Exposure Scaling in Aggregate Pure Premium Models

Why Exposure Changes the Loss Prediction

A pure premium is an expected loss amount per unit of exposure. A policy-period prediction, by contrast, is the expected aggregate loss over that policy's actual exposure. Keeping those quantities separate is essential: an annual rate of $600 per vehicle-year does not predict $600 of loss for every policy, regardless of how long it was insured or how many vehicles it covered.

In an aggregate pure premium model, exposure scales the expected aggregate loss when the risk is otherwise held constant. If a risk has twice the exposure, the model predicts twice the expected loss, but the same expected loss per exposure unit. This statement relies on exposure being measured consistently and on the modeled risk characteristics and expected loss rate being stable over the exposure being scaled.

Definition: Let $e_i$ be the exposure for policy or observation $i$, measured in a stated unit, and let $S_i$ be its aggregate loss over that exposure. The observed pure premium is $S_i/e_i$ when $e_i>0$. The modeled pure premium $m_i$ is the conditional expected loss per exposure unit. Under linear exposure scaling, the expected aggregate loss is $$ E[S_i\mid x_i,e_i]=e_i m_i, $$ where $x_i$ denotes the modeled risk characteristics. Thus $m_i$ and $e_i m_i$ answer different questions: the former is a per-unit rate; the latter is a policy-period expected loss.

Exposure might be measured in vehicle-years, insured-location-years, payroll units, or another basis appropriate to the coverage. A fraction of a year is a valid exposure when the chosen unit is a year: six months of coverage for one vehicle is $0.5$ vehicle-years. The model's predictions are meaningful only when the exposure used in the calculation matches the exposure unit used to define the pure premium.

Writing Exposure into an Aggregate Mean Model

A common specification uses a log link for the conditional mean of aggregate loss. If $m_i$ is modeled by a linear predictor $x_i^\mathsf{T}\beta$, then $$ \log E[S_i\mid x_i,e_i]=\log(e_i)+x_i^\mathsf{T}\beta, $$ or equivalently, $$ E[S_i\mid x_i,e_i]=e_i\exp(x_i^\mathsf{T}\beta). $$ Here $\log(e_i)$ is an offset: its coefficient is fixed at one rather than estimated. The fitted pure premium per exposure unit is $\exp(x_i^\mathsf{T}\beta)$, while the fitted expected aggregate loss is $e_i\exp(x_i^\mathsf{T}\beta)$.

For example, if the fitted per-unit pure premium for a risk is $400 per vehicle-year, the expected aggregate loss for $0.75$ vehicle-years is \(0.75(400)=$300. The exposure offset does not change the risk's per-unit rate; it converts that rate into a mean for the observed exposure period.

The log-link form is useful because it keeps the predicted mean positive and makes multiplicative rate factors convenient. It is not the only way to state the mean relationship: the underlying exposure-scaling idea is $E[S_i\mid x_i,e_i]=e_i m_i$, provided the assumed risk rate is stable over the exposure. A model with a different link or response distribution requires its own appropriate specification; the log-offset equation should not be imposed without checking that it matches the intended model.

Model condition: Linear scaling is appropriate when $e_i$ measures comparable units of risk and the conditional expected loss per unit, given the modeled characteristics, does not change merely because the observation has more exposure. If risk composition, coverage terms, or conditions vary over the period, a single constant $m_i$ may not describe the whole exposure. Split the period or otherwise model the changing risk rather than treating the total exposure as interchangeable.

Why the Mean Scales Linearly

The scaling result follows from additivity of expected loss. Suppose, first, that exposure consists of $n$ equal units and that the loss contribution from each unit has conditional expected value $m$, given the stated risk characteristics. Let $L_j$ be the loss contribution from unit $j$. Aggregate loss is $S=\sum_{j=1}^{n}L_j$. By linearity of expectation,

E[S\mid x] =E\left[\sum_{j=1}^{n}L_j\middle|x\right] =\sum_{j=1}^{n}E[L_j\mid x] =\sum_{j=1}^{n}m =nm.

