Offsets For Partial Terms And Changing Exposure

Why Exposure Must Match the Observation

A frequency model predicts claim counts over a specified observation period. Its exposure must describe the amount of risk observed during that same period. If a policy is in force for only part of a year, or covers a changing number of insured units, treating it as a full year for one unit generally misstates the expected count.

The basic exposure offset from earlier material on frequency models applies when a single frequency rate is appropriate throughout the observation. The harder cases arise when the term is partial or the rate or number of units changes within it. The reliable approach is to calculate the expected count over each relevant segment, then add those segment expectations.

Definition: Exposure is the amount of risk observed, expressed in the units used to define the frequency rate. For a rate per insured-unit-year, one insured unit covered for one year contributes one unit-year; a unit covered for a fraction of a year contributes that fraction. An effective exposure is an exposure adjusted for time-varying rate factors so that multiplying it by a reference rate gives the correct expected count.

Partial Terms Under a Constant Rate

Suppose \(N_i\) is the claim count for observation \(i\), and the annual expected frequency for its risk characteristics is \(\lambda_i\) claims per exposure unit-year. Under a Poisson frequency model with a constant rate over the term, the mean count is the annual rate multiplied by the observed exposure \(E_i\):

N_i \sim \operatorname{Poisson}(\mu_i), \qquad \mu_i = E_i\lambda_i, \qquad \log(\mu_i)=\log(E_i)+\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta}.

Here the frequency rate is modeled as \(\lambda_i=\exp(\mathbf{x}_i^\mathsf{T}\boldsymbol{\beta})\), and \(E_i\) is the number of exposure unit-years. The term \(\log(E_i)\) is the offset: it adjusts the expected count for the amount of observation while its coefficient is fixed at one. If a policy covers one car for five months, then \(E_i=5/12\) car-years, not one car-year.

This calculation assumes the modeled annual rate is constant during the observed part of the year, conditional on the model variables. Prorating an annual rate by time is not generally appropriate if the rate changes by season, calendar period, or another factor during the term. In that case, calculate the contribution from each period separately.

Worked Example: A Five-Month Policy Term

A hypothetical auto policy covers one car for five months. For the car’s rating characteristics, the modeled annual frequency is 0.12 claims per car-year. A rating factor of 1.25 applies throughout the term. Calculate the expected claim count and the probability of at least one claim, assuming a Poisson count and a constant rate over the covered period.

Method. Multiply the annual frequency by the rating factor to obtain the applicable rate, then multiply that rate by the partial-term exposure.

The adjusted annual frequency is \(0.12(1.25)=0.15\) claims per car-year. The exposure is \(5/12=0.4167\) car-years before rounding. Therefore,

\mu=0.15\left(\frac{5}{12}\right)=0.0625.

For a Poisson count, the probability of at least one claim is one minus the probability of zero claims:

P(N\geq 1)=1-P(N=0)=1-e^{-0.0625}=0.0606.

The expected count is 0.0625 claims, and the probability of at least one claim is 0.0606. Probabilities are rounded to four decimal places. Check: the no-claim probability is \(e^{-0.0625}=0.9394\), so the two probabilities sum to 1. The common mistake is to use a full car-year and report a mean of 0.15, overstating the expected count for this five-month term.

Changing Numbers of Insured Units

If the number of insured units changes, exposure is the total unit-time observed. For example, if a fleet has \(n_j\) insured units for \(t_j\) years in segment \(j\), its exposure is \(\sum_j n_jt_j\) unit-years. This total can be used with one rate only when the rate per unit-year is the same in all segments, conditional on the model.

More generally, let \(\lambda_j\) be the annual frequency per insured unit during segment \(j\). If claim counts in disjoint segments are conditionally independent Poisson counts, the total count is also Poisson, with mean equal to the sum of the segment means:

\mu=\sum_j n_jt_j\lambda_j.

The count distribution and additivity here rely on the conditional Poisson-process assumptions: the specified segment rates apply during their segments, and disjoint increments have independent counts. If those assumptions are unsuitable, the expected count can still be calculated by summing segment expectations, but a Poisson distribution for the total does not automatically follow.

Worked Example: A Fleet That Adds Vehicles Midyear

A hypothetical business insures two vehicles for the first four months of a year and three vehicles for the remaining eight months. The modeled frequency is constant at 0.08 claims per vehicle-year. Find the exposure and expected annual claim count, assuming a Poisson process with independent counts over disjoint periods.

Method. Calculate vehicle-years in each period and multiply each amount by the common annual rate.

In the first period, exposure is \(2(4/12)=8/12=2/3\) vehicle-years. In the second, it is \(3(8/12)=24/12=2\) vehicle-years. Total exposure is \(2/3+2=8/3\) vehicle-years. Thus,

\mu=0.08\left(\frac{8}{3}\right)=0.2133\text{ claims, to four decimal places}.

Equivalently, the segment means are \(0.08(2/3)=0.0533\) and \(0.08(2)=0.1600\); their sum is 0.2133. Check: \(0.0533+0.1600=0.2133\), subject to rounding. The common mistake is to count the three vehicles as covered for all twelve months, which would give an incorrect exposure of three vehicle-years.

When Frequency Changes Within the Term

Changing exposure alone is not the whole problem. A policy may cover the same number of units throughout the term while the applicable annual frequency changes by season or by a changing risk characteristic. In that situation, multiplying the unadjusted total exposure by one of the segment rates is generally wrong. The expected count is the sum of rate times exposure for each segment.

