Pooling Homogeneous Experience for Credibility

Why Pool Experience?

A single period or small group may have too few claims to produce a stable observed frequency. If several groups or periods measure the same kind of risk on the same basis, their counts can be combined to estimate a common frequency. The key is to combine claims and exposure together, not to treat every observed rate as equally informative regardless of its exposure volume.

Pooling is appropriate only when the observations are sufficiently comparable for the question being asked. A larger pooled data set does not fix differences in coverage, claim definitions, risk characteristics, or frequency level. Those differences can make the pooled result describe a changing mixture rather than the rate for any particular group.

Definition: For observations indexed by $j=1,\ldots,m$, let $N_j$ be the claim count and $e_j$ the exposure in group or period $j$, expressed in one consistent exposure unit. The pooled claim frequency is $$ \widehat{\lambda}_{\text{pool}}=\frac{\sum_{j=1}^{m}N_j}{\sum_{j=1}^{m}e_j}. $$ It is the total number of claims divided by the total exposure. It estimates one common claim rate per exposure unit when pooling is justified.

Why the Pooled Rate Is Exposure-Weighted

Write the observed frequency for group $j$ as $r_j=N_j/e_j$, assuming $e_j>0$. Since $N_j=e_jr_j$, the pooled rate can also be written as

\widehat{\lambda}_{\text{pool}} =\frac{\sum_{j=1}^{m}e_jr_j}{\sum_{j=1}^{m}e_j}.

Thus each group’s rate is weighted by its share of the total exposure. A rate based on many exposure units has more influence than one based on few exposure units. An ordinary average gives each group equal weight, which generally answers a different question: the average of the group rates, not the frequency across all the combined exposure.

Worked Example: Pooling Three Comparable Periods

A hypothetical specialty insurer records the following claim counts for three periods. Assume the periods use the same exposure unit and cover a comparable risk population. Find the pooled frequency and compare it with the unweighted average of the three period frequencies.

Period	Claims	Exposure
1	36	1,200
2	49	1,400
3	15	600

Method: exposure-weighted pooling. First calculate each period’s frequency:

r_1=\frac{36}{1{,}200}=0.030,\qquad r_2=\frac{49}{1{,}400}=0.035,\qquad r_3=\frac{15}{600}=0.025.

The total claim count is $36+49+15=100$, and the total exposure is $1{,}200+1{,}400+600=3{,}200$. Therefore,

\widehat{\lambda}_{\text{pool}} =\frac{100}{3{,}200} =0.03125.

The unweighted average of the three rates is $(0.030+0.035+0.025)/3=0.030$, not $0.03125$. It gives the smallest period, with 600 exposure units, the same influence as the period with 1,400 units. Common mistake: averaging the period frequencies equally. The correction is to weight each rate by its exposure, or equivalently divide total claims by total exposure. The pooled result is $0.03125$ claims per exposure unit.

A Common-Rate Model and Its Justification

A standard model for pooling claim counts assumes that, conditional on their exposures, counts from different groups or periods are independent Poisson random variables with one common frequency $\lambda$:

N_j\sim\operatorname{Poisson}(\lambda e_j),\qquad j=1,\ldots,m.

Here $e_j$ is the amount of exposure at risk, and $\lambda e_j$ is the expected claim count. The model allows exposure volumes to differ; homogeneity means the expected claim rate per exposure unit is the same. Independence and the Poisson variance assumption also matter when using this model to assess the statistical variability of the pooled count.

The pooled estimator follows directly from the likelihood. Let $N=\sum_jN_j$ and $E=\sum_je_j$. Under the stated model, the likelihood as a function of $\lambda$ is proportional to $e^{-\lambda E}\lambda^N$. For $N>0$, the log-likelihood, apart from terms not involving $\lambda$, is $-\lambda E+N\log(\lambda)$. Its derivative is $-E+N/\lambda$, so setting the derivative equal to zero gives $\widehat{\lambda}=N/E$. The second derivative is $-N/\lambda^2<0$, confirming a maximum.

