Why Model Aggregate Loss Directly?
A pure premium is expected loss per unit of exposure. One way to estimate it is to model claim frequency and claim severity separately, then combine their predictions under appropriate conditions. An alternative is to model each policy period’s aggregate loss per exposure unit directly. This approach targets the expected loss cost without requiring claim counts or claim-level severities as separate responses.
The aggregate approach is useful when the rating target is loss per exposure and policy-period totals are available. It also requires careful construction of the response: losses must refer to a consistent coverage, valuation basis, and policy period, and each response must be paired with the exposure that generated it.
The loss definition is part of the model specification. For example, \(L_i\) might be incurred losses valued at a specified date, net of a stated deductible and subject to a stated limit. If the data mix paid and incurred amounts, different development bases, or different coverage terms without adjustment, the response will not represent one consistent quantity.
Response, Link, and Distribution
An aggregate pure premium GLM models \(Y_i\), not each claim within the policy period. Its linear predictor is \(\eta_i=x_i^\mathsf{T}\beta\), and a link function \(g\) connects that predictor to the expected pure premium:
With an identity link, the modeled mean is additive in the predictors. With a log link, \(\log(\mu_i)=x_i^\mathsf{T}\beta\), so changes in predictors act multiplicatively on the expected pure premium. Under the log link, a coefficient \(\beta_j\) for a one-unit increase in \(x_j\) corresponds to a mean ratio of \(\exp(\beta_j)\), holding other predictors fixed.
The response distribution and link must make sense together. Aggregate losses are nonnegative and can be highly right-skewed, with many zero-loss periods. A Gaussian model can be a useful mean-model approximation in some settings, but its distribution allows negative values and may be unsuitable when that feature matters. A Gamma model is defined for strictly positive responses, so it does not directly accommodate zero-loss policy periods. The family must be selected to match the observed response and the purpose of the model; a link alone does not determine a valid distribution.
The aggregate model does not assert that claim frequency and severity are irrelevant. Rather, their combined effect is represented in the conditional mean of \(Y_i\). If the data contain only policy-period totals, the model can estimate that mean without observing or fitting a separate claim count and claim severity model.
Why Exposure Matters in Fitting
A policy period with more exposure generally contains more information about the underlying expected loss cost than a much shorter period. A common GLM specification represents this by assuming \(\operatorname{Var}(Y_i\mid x_i)=\phi V(\mu_i)/e_i\), where \(V(\mu_i)\) is the variance function for the selected model and \(\phi\) is a dispersion parameter. This is a modeling assumption, not an automatic fact for every exposure measure or data set.
Under this variance structure, the corresponding GLM or quasi-likelihood score gives greater influence to records with more exposure. For a set of observations sharing the same modeled mean, the exposure-weighted estimate of that mean is total loss divided by total exposure. The derivation below shows why.
Here the common positive factor \(1/V(\mu)\) cancels from the estimating equation. Since \(e_iY_i=L_i\), the fitted mean is aggregate loss divided by aggregate exposure. This also explains why simply averaging each policy’s \(Y_i\) can give a different estimate: that average gives equal weight to a short-exposure and a long-exposure record.
In implementation, the analyst must ensure the software’s weight convention matches the intended variance model. Entering exposure as a weight is not interchangeable with using an unweighted average, and software can distinguish frequency weights from prior or variance weights. The assumed relationship between exposure and response variance should be stated and checked rather than inferred merely from the fact that exposure appears in the data.
Worked Example: Exposure-Weighted Pure Premium
A hypothetical insurer has four policy-period records. Losses are on the same valuation and coverage basis, and exposure is measured in vehicle-years. Find the overall aggregate pure premium and compare it with the unweighted average of the four policy-level responses. Assume a common mean pure premium for the records; no interest or mortality assumptions apply.
| Record | Loss \(L_i\) | Exposure \(e_i\) | Response \(Y_i=L_i/e_i\) |
|---|---|---|---|
| A | $900 | 0.5 | $1,800 |
| B | $1,200 | 1.0 | $1,200 |
| C | $3,000 | 2.0 | $1,500 |
| D | $400 | 0.5 | $800 |
Method. The exposure-weighted common mean is total loss divided by total exposure. First, the responses are \(900/0.5=1{,}800\), \(1{,}200/1=1{,}200\), \(3{,}000/2=1{,}500\), and \(400/0.5=800\) dollars per vehicle-year.
Total loss is \(900+1{,}200+3{,}000+400=\$5{,}500\). Total exposure is \(0.5+1+2+0.5=4\) vehicle-years. Therefore the exposure-weighted pure premium is \(\$5{,}500/4=\$1{,}375.00\) per vehicle-year.
The unweighted average of the four responses is \((1{,}800+1{,}200+1{,}500+800)/4=\$5{,}300/4=\$1{,}325.00\) per vehicle-year. The difference arises because the unweighted average gives the half-year records as much influence as the two-year record. The check is that \(4\times \$1{,}375=\$5{,}500\), the observed total loss. The common mistake is averaging policy responses equally when the target is the exposure-weighted aggregate rate.
Worked Model With Rating Classes
A direct aggregate model can include classification variables just as other GLMs do. In the next example, the response is positive for every policy, so a Gamma mean model is a suitable illustration. The Gamma assumption is explicit: observations are independent conditional on rating class, and the response variance is \(\phi\mu_i^2/e_i\), with a common positive dispersion \(\phi\). The log link makes the class comparison multiplicative.