This argument does not require the unit losses to be independent; linearity of expectation holds whether or not they are dependent. It does require the same expected loss per unit across the units being added. For fractional exposure, the model extends this unit-rate relationship proportionally, as is standard when exposure represents a share of the corresponding risk period. This is a modeling assumption, not a guarantee that every short period will realize a proportional share of annual losses.

Identify the exposure basis.
State whether the rate is per vehicle-year, location-year, or another unit. Convert the policy period to that same unit.

Calculate the per-unit prediction.
Use the fitted risk factors or linear predictor to obtain $m_i$, with units of dollars per exposure unit.

Scale to the policy period.
Compute $e_i m_i$ to obtain expected aggregate loss in dollars for the stated exposure.

Check the units and interpretation.
The result should be a period total, not a per-unit rate. Confirm that exposure has not already been included in the prediction.

Worked Predictions

Worked Example: From a Log-Link Rate to a Policy-Period Mean

A hypothetical insurer models aggregate loss for a vehicle risk using a log link. The fitted coefficients are $\beta_0=\log(300)$ and $\beta_1=\log(1.25)$, where $x=1$ indicates a stated vehicle-use class and $x=0$ is the reference class. A policy in the indicated class has $0.6$ vehicle-years of exposure. Find its modeled pure premium per vehicle-year and expected aggregate loss. All amounts are in dollars.

Method. Calculate the per-unit mean from the risk linear predictor, then multiply by exposure once. The offset is not included in the per-unit rate.

The linear predictor for this class, excluding the exposure offset, is $$ x^\mathsf{T}\beta=\log(300)+1\log(1.25)=\log(375), $$ because $300(1.25)=375$. Therefore the modeled pure premium is $$ m=\exp(\log(375))=\$375\text{ per vehicle-year}. $$

Scale this rate by the policy's $0.6$ vehicle-years: $$ E[S\mid x,e]=0.6(375)=\$225. $$

Check. The units are vehicle-years times dollars per vehicle-year, giving dollars. The implied log mean for the aggregate is $\log(0.6)+\log(375)=\log(225)$, which also gives an expected aggregate loss of $225. A common mistake is to report $375 as the policy-period expected loss; that is the annualized per-unit prediction, not the total for $0.6$ vehicle-years.

Worked Example: Combining Expected Losses Across Two Groups

A hypothetical property portfolio has two groups. Group A has $120$ insured-location-years and a modeled pure premium of $480 per location-year. Group B has $75$ insured-location-years and a modeled pure premium of $620 per location-year. Find each group's expected aggregate loss, the portfolio expected aggregate loss, and the portfolio pure premium per location-year.

Method. Scale each group's rate by its own exposure, add the expected losses, and divide the combined expected loss by the combined exposure. The portfolio rate is exposure-weighted; it is not the simple average of the two group rates.

For Group A, $$ E[S_A]=120(480)=\$57{,}600. $$ For Group B, $$ E[S_B]=75(620)=\$46{,}500. $$ Thus the portfolio expected aggregate loss is $$ E[S_A+S_B]=57{,}600+46{,}500=\$104{,}100, $$ and total exposure is $$ 120+75=195\text{ location-years}. $$ The portfolio pure premium is therefore $$ \frac{104{,}100}{195}=\$533.846\ldots\text{ per location-year}. $$ Rounded to the nearest cent, the answer is $533.85 per location-year.

Check. The weighted rate lies between the group rates of $480 and $620, as it should. The unweighted average, $(480+620)/2=$550$, would incorrectly give each group equal influence despite their different exposures. The total-loss check is $195(533.846\ldots)=\$104{,}100$.

Worked Example: Scaling a Rate for Different Policy Durations

A hypothetical insurer's fitted model has the following log-link coefficients for a coverage: $\beta_0=\log(500)$, $\beta_1=\log(1.2)$ for a territory indicator $x_1$, and $\beta_2=\log(0.8)$ for a selected deductible indicator $x_2$. A fleet has $x_1=1$, $x_2=1$, and $1.5$ vehicle-years of exposure. Calculate the modeled pure premium and expected aggregate loss. Then calculate the expected aggregate loss for the same risk characteristics at $0.75$ vehicle-years.

