Gamma Priors for Poisson Claim Rates

Why Put a Prior on a Poisson Claim Rate?

A Poisson model describes claim counts over a stated amount of exposure when claims occur at a stable underlying rate. In practice, an insurer may have limited experience for a particular risk or rating cell. A Bayesian model represents prior information about that rate explicitly, then updates it using the observed claim count and exposure. The result is a posterior distribution: it describes the remaining uncertainty about the rate after the experience is considered.

This tutorial focuses on updating the rate itself and obtaining rate estimates from its posterior distribution. It does not require the prior to be interpreted as literal past claims. It is a probability model for the unknown rate, and its parameters determine both the prior center and the strength of the prior information.

Definition: Use the Gamma shape-rate parameterization \( \lambda\sim\operatorname{Gamma}(\alpha,\beta) \), where \( \alpha>0 \) is the shape and \( \beta>0 \) is the rate parameter. Its density is \( f_\lambda(\lambda)=\frac{\beta^\alpha}{\Gamma(\alpha)}\lambda^{\alpha-1}e^{-\beta\lambda} \), for \( \lambda>0 \). Thus \( E[\lambda]=\alpha/\beta \) and \( \operatorname{Var}(\lambda)=\alpha/\beta^2 \). Here, \(\lambda\) is a claim rate per exposure unit, and \(\beta\) is measured in exposure units.

The word “rate” has two roles in this notation. The unknown Poisson claim rate is \(\lambda\), measured in claims per exposure unit. The Gamma parameter \(\beta\) is the rate parameter of the prior distribution, not a claim rate; with this shape-rate convention, it has exposure units. Keeping the parameterization explicit prevents a common scale-versus-rate error.

Model and Posterior Update

Suppose observations are collected over \(m\) periods. In period \(j\), the known exposure is \(t_j\) and the observed claim count is \(n_j\). Assume that, conditional on one common unknown rate \(\lambda\), the period counts are independent and \( N_j\mid\lambda\sim\operatorname{Poisson}(t_j\lambda) \). The exposure units must be consistent across all periods. For example, if \(\lambda\) is claims per policy-year, each \(t_j\) must be in policy-years.

Write the total observed count as \(N=\sum_{j=1}^{m}n_j\), and the total exposure as \(T=\sum_{j=1}^{m}t_j\). Provided the prior is \( \operatorname{Gamma}(\alpha,\beta) \), the posterior is:

Theorem: Under the conditional Poisson model with known exposures and a common rate \(\lambda\), a Gamma prior updates to \( \lambda\mid\text{data}\sim\operatorname{Gamma}(\alpha+N,\beta+T) \). Consequently, \( E[\lambda\mid\text{data}]=(\alpha+N)/(\beta+T) \) and \( \operatorname{Var}(\lambda\mid\text{data})=(\alpha+N)/(\beta+T)^2 \). The update uses both total count and total exposure.

These results require the underlying rate to be common over the included periods, conditional independence of counts given that rate, and known exposure amounts. If the exposure is zero, it contributes neither count information nor exposure information. If the underlying rate changes across periods, this single-rate update does not describe that change; a different model is needed.

Why the Update Has This Form

The key is that the Poisson likelihood contributes one power of \(\lambda\) for each observed claim and an exponential term determined by exposure. For the independent period counts, the likelihood is:

\begin{aligned} L(\lambda) &=\prod_{j=1}^{m}\frac{e^{-t_j\lambda}(t_j\lambda)^{n_j}}{n_j!}\\ &\propto e^{-\lambda\sum_jt_j}\lambda^{\sum_jn_j}\\ &=e^{-T\lambda}\lambda^N. \end{aligned}

The omitted factors, such as \(t_j^{n_j}/n_j!\), do not depend on \(\lambda\), so they are irrelevant to the posterior density as a function of \(\lambda\). Multiplying the likelihood by the prior density gives a posterior kernel:

\begin{aligned} f_\lambda(\lambda)L(\lambda) &\propto \lambda^{\alpha-1}e^{-\beta\lambda} \lambda^Ne^{-T\lambda}\\ &=\lambda^{\alpha+N-1}e^{-(\beta+T)\lambda}. \end{aligned}

This is the kernel of a Gamma density with shape \(\alpha+N\) and rate parameter \(\beta+T\). Normalizing that kernel gives the posterior distribution stated above. The derivation explains why counts add to the shape and exposure adds to the Gamma rate parameter.

