Two Different Roles for Counts and Exposure
A ratemaking data set may contain both the number of claims and the amount of time or risk exposure represented by a record. These quantities are related to claims, but they do not play the same modeling role. In a grouped severity model, a claim count can tell us how much information is represented by an observed average severity. In a frequency model, exposure scales the expected claim count. Confusing the two changes the quantity being modeled.
The distinction is especially important when the data have been aggregated. A row might report an average severity based on several claims, or it might report a claim count over several policy-years. The first is a severity response with a claim-count weight; the second is a frequency response modeled relative to exposure. A weight does not automatically become an offset, and an offset does not automatically become a weight.
Why a Claim Count Weights a Grouped Severity
Suppose group \(j\) contains \(n_j\) claim severities \(X_{j1},\ldots,X_{jn_j}\), and the recorded response is their arithmetic average \(\bar X_j\). For this to be a valid summary for a common severity model, the claims in the group must have the same modeled conditional distribution, or at least the same conditional mean and dispersion structure specified by that model. The grouping must not hide material differences in predictors that the model is meant to represent.
For example, in a Gamma GLM with mean \(\mu_j\) and dispersion \(\phi\), the variance function is \(\operatorname{Var}(X_{jk})=\phi\mu_j^2\). If claims are conditionally independent within a group, the variance of the group average is smaller when more claims contribute to it.
The second equality uses conditional independence, which makes the covariance terms zero. Thus \(n_j\) claims produce an average with \(1/n_j\) of the variance of a single claim. In a Gamma GLM for the grouped average, the claim count therefore acts as a precision or frequency weight: larger groups provide more information about their conditional mean. With an inverse Gaussian severity model, \(\operatorname{Var}(X_{jk})=\phi\mu_j^3\), and the same independence argument gives \(\operatorname{Var}(\bar X_j)=\phi\mu_j^3/n_j\). The count again scales the information in the group average.
Equivalently, when the model and grouping assumptions hold, a group of \(n_j\) claims contributes the same claim-level likelihood information as \(n_j\) individual observations with the same modeled predictors, with the grouped data retaining their average. This is why grouped severity fits commonly use claim counts as frequency weights. This interpretation requires that the response really is an average of \(n_j\) relevant claims. A count of policies, days, or exposure units is not a substitute for that claim count.
Worked Example: Weighting Two Grouped Severity Averages
A hypothetical insurer has grouped claim severities for two risk cells. Cell A has 2 claims and an average severity of $1,000. Cell B has 6 claims and an average severity of $1,500. Fit an intercept-only Gamma severity model with a log link using the grouped means and claim-count frequency weights. State the fitted mean and compare it with the unweighted result.
Method. For an intercept-only Gamma model, the fitted mean under the claim-level likelihood is the arithmetic average of all claim severities. From grouped means, the equivalent calculation is the claim-count-weighted average.
Cell A represents \(2\times \$1{,}000=\$2{,}000\) of total severity. Cell B represents \(6\times \$1{,}500=\$9{,}000\). The total represented severity is \(\$2{,}000+\$9{,}000=\$11{,}000\), and the total number of claims is \(2+6=8\). Therefore,
As a check, the fitted average lies between the two cell averages and is closer to $1,500 because Cell B contains more claims. If the two cell averages were incorrectly given equal weight, the result would be \((\$1{,}000+\$1{,}500)/2=\$1{,}250\). That treats two claims and six claims as equally informative groups, so it does not reproduce the claim-level average.
Common mistake avoided: the weights are the numbers of claims summarized by the severity averages, not the exposures that generated those claims.
Why Exposure Is an Offset in a Frequency Model
Let \(N_i\) be the claim count in record \(i\), and let \(e_i\) be its exposure measured in a consistent unit, such as policy-years. A standard Poisson frequency model assumes, conditional on the predictors, that \(N_i\) is Poisson with mean \(e_i\lambda_i\), where \(\lambda_i\) is the expected claim rate per exposure unit. If the rate has a log-linear predictor, the model is
The term \(\log(e_i)\) is an offset: it enters the linear predictor with its coefficient fixed at 1. This specification says that doubling exposure doubles the expected number of claims, holding the modeled rate constant. The offset gives the model the correct expected-count scale while allowing the predictors to explain differences in rates. The usual interpretation assumes exposure is measured consistently and that expected claims are proportional to exposure over the range being modeled.
For a common rate \(\lambda\) across records, the Poisson likelihood is proportional to \(\prod_i(e_i\lambda)^{N_i}\exp(-e_i\lambda)/N_i!\). Differentiating its log likelihood with respect to \(\lambda\) gives \(\sum_i N_i/\lambda-\sum_i e_i=0\). Hence the maximum-likelihood estimate is total claims divided by total exposure, not average claims per record. This simple case shows why exposure belongs in the expected count.
Worked Example: Estimating a Frequency Rate With Unequal Exposure
Two hypothetical groups have the same modeled claim rate. Group A has 2 policy-years of exposure and 4 claims; Group B has 8 policy-years and 8 claims. Under a Poisson model with a common rate and an exposure offset, estimate the rate and calculate the expected count for each group.
Method. With a common rate, the offset model gives an expected count of \(e_i\lambda\) for group \(i\). The estimate is total claims divided by total exposure:
For Group A, the fitted expected count is \(2(1.2)=2.4\) claims. For Group B, it is \(8(1.2)=9.6\) claims. These fitted counts sum to \(2.4+9.6=12\), matching the observed total. The fitted rate is the same in both groups; their expected counts differ because their exposures differ.