The zero-count case must be handled separately. If $N=0$ and $E>0$, the likelihood is proportional to $e^{-\lambda E}$, which decreases as $\lambda$ increases over $\lambda\geq 0$. Its maximum is at the boundary $\widehat{\lambda}=0$. Thus the total-count-over-total-exposure formula gives the maximum-likelihood estimate in this case as well, but the derivative argument for a positive interior maximum does not apply.

Result: If the counts are independent Poisson variables with a common rate per exposure unit, then their sum is Poisson with mean $\lambda E$, where $E=\sum_je_j$. The pooled frequency $\sum_jN_j/E$ is the maximum-likelihood estimate of that common rate, including the zero-count boundary case when $E>0$.

Check the Exposure Basis Before Combining Data

Exposure is not just a number to add. Its unit must match across observations. For example, a count measured per vehicle-year cannot be combined directly with exposure measured in vehicle-months until the latter is converted. The same principle applies to earned car-years, policy-years, person-years, or another relevant exposure base: identify the unit first, then express all observations on that basis.

If a limited-fluctuation count standard is being applied, use the pooled claim count as the count volume, provided the standard’s assumptions are appropriate. Do not substitute the total number of exposure units for the number of claims. Exposure determines the estimated rate’s denominator; it is not itself the observed claim count.

Worked Example: Pooling After Converting a Time Unit

A hypothetical portfolio has 21 claims over 700 life-years in one observation and 24 claims over 900 life-years in another. A third observation has 15 claims among 800 covered lives for six months. Assume the three observations are comparable and that a stated full-credibility standard requires 100 claims. Find the pooled frequency and the partial credibility under the square-root rule.

Method: convert exposure, pool counts and exposure, then apply the stated count standard. The third observation covers each life for half a year, so its exposure is $800(0.5)=400$ life-years. Total exposure is $700+900+400=2{,}000$ life-years. Total claims are $21+24+15=60$.

\widehat{\lambda}_{\text{pool}} =\frac{60}{2{,}000} =0.030\ \text{claims per life-year}.

The pooled count is 60. With the hypothetical full-credibility count $N_0=100$, the square-root rule gives

Z=\min\left(1,\sqrt{\frac{60}{100}}\right) =\sqrt{0.6} =0.7746.

The result is below one because 60 claims are below the stated standard. Common mistake: treating the third observation as 800 life-years would produce an incorrect total exposure of 2,400 and an incorrect frequency of $60/2{,}400=0.025$. The correction is to convert six months per life to one-half of a life-year before pooling. The credibility calculation uses 60 claims, not 2,000 exposure units.

Homogeneity: When Pooling Is and Is Not Appropriate

“Homogeneous” does not mean that every group has exactly the same observed frequency. Random variation will usually produce different observed rates, especially when counts are small. It means that the groups are suitable for estimating a common underlying rate on the chosen basis. That judgment requires considering the purpose of the estimate as well as the data.

Same claim definition and coverage. Counts should refer to the same type of claim, policy coverage, and treatment of claim events. For example, counts from contracts with materially different deductibles may not measure the same frequency of reported claims.
Comparable risk characteristics or a defined common target. If one group has a systematically different expected frequency, pooling it with another does not estimate either group’s rate. It estimates a rate for the combined exposure mix.
Comparable time basis. Periods may differ because of changes in exposure, seasonality, frequency trend, underwriting, or claims handling. Before combining them for a particular purpose, address material changes or define the estimand as the rate for the combined periods as observed.
Compatible exposure measurement. The exposure units and the way exposure is earned or measured must be consistent.
Appropriate variability assumptions. The independent Poisson model assumes variance equal to the mean for each count. Dependence, clustering, or extra-Poisson variation can make a Poisson-based credibility calculation too optimistic even if the groups have similar mean rates.

A pooled rate can be valid for a portfolio’s actual combined mix while being inappropriate as the rate for one of its component classes. If the intended estimate is for a fixed mix, use a consistent target mix or otherwise adjust for the composition change; do not confuse a composition-weighted portfolio frequency with a standardized class rate.