In continuous time, let \(n(t)\) be the number of insured units at time \(t\), and let \(\lambda(t)\) be the conditional annual claim rate per unit. With time measured in years, the expected count over the observation period is the integral of the instantaneous expected rate:

\mu=\int n(t)\lambda(t)\,dt.

This expression also explains the offset construction. If the rate is constant, \(\lambda(t)=\lambda\), and the integral becomes \(\lambda\int n(t)\,dt=\lambda E\), where \(E\) is total unit-year exposure. If the rate varies, define effective exposure relative to a chosen reference rate \(\lambda_0>0\) as

E^*=\int n(t)\frac{\lambda(t)}{\lambda_0}\,dt, \qquad \mu=\lambda_0E^*.

Thus, the ordinary total exposure is sufficient when the rate is constant across the relevant segments. If it is not, a model may use segment-specific expected counts or an effective exposure that incorporates the known rate pattern. The time factors and units must be defined consistently; an annual rate is multiplied by time in years, not by a number of months without conversion.

Divide the observation into segments.
Make a new segment whenever the number of insured units or an applicable frequency factor changes.

Measure exposure within each segment.
Multiply the insured-unit count by the segment duration expressed in the rate’s time unit.

Apply the matching rate.
Multiply each segment’s exposure by its annual frequency per unit, including the relevant modeled factors.

Add expected counts.
Sum the segment means. Use one total-exposure offset only if a common rate adequately represents the entire observation.

Worked Example: Seasonal Rates and Changing Fleet Size

A hypothetical delivery firm has one insured van for the first three months, two vans for the next six months, and one van for the final three months. The respective annual frequencies are 0.12, 0.15, and 0.09 claims per van-year. Calculate the expected count for the year and the probability of no claims, assuming a piecewise-constant Poisson rate and independent counts in the three segments.

Method. For each segment, multiply vans by years and then by that segment’s annual rate.

The first segment contributes \(1(3/12)=0.25\) van-years, so its mean is \(0.25(0.12)=0.0300\). The second contributes \(2(6/12)=1\) van-year, so its mean is \(1(0.15)=0.1500\). The final segment contributes \(1(3/12)=0.25\) van-years, so its mean is \(0.25(0.09)=0.0225\). Therefore,

\mu=0.0300+0.1500+0.0225=0.2025.

The total count has a Poisson distribution with this mean under the stated assumptions. Hence \(P(N=0)=e^{-0.2025}=0.8167\), to four decimal places. Check: total exposure is \(0.25+1+0.25=1.5\) van-years, and the exposure-weighted annual rate is \(0.2025/1.5=0.135\); multiplying \(1.5(0.135)\) recovers 0.2025. Using total exposure with any one segment’s rate would not give the correct mean.

Deriving the Segment-Sum Rule

The sum-of-segment-means formula follows from independent Poisson increments. Suppose segment \(j\) has a Poisson count \(N_j\) with mean \(\mu_j\), and segment counts are independent. The probability-generating function of a Poisson count is \(G_j(s)=\exp(\mu_j(s-1))\). Independence means the generating function of the total \(N=\sum_jN_j\) is the product of the segment generating functions:

G_N(s)=\prod_jG_j(s) =\prod_j\exp(\mu_j(s-1)) =\exp\left(\left(\sum_j\mu_j\right)(s-1)\right).

This is the generating function of a Poisson random variable with mean \(\sum_j\mu_j\). Since each segment mean is \(n_jt_j\lambda_j\), the total mean is \(\sum_jn_jt_j\lambda_j\). The result justifies adding segment expectations and, under the stated independence and Poisson assumptions, identifies the distribution of the total count as well.

Common Mistakes and Practical Checks

Using the full policy year for a partial term. Convert the covered duration to a fraction of a year before multiplying by an annual rate.
Counting units without their time in force. A vehicle added halfway through the year contributes only the covered fraction of a vehicle-year.
Using one rate when the rate changes. Compute each segment’s rate-times-exposure contribution before adding the means.
Mixing time units. A rate per year requires duration in years. For example, six months is \(6/12\) years, not 6 years.
Assuming an expected count determines the whole distribution. Segment expectations can be added without establishing a Poisson total; the Poisson conclusion also needs the relevant count-model assumptions.

A useful final check is dimensional: insured units multiplied by years gives unit-years, and unit-years multiplied by claims per unit-year gives claims. Also check that segment durations and unit counts cover exactly the observation period, with no overlap or omitted time. This topic is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

For each problem, use exposure in years and retain unrounded values until the final answer.

A hypothetical policy covers one insured unit for seven months. Its constant annual frequency is 0.24 claims per unit-year. Find the expected count and the probability of no claims under a Poisson model. (Answer: mean \(0.1400\); probability of no claims \(0.8694\))
A fleet has one unit insured for two months and four units insured for the remaining ten months. The annual rate is constantly 0.06 claims per unit-year. Find total unit-year exposure and the expected count. (Answer: \(3.5\) unit-years; mean \(0.2100\))
A policy covers two units for three months at an annual frequency of 0.10 claims per unit-year, then one unit for nine months at 0.16. Find the expected count. (Answer: \(0.1700\) claims)
A single unit is covered for six months. Its reference annual frequency is 0.20, and a constant multiplicative rating factor of 1.40 applies. Find the mean count and the probability of at least one claim under a Poisson model. (Answer: mean \(0.1400\); probability \(0.1306\))
One unit is covered for four months with an annual frequency of 0.08, followed by eight months with an annual frequency of 0.20. Assuming a piecewise-constant Poisson rate, calculate expected count and probability of no claims. (Answer: mean \(0.1600\); probability of no claims \(0.8521\))

What You'll Learn