Worked Example: Estimating Class Pure Premiums With a Log Link
A hypothetical portfolio has the following policy-period totals, all strictly positive. Fit a saturated two-class Gamma aggregate model with log link and exposure variance weights. Estimate the class means, the class coefficient, and expected loss for a new Class B risk with 1.5 exposure units. No interest or mortality assumptions apply.
| Class | Policy period | Loss \(L_i\) | Exposure \(e_i\) | Response \(Y_i\) |
|---|---|---|---|---|
| A | 1 | $900 | 1 | $900 |
| A | 2 | $2,400 | 2 | $1,200 |
| B | 1 | $1,600 | 1 | $1,600 |
| B | 2 | $2,000 | 1 | $2,000 |
Step 1: Estimate Class A’s mean. Its total loss is \(\$900+\$2{,}400=\$3{,}300\), and its total exposure is \(1+2=3\). Thus \(\widehat{\mu}_A=\$3{,}300/3=\$1{,}100\) per exposure unit.
Step 2: Estimate Class B’s mean. Its total loss is \(\$1{,}600+\$2{,}000=\$3{,}600\), and its total exposure is \(1+1=2\). Thus \(\widehat{\mu}_B=\$3{,}600/2=\$1{,}800\) per exposure unit.
Step 3: Express the estimates as GLM coefficients. Let \(I_B=1\) for Class B and \(0\) for Class A, with \(\log(\mu)=\beta_0+\beta_1 I_B\). Then \(\widehat{\beta}_0=\log(1{,}100)\), and \(\widehat{\beta}_1=\log(1{,}800/1{,}100)=\log(18/11)\), approximately \(0.4925\). The multiplier for Class B relative to Class A is \(\exp(\widehat{\beta}_1)=18/11\), approximately \(1.636\).
Step 4: Predict the new policy period. Its expected pure premium is \(\$1{,}800\) per exposure unit. At 1.5 exposure units, its expected aggregate loss is \(1.5\times \$1{,}800=\$2{,}700\). The check is that multiplying the estimated Class B rate by the stated exposure gives the predicted loss. The common mistake is to use the response total as the rate without dividing by exposure.
Prediction and Interpretation
The model prediction \(\widehat{\mu}_i\) is an estimated expected loss per exposure unit for a policy with the specified characteristics. To obtain expected aggregate loss for exposure \(e_i\), multiply the mean prediction by that exposure:
For a log-link model, coefficient interpretation is on a ratio scale. For instance, a coefficient of \(\log(1.2)\) for a binary rating factor implies a modeled pure premium 1.2 times as large when that factor is present, all else held constant. It does not mean that the aggregate loss rises by 20 percent independently of exposure: predicted aggregate loss also depends on the policy’s exposure.
Worked Example: Applying Aggregate Model Factors
A hypothetical fitted aggregate model has a base pure premium of $500 per exposure unit. A territory factor of 1.20 and a vehicle-use factor of 0.90 apply multiplicatively. A new policy has 0.75 exposure units. Calculate its modeled expected pure premium and aggregate loss. Treat the given fitted factors as fixed; no separate claim-frequency or severity distributions are assumed, and interest and mortality are not applicable.
Step 1: Apply the factors to the base rate. The expected pure premium is \(\$500\times1.20\times0.90\). First, \(\$500\times1.20=\$600\); then \(\$600\times0.90=\$540\) per exposure unit.
Step 2: Scale to the policy’s exposure. Expected aggregate loss is \(0.75\times \$540=\$405.00\).
Check. The factors produce a rate 1.08 times the base, since \(1.20\times0.90=1.08\), and \(\$500\times1.08=\$540\). Multiplying by 0.75 gives \(\$405\). The common mistake is to apply exposure to one rating factor or to the base but not to the final pure premium.
What the Aggregate Model Does—and Does Not—Provide
The direct model’s main advantage is alignment between the response and the pricing target: loss per exposure unit. It can use policy-period data even when individual claim detail is unavailable or when a separate frequency-severity decomposition is not desired. It also estimates the combined mean effect of claim occurrence and claim size as reflected in aggregate losses.
The model does not, by itself, estimate expected claim counts, average claim sizes, or the distribution of claim-level severity. Two portfolios can have the same expected aggregate loss per exposure while having very different claim frequencies and severities. If those components are needed for another decision, an aggregate mean model alone does not supply them.
Nor does a mean prediction describe uncertainty completely. A chosen distribution and variance structure determine how the model treats variability around the mean, including the implications of zero-loss periods and large losses. Residual checks should be performed at the policy-period response level, and predicted rates should be compared with observed aggregate losses by relevant rating groups. A model that predicts the portfolio total well can still miss important class-level patterns.
For an interpretation aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus, state the response, the exposure unit, the conditional mean, and the model assumptions. Then explain whether the target is a rate per unit or an expected total for a particular policy period. That discipline prevents most specification and interpretation errors.
Check Your Understanding
Work each problem by identifying the aggregate response and the exposure basis before calculating.
- A hypothetical policy has $2,100 of aggregate loss over 1.5 exposure units. What is its observed pure premium response? (Answer: $1,400.00 per exposure unit)
- Two policy periods have losses of $700 and $2,400, with exposures of 0.5 and 1.5 units. Find the exposure-weighted common pure premium, then the unweighted average of the two policy responses. (Answer: weighted $1,550.00; unweighted $1,500.00 per exposure unit)
- A log-link aggregate model predicts a pure premium of $820 per exposure unit for a policy with 0.4 exposure units. What is its predicted expected aggregate loss? (Answer: $328.00)
- A hypothetical positive-response Gamma model has two records in one class: losses of $1,200 at exposure 1 and $2,100 at exposure 2. Under the exposure-weighted common-mean fit, what is the estimated class pure premium? (Answer: $1,100.00 per exposure unit)
- A fitted log-link model gives a base rate of $600, a factor of 1.25 for a selected territory, and a factor of 0.80 for vehicle use. For a policy with 0.5 exposure units, calculate the expected pure premium and aggregate loss. (Answer: $600.00 per exposure unit and $300.00 expected aggregate loss)