Estimating Poisson Frequency From Observed Counts

Why Pool Counts and Exposure?

When a portfolio is divided into groups, each group may have a different amount of exposure and a different observed claim count. If the groups are assumed to share one underlying claim frequency, the estimate should give more weight to groups representing more exposure. The maximum likelihood method makes this weighting precise: under the common-rate Poisson model, the estimate is total claims divided by total exposure.

The result depends on the model, not just on an arithmetic convention. In particular, exposures must be measured in consistent units, and the groups must be treated as having the same frequency per exposure unit. The calculation does not establish that the common-frequency assumption is appropriate; it estimates the rate conditional on that assumption.

Definition: Suppose group $j$ has known exposure $e_j$ and observed claim count $n_j$. A common Poisson frequency model assumes that, conditional on the exposures, the counts $N_j$ are independent and $$ N_j \sim \operatorname{Poisson}(\lambda e_j), $$ where $\lambda$ is the common expected number of claims per exposure unit. Exposures are nonnegative and expressed in the same units. For groups included in a log-likelihood calculation, require $e_j>0$. A zero-exposure group with zero claims may be omitted; a zero-exposure group with a positive count has probability zero under the model.

Exposure might be earned policy-years, vehicle-years, or another denominator appropriate to the portfolio. If exposure is measured in hundreds of policy-years, then $\lambda$ is claims per hundred policy-years—not claims per individual policy-year. The units of the reported frequency always follow from the units used for exposure.

Deriving the Maximum Likelihood Estimate

For a Poisson random variable with mean $\lambda e_j$, the probability of observing $n_j$ claims in group $j$ is $\exp(-\lambda e_j)(\lambda e_j)^{n_j}/n_j!$. By independence, the likelihood for all included groups is the product of these probabilities. Assume first that every included group has positive exposure and that $\lambda>0$, so the logarithms below are defined.

L(\lambda) =\prod_{j=1}^{m} \frac{\exp(-\lambda e_j)(\lambda e_j)^{n_j}}{n_j!}

Taking the logarithm converts the product into a sum. Because every $e_j>0$, the term $\log e_j$ is defined, including when $n_j=0$; its coefficient is then zero.

\ell(\lambda) =-\lambda\sum_{j=1}^{m}e_j +\left(\sum_{j=1}^{m}n_j\right)\log\lambda +\sum_{j=1}^{m}n_j\log e_j -\sum_{j=1}^{m}\log(n_j!)

Only the first two terms depend on $\lambda$. Differentiating and setting the score equal to zero gives the estimate when there is at least one claim and positive total exposure.

\frac{d\ell}{d\lambda} =-\sum_{j=1}^{m}e_j+\frac{\sum_{j=1}^{m}n_j}{\lambda}=0 \quad\Longrightarrow\quad \widehat{\lambda} =\frac{\sum_{j=1}^{m}n_j}{\sum_{j=1}^{m}e_j}

The second derivative is $-(\sum_j n_j)/\lambda^2$, which is negative when the total count is positive. Thus the stationary point is a maximum. Equivalently, the log-likelihood rises before that point and falls after it. If the total observed count is zero and the parameter space includes $\lambda=0$, the likelihood is maximized at $\widehat{\lambda}=0$; the displayed log-likelihood derivation applies for $\lambda>0$, with the boundary handled separately.

Why the estimator is pooled: The likelihood’s exposure term is $-\lambda\sum_j e_j$, while its count term involving the rate is $(\sum_j n_j)\log\lambda$. Consequently, the score depends on the grouped data through total claims and total exposure. Dividing one group’s rate by the number of groups, or taking an unweighted mean of group rates, does not generally maximize this likelihood.

Worked Applications

Worked Example: Estimating Frequency Across Three Exposure Groups

A hypothetical insurer observes claim counts for three groups of comparable policies. The groups are assumed to share one Poisson frequency per policy-year. Estimate that frequency and calculate the fitted expected count for each group.

Group	Exposure (policy-years)	Observed claims
A	100	12
B	150	21
C	250	27

Method. Use the maximum likelihood estimator for independent Poisson counts with a common frequency and known exposures. First sum the counts and exposures: $12+21+27=60$ claims, and $100+150+250=500$ policy-years.

The estimated frequency is $\widehat{\lambda}=60/500=0.12$ claims per policy-year.

The fitted expected count in each group is its exposure times the estimated rate. Group A has $100(0.12)=12$ expected claims; group B has $150(0.12)=18$; and group C has $250(0.12)=30$. The fitted total is $12+18+30=60$, equal to the observed total, as the likelihood estimate requires.

Check and trap avoided. The rate has policy-years in its denominator, so it is not 12 claims per policy-year; it is $0.12$ per policy-year (or 12 per 100 policy-years). The fitted counts sum correctly to 60, and the calculation avoids confusing the total number of claims with a frequency.

Worked Example: A Zero-Exposure Group

A hypothetical specialty portfolio has the following grouped experience. Estimate the common frequency per insured-unit-year, and determine whether every row belongs in the log-likelihood expansion.

Group	Exposure (insured-unit-years)	Observed claims
A	0	0
B	80	4
C	120	14

Method. Under the model, a group with zero exposure has Poisson mean $\lambda(0)=0$. Its claim count must therefore be zero with probability one. Group A contributes a likelihood factor of 1 and can be omitted. In particular, it must not be inserted into the expanded log-likelihood as $0\log 0$, which is undefined.

For groups B and C, total claims are $4+14=18$, and total exposure is $80+120=200$ insured-unit-years. Thus $\widehat{\lambda}=18/200=0.09$ claims per insured-unit-year.