Worked Example: A Changing Mix of Risk Classes

A hypothetical insurer has low-frequency and high-frequency vehicle classes. Their expected frequencies are $0.020$ and $0.080$ claims per vehicle-year, respectively. In Period A, the low-frequency class has 18 claims over 900 vehicle-years and the high-frequency class has 8 claims over 100 vehicle-years. In Period B, the corresponding observations are 2 claims over 100 vehicle-years and 72 claims over 900 vehicle-years. Calculate the period and overall raw rates, then decide whether the overall rate estimates either class’s frequency.

Method: inspect rates within class before pooling across classes. In Period A, total claims are $18+8=26$ and total exposure is $900+100=1{,}000$, giving $26/1{,}000=0.026$. In Period B, the totals are $2+72=74$ claims and $100+900=1{,}000$ vehicle-years, giving $74/1{,}000=0.074$.

Across both periods, there are $26+74=100$ claims and $2{,}000$ vehicle-years. The raw overall rate is $100/2{,}000=0.050$. But the class-specific calculations are:

\text{Low-frequency class: }\frac{18+2}{900+100}=\frac{20}{1{,}000}=0.020, \qquad \text{High-frequency class: }\frac{8+72}{100+900}=\frac{80}{1{,}000}=0.080.

The class rates are stable across the periods; the overall period rates differ because the exposure mix changes. In each period, the raw combined rate reflects that period’s mix. The overall $0.050$ is the rate for the combined exposure as observed, with equal total exposure in the two classes. It is not the rate for either class. Common mistake: treating the overall $0.050$ as a homogeneous estimate for the low-frequency class, or interpreting the period-rate change as a change in both class rates. The correction is to retain the class distinction when estimating class rates. Pool within a class when its observations are comparable.

A Practical Pooling Sequence

A disciplined sequence helps prevent a precise-looking pooled rate from answering the wrong question.

Define the target rate.
Specify the risk population, claim definition, coverage, and exposure unit whose frequency is to be estimated.

Check comparability.
Review risk mix, time period, coverage, claim recording, and exposure measurement. Separate or adjust observations that do not represent the same target.

Put exposure on one basis.
Convert all periods and groups to the same unit before adding their exposures.

Combine counts and exposures.
Calculate total claims divided by total exposure. Use the pooled count, not the exposure total, if applying a claim-count credibility standard.

Interpret the result for its target.
State whether it is a class rate or a rate for the observed combined mix, and verify that the statistical assumptions suit any credibility calculation.

The main diagnostic is whether pooling preserves the question being answered. Combining genuinely comparable experience increases the volume supporting a common estimate. Combining unlike experience may instead hide class differences or changes in mix. This distinction is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.

Check Your Understanding

For each question, assume pooling is appropriate unless the problem asks you to assess that assumption.

A hypothetical coverage has 14 claims over 500 exposure units in one period and 51 claims over 1,500 exposure units in another. Find the pooled frequency and the unweighted average of the two period frequencies. (Answer: pooled frequency $65/2{,}000=0.0325$; unweighted average $(0.028+0.034)/2=0.031$.)
A group has 10 claims over 400 policy-years. Another has 12 claims among 600 policies covered for three months each. Find total exposure and pooled frequency in claims per policy-year. (Answer: total exposure $400+150=550$ policy-years; frequency $22/550=0.0400$.)
A pooled homogeneous count is 81 claims, and a stated full-credibility standard is 144 claims. Under the square-root rule, what is the partial credibility? (Answer: $Z=\sqrt{81/144}=0.7500$.)
Two comparable observations have zero claims over 400 exposure units and zero claims over 600 exposure units. Find the pooled maximum-likelihood frequency and state why a zero estimate does not prove that the true rate is zero. (Answer: $0/1{,}000=0$; it is the boundary maximum-likelihood estimate from these observations, but sampling variation permits a positive underlying rate.)
Two risk classes have frequencies of $0.010$ and $0.060$ claims per unit. A period contains 800 exposure units from the first class and 200 from the second. Using those class frequencies, find the combined expected frequency and explain whether it estimates the first class’s rate. (Answer: $(800(0.010)+200(0.060))/1{,}000=0.020$; it is the combined-mix rate, not the first class’s $0.010$ rate.)

What You'll Learn