Posterior Rate Estimates

A standard actuarial estimate of the unknown claim rate is its posterior mean. Under squared-error loss, this is the Bayes estimate because the posterior mean minimizes expected squared estimation error conditional on the data. The estimate is a rate per exposure unit. If the rating basis is one policy-year, for instance, it is claims per policy-year; it should not be reported as a total count without specifying an exposure basis.

A different summary is the posterior mode. If the posterior shape \(a=\alpha+N\) is greater than 1, its mode is \((a-1)/(\beta+T)\). If \(a\leq1\), the density has its mode at the boundary \(\lambda=0\). The mode and mean are generally different, so the requested estimator or loss criterion matters. The maximum likelihood estimate from the Poisson data alone is \(N/T\) when \(T>0\); it does not include the prior information.

Worked Example: Updating Across Two Observation Periods

A hypothetical portfolio has a claim rate \(\lambda\) per policy-year with prior distribution \( \operatorname{Gamma}(2,3) \), using the shape-rate convention. In one period, 2 claims occur over 1.5 policy-years; in another, 3 claims occur over 2.5 policy-years. Find the posterior distribution, its mean and variance, and a squared-error actuarial estimate of the claim rate.

Method and sufficient totals. Add the observed counts and exposures: \(N=2+3=5\) claims and \(T=1.5+2.5=4\) policy-years. The two period counts are conditionally independent Poisson counts with the same rate, so the Gamma-Poisson update applies.

Posterior parameters. Add total claims to the prior shape and total exposure to the prior rate parameter: \( \alpha_{\text{post}}=2+5=7 \) and \( \beta_{\text{post}}=3+4=7 \). Therefore \( \lambda\mid\text{data}\sim\operatorname{Gamma}(7,7) \).

Posterior mean and variance. The rate estimate under squared-error loss is \( E[\lambda\mid\text{data}]=7/7=1 \) claim per policy-year. The posterior variance is \( \operatorname{Var}(\lambda\mid\text{data})=7/7^2=7/49=1/7 \), or approximately \(0.1429\) squared claims per policy-year.

Check. The observed experience rate is \(5/4=1.25\) claims per policy-year, while the prior mean is \(2/3\). The posterior mean of 1 lies between these values, as expected for this example. The common mistake to avoid is adding the two exposures to the count instead of to the prior rate parameter; the posterior parameters must be shape 7 and rate 7.

Worked Example: Unequal Exposures and the Rate Estimate

A hypothetical specialty insurer assigns a prior \( \lambda\sim\operatorname{Gamma}(3,4) \) to claims per equipment-year. It observes 1 claim over 2 equipment-years, 4 claims over 5 equipment-years, and 0 claims over 1 equipment-year. Calculate the posterior distribution, posterior mean and variance, and the data-only maximum likelihood estimate.

Method and totals. The observations share one conditional Poisson rate, so aggregate the counts and exposures: \(N=1+4+0=5\), and \(T=2+5+1=8\) equipment-years. In particular, the zero-claim period still contributes its full exposure.

Posterior distribution. The shape becomes \(3+5=8\), and the Gamma rate parameter becomes \(4+8=12\). Thus \( \lambda\mid\text{data}\sim\operatorname{Gamma}(8,12) \).

Posterior summaries. The posterior mean rate is \(8/12=2/3\), or about \(0.6667\) claims per equipment-year. The posterior variance is \(8/12^2=8/144=1/18\), or about \(0.0556\) squared claims per equipment-year.

Data-only comparison and check. Since \(T=8>0\), the maximum likelihood estimate is \(N/T=5/8=0.625\) claims per equipment-year. It differs from the posterior mean because it excludes the prior. The posterior estimate, \(0.6667\), is between the prior mean \(3/4=0.75\) and the observed rate \(0.625\). Avoid treating the three periods as equally weighted rates: the update is based on total claims and total exposure, not an unweighted average of period rates.

Worked Example: Zero Claims Over Positive Exposure

A hypothetical cover has prior claim rate \( \lambda\sim\operatorname{Gamma}(5,2) \), measured per device-year. The portfolio records zero claims during 6 device-years. Find the posterior distribution, mean, variance, and mode, and explain the effect of the zero-count experience.

