Exposure Scaling in Frequency-Severity Models

Why Exposure Must Be Kept Separate From Severity

A policy in force for a fraction of a year generally contributes less expected claim count than an otherwise identical policy in force for a full year. That reduction is an exposure effect on frequency. It does not mean that each claim is expected to be smaller. Severity describes the amount of an individual claim, so its unit is dollars per claim; frequency describes claims per exposure unit.

Keeping these quantities distinct is essential when using frequency-severity models to price policies with different terms, insured units, or periods of coverage. First calculate the expected number of claims from the exposure and frequency rate. Then combine that count with the expected amount per claim. Scaling the severity by exposure as well would apply the exposure adjustment twice.

Definition: Let $e$ denote exposure measured in a stated unit, such as vehicle-years or policy-years. Let $\lambda$ be the expected number of claims per unit of exposure, and let $X$ be the amount of one claim, with mean severity $\mu=E[X]$ dollars per claim. Under a frequency model whose expected claim count is $e\lambda$, the expected aggregate loss is $e\lambda\mu$. Exposure scales the expected count; $\mu$ remains a per-claim quantity.

Model Assumptions and the Exposure Formula

A standard frequency-severity specification assumes that, conditional on the risk characteristics and exposure, the claim count $N$ has a mean proportional to exposure. A common choice is a Poisson frequency model: $N\mid e\sim\operatorname{Poisson}(e\lambda)$. The claim amounts $X_1,X_2,\ldots$ are assumed to have the specified severity distribution with finite mean $\mu$, and to be identically distributed and independent of $N$, conditional on the modeled risk characteristics. Aggregate loss is $S=\sum_{j=1}^{N}X_j$, with $S=0$ when $N=0$.

For the basic formula, the frequency rate and severity mean are held constant over the exposure being measured. The Poisson distribution is a standard way to model counts, but the expected-loss calculation itself only requires $E[N\mid e]=e\lambda$ and a common conditional severity mean $\mu$. If the expected count is not proportional to exposure, or expected severity changes over the covered period, the simple scaling formula needs adjustment.

E[N\mid e]=e\lambda,\qquad E[S\mid e]=e\lambda\mu

The severity distribution can be any appropriate positive-loss distribution with a finite mean. For example, if $X$ is Gamma with shape $\alpha$ and scale $\theta$, then $E[X]=\alpha\theta$. If $X$ is exponential with mean $m$, then $E[X]=m$. Those parameters describe the amount of one claim; they are not multiplied by the policy exposure.

Unit check: If $e$ is measured in vehicle-years, then $\lambda$ must be claims per vehicle-year. The product $e\lambda$ is an expected number of claims, and $e\lambda\mu$ is dollars of expected aggregate loss. A rate stated per policy-year cannot be multiplied directly by vehicle-years unless the exposure basis is reconciled.

Why the Formula Holds

The expected aggregate loss follows by conditioning on the number of claims. Given $N=n$, the aggregate is the sum of $n$ claim amounts. Under the stated common-mean severity assumption, its conditional expectation is $n\mu$. Averaging this result over the random count gives the expected count times the mean amount per claim.

E[S\mid N=n,e]=n\mu

E[S\mid e] =E\!\left[E[S\mid N,e]\mid e\right] =E[N\mu\mid e] =\mu E[N\mid e] =\mu(e\lambda) =e\lambda\mu

This argument explains why exposure does not appear inside the severity mean. The expected number of terms in the aggregate sum changes with exposure; the expected value of each term does not, under the assumption that severity is stable for the risk being modeled.

Set the exposure basis.
Record the amount of coverage in the same units used to state the frequency rate.

Calculate expected claims.
Multiply exposure by the frequency rate: $E[N\mid e]=e\lambda$.

Calculate mean severity.
Use the severity distribution’s parameters to obtain $\mu=E[X]$, in dollars per claim.

Calculate expected aggregate loss.
Multiply expected claims by mean severity: $E[S\mid e]=e\lambda\mu$.

Worked Applications

Worked Example: Scaling a Fractional-Year Policy

A hypothetical policy covers one vehicle for 9 months. Its expected claim frequency is 0.12 claims per vehicle-year. Conditional claim severity is Gamma with shape 2 and scale $\$2{,}500$. Calculate expected claim count and expected aggregate loss for the policy term. No interest or mortality assumption is involved; the calculation concerns property-and-casualty loss over the stated exposure period.

