Exposure Scaling in Gamma-Poisson Credibility

Why Exposure Must Scale the Update

A claim count from a short observation period does not carry the same frequency information as the same count from a long period. If a hypothetical policy produces 3 claims in one quarter, that experience is not equivalent to 3 claims in a full year. In a Gamma-Poisson model, the observation period enters through its exposure: the total claim count updates one prior parameter, while total exposure updates the other.

The model in this tutorial assumes one unknown claim rate shared across the observed periods. Each period may have a different known exposure, such as a fraction of a policy-year, but its expected count is proportional to that exposure. The result is useful for Bayesian updating when experience is recorded over unequal durations or across policies with unequal exposure.

Definition: Let $\lambda$ be a claim rate per unit of exposure. Under the shape-rate convention, a Gamma prior with parameters $\alpha>0$ and $\beta>0$ has density $$ \pi(\lambda)=\frac{\beta^\alpha}{\Gamma(\alpha)}\lambda^{\alpha-1}e^{-\beta\lambda},\qquad \lambda>0. $$ For period $j$, let $e_j\geq 0$ be its known exposure and $N_j$ its claim count. The model is $N_j\mid\lambda\sim\operatorname{Poisson}(e_j\lambda)$, with counts conditionally independent given $\lambda$. The exposure unit used to define $\lambda$ must also be used in $\beta$.

The Exposure-Scaled Posterior

For observed periods $j=1,\ldots,m$, write $n_j$ for the observed count. Define the total count $n=\sum_{j=1}^m n_j$ and total exposure $E=\sum_{j=1}^m e_j$. Under the model above, the posterior distribution is

Theorem: If the prior is $\operatorname{Gamma}(\alpha,\beta)$ in the shape-rate convention, and the conditionally independent observations satisfy $N_j\mid\lambda\sim\operatorname{Poisson}(e_j\lambda)$, then $$ \lambda\mid\boldsymbol{n}\sim\operatorname{Gamma}\left(\alpha+\sum_{j=1}^m n_j,\ \beta+\sum_{j=1}^m e_j\right). $$ Thus the data enter the posterior parameters through total claims and total exposure. The posterior mean is $$ E[\lambda\mid\boldsymbol{n}]=\frac{\alpha+\sum_{j=1}^m n_j}{\beta+\sum_{j=1}^m e_j}. $$

Why Total Count and Total Exposure Appear

The likelihood for one period, apart from its dependence on $\lambda$, is proportional to $\lambda^{n_j}e^{-e_j\lambda}$. Conditional independence means that the joint likelihood is the product of these period likelihoods. Multiplying the factors adds the count exponents and the exposure terms:

L(\lambda\mid\boldsymbol{n}) \propto \prod_{j=1}^m \lambda^{n_j}e^{-e_j\lambda} =\lambda^{\sum_j n_j}e^{-\lambda\sum_j e_j}.

Multiplying this likelihood by the prior density gives a kernel proportional to $\lambda^{\alpha-1+\sum_j n_j}e^{-(\beta+\sum_j e_j)\lambda}$. This is the kernel of a Gamma distribution with shape $\alpha+\sum_j n_j$ and rate $\beta+\sum_j e_j$, proving the update. The factorial terms in the Poisson probabilities do not involve $\lambda$, so they do not alter the posterior parameters.

This result also explains the role of an observation with zero claims. If its exposure is positive, it contributes no amount to the shape increment, but it does contribute to the rate increment. A period with both zero count and zero exposure contributes nothing. Treating every period as if it had equal exposure would discard information about how long the counts were observed.

Worked Example: Two Unequal Policy Periods

A hypothetical insurer models a policy’s annual claim rate $\lambda$ with prior $\operatorname{Gamma}(2,1)$, using the shape-rate convention. The policy has 1 claim during 0.5 policy-years of exposure and 6 claims during a separate 1.5 policy-years. Find the posterior distribution and posterior mean annual rate.

Method. Apply the Gamma-Poisson update using total count and total policy-year exposure. The exposure is measured in policy-years, so the prior rate parameter is also on that exposure scale.

Step 1: Total the observations. The total claim count is $n=1+6=7$. The total exposure is $E=0.5+1.5=2.0$ policy-years.

Step 2: Update the parameters. The posterior shape is $2+7=9$, and the posterior rate is $1+2.0=3.0$.

\lambda\mid\boldsymbol{n}\sim\operatorname{Gamma}(9,3),\qquad E[\lambda\mid\boldsymbol{n}]=\frac{9}{3}=3.

The posterior mean is 3 claims per policy-year. As a check, both posterior parameters are positive, and the rate parameter increased by exactly the 2.0 policy-years observed. A common mistake would be to average the two period rates, $1/0.5=2$ and $6/1.5=4$, giving 3 without using the prior; that is not the specified Bayesian update. In this particular data set the numerical value happens to match the posterior mean, but that coincidence does not make the method valid.

Worked Example: A Claim-Free Period Still Adds Exposure

A hypothetical portfolio segment has a prior annual claim rate $\lambda\sim\operatorname{Gamma}(4,2)$. It records 0 claims in 0.25 policy-years, 2 claims in 0.75 policy-years, and 8 claims in 2.00 policy-years. Determine the posterior distribution and posterior mean.

Method. Add all counts and all exposures, including exposure in a period with no claims.