The fitted expected counts are $80(0.09)=7.2$ for B and $120(0.09)=10.8$ for C; together they equal 18. Group A has fitted mean zero, consistent with its zero exposure and zero claims.

Check and trap avoided. The zero-exposure, zero-claim row neither adds exposure nor claims, so omitting it leaves the estimate unchanged. If that row instead had even one observed claim, its likelihood under this model would be zero for every value of $\lambda$: the data would contradict the stated exposure/count model, not provide a rate contribution.

Worked Example: Why an Unweighted Average Is Not the Pooled Estimate

A hypothetical portfolio has three classes, all assumed to have a common Poisson frequency per policy-year for this calculation. Find the pooled maximum likelihood estimate, compare it with the unweighted average of the class-specific observed rates, and estimate expected claims for a future 80 policy-years.

Class	Exposure (policy-years)	Claims
1	40	2
2	60	9
3	100	19

Method. The class-specific observed rates are $2/40=0.05$, $9/60=0.15$, and $19/100=0.19$ claims per policy-year. Their unweighted average is $(0.05+0.15+0.19)/3=0.13$.

That average gives each class equal weight even though the classes have different exposures. The likelihood estimate instead uses the pooled totals: $2+9+19=30$ claims and $40+60+100=200$ policy-years. Therefore, $\widehat{\lambda}=30/200=0.15$ claims per policy-year.

For a future 80 policy-years under the same assumed frequency, the fitted expected count is $80(0.15)=12$ claims. As a check, applying the estimate to the observed total exposure gives $200(0.15)=30$ fitted claims.

Check and trap avoided. The value $0.13$ is the arithmetic mean of the three class rates, not the common-rate Poisson maximum likelihood estimate. The pooled value $0.15$ weights each class rate in proportion to exposure: $(40(0.05)+60(0.15)+100(0.19))/200=30/200=0.15$.

Conditions, Units, and Boundary Cases

The estimator is simple, but it is meaningful only when its modeling conditions and data definitions are clear. The groups must be treated as independent conditional on their exposures, and the expected count in each group must be proportional to exposure with the same $\lambda$. The exposure denominator must represent comparable risk time or another consistently defined exposure measure. If groups truly have different frequencies, the common-rate model is a simplifying assumption; the pooled estimate is then an exposure-weighted overall rate, not a claim that every group has the same observed rate.

A zero-exposure row needs explicit treatment. If $e_j=0$ and $n_j=0$, then $P(N_j=0)=1$, so the row can be dropped from the likelihood. If $e_j=0$ and $n_j>0$, then $P(N_j=n_j)=0$ under the model. Do not use the log-likelihood expansion with such a row or attempt to repair it by evaluating $\log 0$; investigate the exposure definition, count assignment, or model instead.

If total exposure is zero, the ratio estimator is undefined. If total exposure is positive but there are no claims, the maximum likelihood estimate on the nonnegative parameter space is zero. If a problem restricts the rate to be strictly positive, there is no positive interior maximizer in that all-zero-count case; the likelihood approaches its supremum as the rate approaches zero.

Check the units and rows.
Confirm that exposures use one consistent denominator. Omit zero-exposure rows only when their observed counts are also zero; a positive count at zero exposure is incompatible with the model.

Pool the data.
Add the included claim counts and exposures separately. Do not average class rates unless the task specifically asks for an unweighted average.

Calculate and interpret.
Divide total claims by total exposure, state the frequency units, and multiply by exposure to check fitted expected counts.

For exam-style calculations, a common error is to divide total claims by the number of groups, or to average rates without accounting for different exposure volumes. Another is to change exposure units without changing the frequency units: if exposure is converted from policy-years to hundreds of policy-years, the numeric rate must be reported per hundred policy-years. A final check is dimensional: claims divided by policy-years yields claims per policy-year, while claims divided by vehicle-years yields claims per vehicle-year.

Key takeaway: For independent Poisson group counts with known exposures and a common frequency, the maximum likelihood estimate is total claims divided by total exposure. Apply that result only after checking the common-rate assumptions, exposure units, and zero-exposure rows.

Check Your Understanding

For each problem, use the common Poisson frequency model only where the stated conditions permit it. State the rate in the requested exposure units.

A hypothetical delivery fleet reports 7 claims over 70 vehicle-years and 15 claims over 180 vehicle-years. Estimate the common frequency per vehicle-year. (Answer: $22/250=0.088$ claims per vehicle-year.)
A hypothetical equipment portfolio has 0 claims and 0 exposure in one reporting cell, 3 claims over 50 machine-years in another, and 9 claims over 100 machine-years in a third. Find the pooled estimate and state how the first cell is handled. (Answer: Omit the zero-exposure, zero-claim cell; $12/150=0.08$ claims per machine-year.)
Two hypothetical districts have 8 claims over 40,000 household-years and 12 claims over 60,000 household-years. Express the pooled estimate per 10,000 household-years. (Answer: $20/100{,}000=0.0002$ per household-year, or 2 claims per 10,000 household-years.)
A hypothetical portfolio has group rates of $6/30$ and $15/120$ claims per policy-year. Find the common-rate estimate and explain why the unweighted mean of those two rates is not the maximum likelihood estimate. (Answer: $(6+15)/(30+120)=21/150=0.14$ claims per policy-year; the unweighted mean $0.1625$ gives the groups equal weight despite unequal exposure.)
A hypothetical portfolio records 0 claims over 90 policy-years and 0 claims over 110 policy-years. On the nonnegative rate parameter space, what is the maximum likelihood estimate, and what if total exposure were zero instead? (Answer: With 200 policy-years and zero claims, $\widehat{\lambda}=0$; with zero total exposure, the ratio is undefined and the data provide no exposure-based frequency estimate.)

What You'll Learn