_j$ the exposure measured in a consistent unit, such as policy-years. Across $m$ periods, total claims and total exposure are $$ N=\sum_{j=1}^{m}N_j, \qquad E=\sum_{j=1}^{m}e_j. $$ The pooled claim-frequency estimate is $\widehat{\lambda}=N/E$, measured in claims per exposure unit.

The quotient $N/E$ is an exposure-weighted average of the period frequencies $N_j/e_j$. It is not generally the ordinary arithmetic average of those frequencies. In particular, a short period with little exposure should not receive the same weight as a long period with much more exposure.

Combining Counts and Exposure

Suppose counts in different periods are independent and, for each period, follow a Poisson model with a common claim frequency $\lambda$ per exposure unit: $N_j\sim\operatorname{Poisson}(\lambda e_j)$. This assumes that the frequency level is stable across the periods being pooled, conditional on their exposures. Under this model, larger exposure produces a larger expected count, but it does not change the underlying frequency parameter.

Worked Example: Pooling Three Periods With Different Exposure

A hypothetical insured group has the following experience. Assume that the same claim-frequency level applies in all three periods and that exposure is measured in policy-years.

Period	Exposure $e_j$	Claims $N_j$	Period frequency $N_j/e_j$
1	50	19	0.3800
2	80	30	0.3750
3	70	24	0.3429

Step 1: Add the claims. $N=19+30+24=73$ claims.

Step 2: Add the exposure. $E=50+80+70=200$ policy-years.

Step 3: Divide total claims by total exposure. $\widehat{\lambda}=73/200=0.3650$ claims per policy-year.

Step 4: Check the weighting. The unweighted arithmetic average of the three displayed period frequencies is $(0.3800+0.3750+0.3429)/3\approx0.3660$, not exactly the pooled estimate. The pooled calculation weights each period frequency by its exposure: $(50(0.3800)+80(0.3750)+70(0.3429))/200=73/200=0.3650$, subject to rounding in the displayed period frequencies.

The pooled frequency is therefore 0.3650 claims per policy-year. The method is to pool counts and exposure, not to give every period equal weight. The calculation avoids the common mistake of averaging annual frequencies without regard to their exposure.

Why Total Exposure Determines Precision

The connection between combined exposure and credibility follows directly from the Poisson model. Independent Poisson counts add to a Poisson count, with the means adding as well. Therefore, the total count $N$ has mean $\lambda E$ and variance $\lambda E$. Dividing by the fixed total exposure $E$ gives an unbiased frequency estimate, with variance that decreases as total exposure increases.

N=\sum_{j=1}^{m}N_j\sim\operatorname{Poisson}\left(\lambda\sum_{j=1}^{m}e_j\right) =\operatorname{Poisson}(\lambda E), \qquad E[\widehat{\lambda}]=\lambda, \qquad \operatorname{Var}(\widehat{\lambda})=\frac{\lambda}{E}.

The variance statement follows because $\widehat{\lambda}=N/E$, so $\operatorname{Var}(N/E)=\operatorname{Var}(N)/E^2=\lambda E/E^2=\lambda/E$. This also explains why exposure, rather than the number of calendar periods, measures the amount of frequency information: doubling exposure halves the estimator's variance under the model.

A common limited-fluctuation full-credibility standard requires the observed count to be within a relative error $r$ of its mean with approximate probability $p$. Under a normal approximation to the Poisson count, let $z$ be the standard normal quantile satisfying $P(|Z|\le z)=p$, where $Z\sim N(0,1)$. The full standard is reached when the expected count is at least $N_0=(z/r)^2$. For a common frequency $\lambda$, this corresponds to the following required total exposure:

\lambda E_0\ge\left(\frac{z}{r}\right)^2, \qquad E_0=\frac{z^2}{r^2\lambda}.

This is an approximate standard, not an exact Poisson probability calculation. It is most suitable when the expected count is large enough for the normal approximation to be reasonable. When total exposure is below $E_0$, the conventional square-root limited-fluctuation factor is $Z_c=\sqrt{E/E_0}$; it is capped at one. Combining period exposures first is essential, because the standard is applied to the experience accumulated across the periods, not separately to each period and then averaged.

Worked Example: Credibility From Combined Exposure

A hypothetical portfolio is assumed to have a common expected frequency of $\lambda=0.4$ claims per policy-year. The full standard uses relative error $r=0.20$ and $z=2$, corresponding to an approximate central probability of 95.45%. Two periods contribute 40 and 60 policy-years, with 18 and 26 claims, respectively. Find the pooled frequency and partial credibility.

Step 1: Combine the periods. The total exposure is $E=40+60=100$ policy-years, and the total claim count is $N=18+26=44$.

Step 2: Calculate the pooled frequency. $\widehat{\lambda}=44/100=0.44$ claims per policy-year.

Step 3: Find the full-credibility count standard. $N_0=(z/r)^2=(2/0.20)^2=10^2=100$ expected claims.