Method and update. Here \(N=0\) and \(T=6\). The zero count leaves the shape unchanged, while the positive exposure increases the Gamma rate parameter: \( \alpha_{\text{post}}=5+0=5 \) and \( \beta_{\text{post}}=2+6=8 \). Hence \( \lambda\mid\text{data}\sim\operatorname{Gamma}(5,8) \).

Posterior mean and variance. The posterior mean is \(5/8=0.625\) claims per device-year. Its variance is \(5/8^2=5/64=0.078125\) squared claims per device-year.

Posterior mode. The shape is greater than 1, so the mode is \((5-1)/8=0.5\) claims per device-year. The mean and mode differ; for squared-error estimation, use the posterior mean, not the mode.

Check and interpretation. The prior mean was \(5/2=2.5\) claims per device-year. Zero claims across positive exposure lower the posterior mean to \(0.625\). The count being zero does not make the exposure irrelevant: without the 6 device-years, the posterior would remain the prior. A common mistake is to set both posterior parameters equal to their prior values whenever \(N=0\); the correct rate parameter is \(2+6=8\).

Common Exam Traps and Practical Checks

These updates are short once the model and parameterization are clear, but small interpretation errors can change the answer substantially. The following checks help distinguish valid calculations from plausible-looking ones.

Use the Gamma rate parameterization stated in the problem. Here the second parameter is \(\beta\), so the mean is \(\alpha/\beta\), and exposure adds to \(\beta\). A scale parameterization would use different formulas.
Keep count and exposure in their proper roles. Total count updates the shape; total exposure updates the Gamma rate parameter. Counts cannot replace exposure, even when every observation has a positive count.
Use consistent exposure units. If the rate is per 100 policy-years, convert exposure to units of 100 policy-years before updating. Mixing policy-years and blocks of 100 policy-years changes the implied rate.
Do not discard zero-count periods. A zero count contributes no increment to the shape, but its positive exposure still increments the rate parameter.
Check that all observations can share one rate. Combining periods assumes the same underlying \(\lambda\) governs them, conditional on that rate. A systematic change in the risk rate is not fixed merely by summing counts and exposures.
Label the estimate and its units. The posterior mean estimates a rate per exposure unit. It is not itself a total claim count, and the posterior variance is in squared rate units.

Key takeaway: With a Gamma shape-rate prior and conditionally independent Poisson counts sharing one rate, update by adding total claims to the shape and total exposure to the rate parameter. The posterior mean \((\alpha+N)/(\beta+T)\) is the squared-error actuarial rate estimate.

Check Your Understanding

For each problem, use the Gamma shape-rate convention and state posterior parameters before calculating summaries.

A hypothetical fleet has prior \( \operatorname{Gamma}(2,5) \). It records 3 claims over 2 vehicle-years. Find the posterior mean and variance. (Answer: posterior \( \operatorname{Gamma}(5,7) \); mean \(5/7=0.7143\), variance \(5/49=0.1020\).)
A device portfolio has prior \( \operatorname{Gamma}(4,3) \). It records 2 claims over 1 device-year and 5 claims over 3 device-years. Find the posterior distribution and mean rate. (Answer: posterior \( \operatorname{Gamma}(11,7) \); mean \(11/7=1.5714\) claims per device-year.)
A hypothetical policy class has prior \( \operatorname{Gamma}(1.5,2) \) and no claims over 4 exposure units. Find the posterior mean and variance. (Answer: posterior \( \operatorname{Gamma}(1.5,6) \); mean \(0.25\), variance \(1.5/36=0.0417\).)
A book of risks has prior \( \operatorname{Gamma}(6,4) \) and observes 8 claims over 5 exposure units. Find the posterior mean, variance, and mode. (Answer: posterior \( \operatorname{Gamma}(14,9) \); mean \(14/9=1.5556\), variance \(14/81=0.1728\), mode \(13/9=1.4444\).)
A hypothetical coverage group has prior \( \operatorname{Gamma}(3,6) \), with 9 claims over 6 exposure units. Find the posterior mean and compare it with the data-only maximum likelihood estimate. (Answer: posterior \( \operatorname{Gamma}(12,12) \); posterior mean \(1\), and maximum likelihood estimate \(9/6=1.5\) claims per exposure unit.)

What You'll Learn