Method. Convert the term to vehicle-years, calculate the expected count, and then multiply by the mean amount per claim. The Gamma mean is shape times scale:

$e=9/12=0.75$ vehicle-year.

$\mu=2(\$2{,}500)=\$5{,}000$ per claim.

The expected claim count is $e\lambda=0.75(0.12)=0.09$ claims.

The expected aggregate loss is $0.09(\$5{,}000)=\$450.00$. Equivalently, $0.75(0.12)(\$5{,}000)=\$450.00$.

Check. The policy has three-quarters of a vehicle-year, so its expected count is three-quarters of the annual rate: $0.75(0.12)=0.09$. Severity remains $\$5{,}000$ per claim, not $0.75(\$5{,}000)$. The common mistake avoided here is scaling severity by the policy term as well as scaling frequency.

Worked Example: Combining Unequal-Exposure Groups

Three hypothetical vehicle groups have the following assumptions. Group A has 0.5 vehicle-year of exposure, frequency 0.12 claims per vehicle-year, and Gamma severity with shape 2 and scale $\$1{,}000$. Group B has 1.5 vehicle-years, frequency 0.08 claims per vehicle-year, and exponential severity with mean $\$3{,}000$. Group C has 0.5 vehicle-year, frequency 0.20 claims per vehicle-year, and Gamma severity with shape 4 and scale $\$500$. Find expected aggregate loss for all groups and expected loss per vehicle-year. No interest or mortality assumption is involved.

Method. Calculate each group’s expected count and mean severity separately, then add group expected losses. The frequency rates are per vehicle-year. For a Gamma distribution, mean severity is shape times scale; for the stated exponential distribution, the mean is given directly.

Group A. Mean severity is $2(\$1{,}000)=\$2{,}000$ per claim. Expected count is $0.5(0.12)=0.06$. Expected loss is $0.06(\$2{,}000)=\$120.00$.

Group B. Mean severity is $\$3{,}000$ per claim. Expected count is $1.5(0.08)=0.12$. Expected loss is $0.12(\$3{,}000)=\$360.00$.

Group C. Mean severity is $4(\$500)=\$2{,}000$ per claim. Expected count is $0.5(0.20)=0.10$. Expected loss is $0.10(\$2{,}000)=\$200.00$.

Total expected aggregate loss is $\$120.00+\$360.00+\$200.00=\$680.00$. Total exposure is $0.5+1.5+0.5=2.5$ vehicle-years. Expected loss per vehicle-year is $\$680.00/2.5=\$272.00$ per vehicle-year.

Check. The total expected claim count is $0.06+0.12+0.10=0.28$, and the total expected loss is the sum of the three independently calculated group expectations, $\$680.00$. The common mistake avoided is averaging group frequencies or severities without accounting for their different exposures and risk characteristics.

Worked Example: Converting Months to Vehicle-Years Before Scaling

A hypothetical fleet has two coverage blocks. Block 1 covers a vehicle for 10 months, with frequency 0.18 claims per vehicle-year and Gamma severity with shape 3 and scale $\$800$. Block 2 covers vehicles for a total of 18 vehicle-months, with frequency 0.10 claims per vehicle-year and Gamma severity with shape 2 and scale $\$4{,}000$. Calculate the expected aggregate loss for each block and for the fleet. No interest or mortality assumption is involved.

Method. Convert both exposure amounts to vehicle-years before multiplying by annual frequency rates. Use each block’s own severity mean. The Gamma parameters are shape and scale, so the mean is their product.

Block 1. Exposure is $10/12=5/6$ vehicle-year. Mean severity is $3(\$800)=\$2{,}400$ per claim. Expected count is $(5/6)(0.18)=0.15$. Expected loss is $0.15(\$2{,}400)=\$360.00$.

Block 2. Exposure is $18/12=1.5$ vehicle-years. Mean severity is $2(\$4{,}000)=\$8{,}000$ per claim. Expected count is $1.5(0.10)=0.15$. Expected loss is $0.15(\$8{,}000)=\$1{,}200.00$.

Fleet expected aggregate loss is $\$360.00+\$1{,}200.00=\$1{,}560.00$. Total exposure is $5/6+1.5=7/3$ vehicle-years. The implied expected loss per vehicle-year is $\$1{,}560.00/(7/3)=\$668.57$, rounded to the nearest cent.