Step 1: Calculate totals. The total count is $0+2+8=10$. The total exposure is $0.25+0.75+2.00=3.00$ policy-years.

Step 2: Update. The posterior shape is $4+10=14$, and the posterior rate is $2+3=5$.

\lambda\mid\boldsymbol{n}\sim\operatorname{Gamma}(14,5),\qquad E[\lambda\mid\boldsymbol{n}]=\frac{14}{5}=2.8.

The posterior mean is 2.8 claims per policy-year. The zero-claim period contributes 0 to the shape increment and 0.25 to the rate increment. If that period were incorrectly omitted, the rate parameter would instead be calculated as $2+2.75=4.75$, producing a posterior mean of $14/4.75\approx2.9474$. That omission overstates the posterior mean because it ignores observed claim-free exposure.

Worked Example: Include a Zero-Exposure Record Correctly

A hypothetical policy has prior annual claim rate $\lambda\sim\operatorname{Gamma}(2,1)$. The records show 0 claims in 0.5 policy-years, 6 claims in 1.5 policy-years, and a third entry with 0 claims and 0 exposure because the policy was not in force. Find the posterior mean annual rate.

Method. Use the likelihood for the periods actually at risk. A record with zero exposure has Poisson mean $0\lambda=0$, so its claim count must be zero under this model and it supplies no information about $\lambda$.

Step 1: Identify informative totals. The total count is $0+6+0=6$. The total exposure is $0.5+1.5+0=2.0$ policy-years.

Step 2: Update the prior. The posterior shape is $2+6=8$, and the posterior rate is $1+2=3$.

\lambda\mid\boldsymbol{n}\sim\operatorname{Gamma}(8,3),\qquad E[\lambda\mid\boldsymbol{n}]=\frac{8}{3}=2.6667.

The posterior mean is 2.67 claims per policy-year, rounded to two decimal places. As a check, the zero-exposure record changes neither posterior parameter. A common error is to treat the zero count in that record as claim-free exposure; it is not, because the policy was not at risk during any exposure time.

Assumptions, Units, and Limits of the Update

The update is exact under the stated model, but its interpretation depends on the assumptions. In particular, a single $\lambda$ must describe the claim rate across the included periods. Conditional on that rate, counts are Poisson and independent, and exposures are known. If the underlying rate changes by period, if observations are dependent beyond the shared rate, or if exposure is measured inconsistently, simply aggregating counts and exposure need not give an appropriate model.

Units deserve explicit attention. If $\lambda$ is claims per policy-year, each $e_j$ must be in policy-years and $\beta$ must be expressed in policy-years. For example, a six-month observation is $0.5$ policy-years, not 6 policy-years or 6 annual observations. Alternatively, a model can define a monthly rate, but then all exposures and the prior rate parameter must be in months. A numerical exposure cannot be carried unchanged when the time unit changes.

Also distinguish the prior rate parameter $\beta$ from the claim rate $\lambda$. Under the shape-rate convention, the prior mean is $\alpha/\beta$, and adding exposure to $\beta$ is what makes the posterior mean account for time at risk. Under a shape-scale convention, the formula would be written differently; do not substitute scale-form parameters into the shape-rate update.

Key takeaway: For a shared Poisson claim rate and a Gamma(shape-rate) prior, update the shape by adding total claims and update the rate by adding total exposure. Check that the count, exposure, and prior parameter all use consistent units before calculating the posterior.

Check Your Understanding

For each problem, use the Gamma shape-rate convention and state the posterior distribution or mean requested.

A hypothetical account has prior $\lambda\sim\operatorname{Gamma}(3,2)$ per policy-year. It has 4 claims over 2 policy-years. Find the posterior distribution and posterior mean. (Answer: $\operatorname{Gamma}(7,4)$; posterior mean $1.75$ claims per policy-year.)
A hypothetical risk has prior $\lambda\sim\operatorname{Gamma}(2,3)$ per policy-year. It has 1 claim over 0.5 policy-years and 2 claims over 1.5 policy-years. Find the posterior distribution and mean. (Answer: $\operatorname{Gamma}(5,5)$; posterior mean $1.00$ claim per policy-year.)
A hypothetical group has prior $\lambda\sim\operatorname{Gamma}(5,2)$ per policy-year. It records 0 claims over 1 policy-year and 9 claims over 3 policy-years. Find the posterior distribution and mean. (Answer: $\operatorname{Gamma}(14,6)$; posterior mean $2.3333$ claims per policy-year.)
A hypothetical policy has prior $\lambda\sim\operatorname{Gamma}(2,1)$ per policy-year. It records 3 claims in 6 months and 2 claims in 18 months. Convert the periods to policy-years, then find the posterior distribution and mean. (Answer: total exposure $2$ policy-years; $\operatorname{Gamma}(7,3)$; posterior mean $2.3333$ claims per policy-year.)
A hypothetical book has prior $\lambda\sim\operatorname{Gamma}(4,1.5)$ per policy-year. The data are 2 claims over 0.5 policy-years, 0 claims over 0.5 policy-years, and 5 claims over 2 policy-years. Find the posterior distribution and mean, explaining whether the claim-free period contributes. (Answer: $\operatorname{Gamma}(11,4.5)$; posterior mean $2.4444$ claims per policy-year; the claim-free period adds 0.5 to the rate parameter through exposure.)

This treatment is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

What You'll Learn