A wrong calculation would treat the two observed counts as equally exposed and report \((4+8)/2=6\) claims per group. That is an average count per record, not a rate per policy-year, and it fails to account for the fourfold exposure difference. The offset corrects the modeled mean; it is not a severity weight.
Putting the Roles Together
A portfolio analysis can use both mechanisms, but in different equations. A frequency model uses exposure to model claim counts per exposure unit. A severity model uses the number of claims represented by each group average to reflect the information in that average. The frequency response is a count; the severity response is an amount per claim. Their units and likelihood contributions differ.
The next example keeps these roles separate, then checks that the resulting aggregate loss per exposure agrees with the direct total-loss calculation. All figures are hypothetical.
Worked Example: Separate Frequency Offsets From Severity Weights
A hypothetical portfolio has two classes. Class A has 100 policy-years, 20 claims, and an average claim severity of $800. Class B has 200 policy-years, 30 claims, and an average claim severity of $1,200. Estimate each class's frequency, calculate the claim-count-weighted portfolio severity, and compute the portfolio's aggregate loss per policy-year.
Step 1: Estimate class frequencies. Divide each claim count by its exposure. Class A has \(20/100=0.20\) claims per policy-year. Class B has \(30/200=0.15\) claims per policy-year. In a frequency GLM, these exposure amounts would enter through \(\log(100)\) and \(\log(200)\) offsets for the respective records.
Step 2: Weight severity by claims. Class A represents \(20\times \$800=\$16{,}000\) of claim severity; Class B represents \(30\times \$1{,}200=\$36{,}000\). Across \(20+30=50\) claims, the weighted average severity is
The weights are 20 and 30 because those are the claims represented by the severity averages. The exposures of 100 and 200 do not replace those weights.
Step 3: Calculate the aggregate loss per policy-year. The observed total severity amount is \(\$16{,}000+\$36{,}000=\$52{,}000\), and total exposure is \(100+200=300\) policy-years. Thus the aggregate loss per policy-year is \(\$52{,}000/300=\$173.333\ldots\), or \(\$173.33\) to the nearest cent. As a cross-check, the overall frequency is \(50/300=1/6\) claim per policy-year, and \((1/6)\times \$1{,}040=\$173.333\ldots\).
Common mistake avoided: weighting severity by exposure would produce \((100(800)+200(1200))/300=\$1{,}066.67\), which is not the average amount per claim. Exposure weights policy-years; claim counts weight claim severities.
Modeling Conditions and Exam Traps
The weights and offsets are not interchangeable because they encode different assumptions:
| Model component | Quantity represented | How it enters | Typical interpretation |
|---|---|---|---|
| Claim-count weight in grouped severity | Number of claim severities summarized by a group mean | Weights the severity likelihood or precision contribution | A group mean based on more claims is more informative |
| Exposure offset in frequency | Time or risk units over which claims could occur | Adds \(\log(e_i)\) to the log expected-count predictor with coefficient fixed at 1 | Expected claims scale with exposure at a given rate |
A claim-count weight is appropriate only if the group response and count match: for instance, an average of 30 claim severities paired with a weight of 30. A weight of 30 would be unjustified if the response were instead a single policy-level average, or if the underlying observations were not 30 comparable claims. Likewise, an exposure offset requires an exposure quantity that measures the opportunity for claims; using claim count as that exposure would make the predictor depend on the outcome it is meant to model.
A common exam error is to see a count in a data table and apply it mechanically. First identify what each row's response represents. If it is a grouped severity average, ask how many claims produced that average. If it is a frequency count, ask how much exposure generated that count. Then match the quantity to the model component. This distinction is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Determine whether the modeled value is a severity amount or a claim count.
For a severity average, find the number of claims represented. For a frequency count, find the exposure units represented.
Use claim count as a grouped-severity weight when the severity assumptions justify it; use exposure as a frequency offset when expected counts scale with exposure.
Severity weights should count claims, frequency rates should have units of claims per exposure, and fitted expected counts should scale with exposure.
Check Your Understanding
For each problem, identify the appropriate role of the count or exposure and calculate the requested quantity.
- A severity cell has 4 claims with an average severity of $900, and a second cell has 12 claims with an average severity of $1,100. Under an intercept-only grouped severity fit with claim-count weights, calculate the fitted mean. (Answer: $1,050.)
- A frequency record has 3 policy-years and 5 claims; another has 7 policy-years and 9 claims. Under a common-rate Poisson model with an exposure offset, calculate the estimated rate and the fitted expected count for the second record. (Answer: 1.4 claims per policy-year; 9.8 expected claims.)
- Two severity-group averages are $700 based on 5 claims and $1,300 based on 15 claims. Calculate the claim-count-weighted average severity, then state why an unweighted average is not equivalent. (Answer: $1,150; an unweighted average is $1,000 and gives the two group means equal influence despite different claim counts.)
- A portfolio class has 80 policy-years and 16 claims, with an average severity of $750. Another has 120 policy-years and 24 claims, with an average severity of $1,000. Calculate the two class frequencies, the claim-count-weighted overall severity, and total observed loss per policy-year. (Answer: 0.20 and 0.20 claims per policy-year; $900 average severity; $180 loss per policy-year.)
- A data table reports a claim count of 18, exposure of 60 policy-years, and a grouped average severity of $2,000 for one cell. State the frequency rate, the severity weight, and the expected count under the fitted frequency rate for this exposure. (Answer: 0.30 claims per policy-year; severity weight 18; expected count 18 claims.)