Check. The total expected count is $0.15+0.15=0.30$. The fleet’s expected loss is not $0.30$ times either group’s severity mean, because the groups have different severities. The group-by-group calculation correctly gives $\$360.00+\$1{,}200.00=\$1{,}560.00$. The common mistake avoided is treating vehicle-months as vehicle-years when the frequency rate is annual.

Unequal Exposures and Portfolio Rates

For a portfolio divided into risk groups, let group $g$ have exposure $e_g$, frequency rate $\lambda_g$, and mean severity $\mu_g$. Under the same exposure-proportional frequency and stable-severity assumptions, the expected total loss is the sum of the group expectations. Dividing by total exposure gives the expected loss per exposure unit for the portfolio. This is an exposure-weighted portfolio result, not generally a simple average of the group rates.

E[S_{\mathrm{portfolio}}] =\sum_g e_g\lambda_g\mu_g,\qquad \text{expected loss per exposure unit} =\frac{\sum_g e_g\lambda_g\mu_g}{\sum_g e_g}

The denominator must use the same exposure basis as the group frequency rates. A group with more exposure contributes more to the total expected loss, but its per-claim severity is still applied only to its expected claim count. If expected frequency or severity varies within a group, divide the data into sufficiently homogeneous cells or use predictions that reflect those differences before aggregating.

Common Mistakes and Exam Checks

A frequent incorrect calculation is to multiply exposure into both frequency and severity. For example, with $e=0.5$, $\lambda=0.12$ claims per vehicle-year, and $\mu=\$2{,}000$ per claim, the incorrect calculation $0.5(0.12)\{0.5(\$2{,}000)\}=\$60$ scales severity twice. The correct expected loss is $0.5(0.12)(\$2{,}000)=\$120$. Expected count is $0.06$, and each claim still has mean severity $\$2{,}000$.

Check the exposure unit. Convert months to years, or otherwise reconcile the basis, before applying a rate stated per year.
Check the severity parameters. A Gamma scale is not generally its mean; for shape $\alpha$ and scale $\theta$, the mean is $\alpha\theta$.
Check for group differences. Multiply each group’s own exposure, frequency, and severity predictions before summing expected losses.
Check the interpretation. Expected claim count can be fractional; it is a mean, not a prediction that a fraction of a claim will occur.
Check the scope of linear scaling. The formula presumes expected count proportional to exposure and a stable per-claim severity mean. A changing risk profile or coverage condition may invalidate that assumption.

Key takeaway: Exposure scales the expected number of claims. The mean severity remains dollars per claim. For each homogeneous risk group, calculate $e\lambda\mu$, then add expected losses across groups; divide by total exposure only when a portfolio rate per exposure unit is required.

Check Your Understanding

For each problem, use the stated exposure unit and keep the severity mean on a per-claim basis.

A hypothetical policy has 0.4 vehicle-year of exposure, frequency 0.25 claims per vehicle-year, and Gamma severity with shape 2 and scale $\$1{,}200$. Find expected claim count and expected aggregate loss. (Answer: 0.10 claims; $\$240.00$)
A hypothetical account has 1.2 policy-years of exposure, frequency 0.07 claims per policy-year, and exponential severity with mean $\$4{,}000$. Find expected aggregate loss. (Answer: $\$336.00$)
Group A has 0.75 vehicle-year, frequency 0.12 claims per vehicle-year, and mean severity $\$3{,}000$. Group B has 2 vehicle-years, frequency 0.06 claims per vehicle-year, and mean severity $\$5{,}000$. Find total expected loss and expected loss per vehicle-year. (Answer: $\$870.00$ total; $\$316.36$ per vehicle-year)
A hypothetical vehicle is insured for 10 months. Its annual frequency rate is 0.18 claims per vehicle-year, and its severity is Gamma with shape 2 and scale $\$1{,}000$. Find expected count and expected loss for the term. (Answer: 0.15 claims; $\$300.00$)
Group A has 0.5 vehicle-year, frequency 0.20 claims per vehicle-year, and Gamma severity with shape 3 and scale $\$500$. Group B has 1.25 vehicle-years, frequency 0.08 claims per vehicle-year, and exponential severity with mean $\$3{,}500$. Find total expected loss and the portfolio expected loss per vehicle-year. (Answer: $\$500.00$ total; $\$285.71$ per vehicle-year)

What You'll Learn