Method. Apply each multiplicative factor to obtain the per-vehicle-year rate. Use the same rate for both durations because the modeled characteristics are unchanged, and scale separately by each exposure.

The per-unit mean is $$ m=\exp\{\log(500)+\log(1.2)+\log(0.8)\} =500(1.2)(0.8)=\$480\text{ per vehicle-year}. $$ For $1.5$ vehicle-years, the expected aggregate loss is $$ 1.5(480)=\$720. $$ For $0.75$ vehicle-years, it is $$ 0.75(480)=\$360. $$

Check. The exposure ratio is $1.5/0.75=2$, and the expected-loss ratio is $720/360=2$. A common mistake is to treat the model output of $480 as the expected aggregate loss for either policy. It is a rate per vehicle-year; the two period totals differ because their exposures differ.

Interpretation, Limits, and Common Traps

Exposure is a scale for the expected aggregate loss, not an additional rating characteristic that changes the per-unit pure premium by itself. In a log-link aggregate mean model, the offset $\log(e_i)$ provides exactly this scale. If the goal is to report a rate, remove the exposure scale and report the modeled per-unit mean. If the goal is to predict a policy-period expected loss, retain the exposure scale.

The simple conversion has important boundaries. First, units must agree: a rate per vehicle-year cannot be multiplied by payroll exposure or by a number of months that has not been converted to years. Second, exposure should represent the quantity of risk actually covered. Third, the constant-rate assumption can fail if the insured risk changes during the period. For example, a location may be added partway through a policy, or a coverage limit may change; treating all months as identical exposure may then misstate the mean unless the exposure is appropriately segmented.

An observation with zero exposure cannot have an observed pure premium $S_i/e_i$, since division by zero is undefined. It is not a usable per-unit rate observation. In the simple mean relationship, zero exposure corresponds to zero expected aggregate loss, but that statement does not create a rate estimate for a zero-exposure policy. Do not replace a zero denominator with an arbitrary small exposure.

A second frequent error is double-counting exposure. If a prediction is already an expected aggregate loss with the offset included, multiplying by exposure again is incorrect. Conversely, if the output is a pure premium per exposure unit, failing to multiply by the policy's exposure understates or overstates the period total whenever exposure differs from one unit.

Key takeaway: Read every prediction together with its units. A per-unit pure premium $m_i$ is a rate; expected aggregate loss for exposure $e_i$ is $e_i m_i$. Under a log link, $\log(e_i)$ is an offset in the aggregate mean, not an estimated risk factor. Exposure scales the mean only under the model's stated stable-rate assumption.

Check Your Understanding

For each problem, distinguish the modeled per-unit pure premium from the period expected aggregate loss. Amounts are hypothetical.

A fitted model gives $540 per vehicle-year for a risk insured for 9 months. Convert the duration to vehicle-years and find expected aggregate loss. (Answer: 0.75 vehicle-years; $405)
A log-link model has $\beta_0=\log(400)$, with a single class factor of $1.1$. A policy has $0.5$ location-years. Find the per-location-year pure premium and expected aggregate loss. (Answer: $440 per location-year; $220)
Group X has 40 policy-years at $700 per policy-year. Group Y has 60 policy-years at $500 per policy-year. Find the combined expected aggregate loss and combined pure premium. (Answer: $58,000; $580 per policy-year)
A model's reported prediction is already the expected aggregate loss of $315 for a policy with $0.7$ vehicle-years. What is the implied pure premium per vehicle-year, and should the prediction be multiplied by $0.7$ again? (Answer: $450 per vehicle-year; no, $315 is already the aggregate prediction)
A risk has a fitted rate of $800 per location-year. It is insured for 4 months, then a second identical location is added for the remaining 8 months. Assuming the rate is stable and exposure is location-years, calculate total exposure and expected aggregate loss. (Answer: $1(4/12)+2(8/12)=20/12=1.6667$ location-years; $1,333.33)

What You'll Learn