Step 4: Convert the count standard to exposure. $E_0=N_0/\lambda=100/0.4=250$ policy-years.

Step 5: Calculate partial credibility. Since $E=100<250$, the uncapped factor is $Z_c=\sqrt{100/250}=\sqrt{0.4}\approx0.6325$. The cap at one does not change it, so $Z_c=0.6325$.

The portfolio has 100 policy-years, not 100 expected claims; under the assumed frequency, that exposure corresponds to 40 expected claims. The calculation uses the assumed frequency to translate the count standard into exposure. Confusing actual claims with exposure, or treating the actual count as the expected count for this conversion, would use the wrong basis.

Combining Periods When Exposure Units Differ

Counts can be added directly, but the exposures must first be expressed in the same unit. For example, policy-months can be converted to policy-years by dividing by 12. After conversion, the same pooled-frequency and credibility calculations apply. Keeping the common frequency assumption in view is just as important as matching units: adding quantities does not by itself make different frequency levels comparable.

Worked Example: Pooling Policy-Month Exposure

A hypothetical account has 180, 240, and 300 policy-months of exposure in three periods, with 8, 10, and 15 claims. Assume a common frequency across all periods. For the full-credibility standard, use $\lambda=0.5$ claims per policy-year, $r=0.25$, and $z=2$. Find the pooled annual frequency and partial credibility.

Step 1: Convert and combine exposure. Total exposure is $180+240+300=720$ policy-months. Since 12 policy-months equal one policy-year, $E=720/12=60$ policy-years.

Step 2: Combine the observed counts. $N=8+10+15=33$ claims.

Step 3: Calculate frequency in claims per policy-year. $\widehat{\lambda}=33/60=0.55$ claims per policy-year.

Step 4: Calculate required full-credibility exposure. $N_0=(2/0.25)^2=8^2=64$ expected claims, so $E_0=64/0.5=128$ policy-years.

Step 5: Calculate the credibility factor. The combined exposure is below the standard. Thus $Z_c=\sqrt{60/128}=\sqrt{0.46875}\approx0.6847$, with no adjustment from the cap at one.

The result is 0.55 claims per policy-year based on 60 policy-years, with partial credibility 0.6847 under the stated standard. Dividing the claim count by 720 without converting the exposure would instead produce a monthly frequency and would not match the annual frequency used in the standard.

When Pooling Is Appropriate

The arithmetic of pooling is simple, but its assumptions deserve attention. If the true frequency differs by period, a single pooled frequency may conceal changes in the underlying risk. For instance, if a rating change, coverage change, or material shift in the insured group makes the periods non-comparable, raw counts and exposure do not automatically provide a valid estimate for every period. The analyst needs a justified common-frequency assumption or an appropriate adjustment before combining them.

Under a common frequency assumption, independent Poisson counts, and a shared exposure basis, pooling has two useful properties: the total count remains Poisson, and the pooled estimator uses all exposure in its precision calculation. The period labels still help diagnose changes, but they do not alter the pooled calculation itself.

Key takeaway: For comparable periods, add claim counts and add exposure separately, then divide total claims by total exposure. For limited-fluctuation credibility, compare that combined exposure with the full-credibility exposure standard and cap the square-root factor at one. Do not average period frequencies equally unless the periods have equal exposure.

Check Your Understanding

For each problem, assume the stated periods are comparable wherever a common frequency is specified.

A hypothetical group has exposures of 25, 35, and 40 policy-years, with 7, 11, and 10 claims. Find total exposure, total claims, and the pooled frequency. (Answer: 100 policy-years, 28 claims, and 0.2800 claims per policy-year.)
Two periods have exposures of 80 and 120 policy-years and counts of 32 and 48. Find the pooled frequency and explain whether an unweighted average of the two period frequencies gives the same result. (Answer: 0.4000 claims per policy-year; yes, because both period frequencies are 0.4000.)
A common expected frequency is 0.25 claims per policy-year. Use $r=0.25$ and $z=2$. Three periods contribute 90, 110, and 56 policy-years. Find the full-credibility exposure and the capped partial-credibility factor. (Answer: $E_0=256$ policy-years and $Z_c=1.0000$.)
A hypothetical group has 30 and 70 policy-years of exposure, with 9 and 28 claims. For $\lambda=0.5$, $r=0.25$, and $z=2$, find the pooled frequency and the partial-credibility factor. (Answer: 0.3700 claims per policy-year and $Z_c=\sqrt{100/128}\approx0.8839$.)
Three periods have 12, 18, and 30 policy-years of exposure, with 5, 4, and 9 claims. Under a common frequency assumption, calculate the pooled estimate and state one condition needed to justify pooling. (Answer: 18 claims over 60 policy-years, giving 0.3000 claims per policy-year; for example, the periods must have a common underlying frequency and a comparable exposure basis.)

What You'll Learn