Bayesian Updating Across Multiple Periods

Why Multiple Periods Can Be Pooled

A policyholder’s claims experience is often observed over several periods, and those periods may have different exposure amounts. A short period with little exposure should not carry the same weight as a long period with substantial exposure. In the Gamma-Poisson model, the appropriate update uses total claims and total exposure, rather than an unweighted average of the period-by-period frequencies.

The key assumption is that the periods share one underlying claim rate. Conditional on that rate, counts from distinct periods are independent Poisson observations, with each period’s expected count equal to the rate multiplied by its exposure. The periods may have different exposures; the rate itself is assumed not to change across them.

Definition: Suppose the unknown claim rate is \(\lambda\) claims per unit of exposure, and the prior is \(\lambda\sim\operatorname{Gamma}(\alpha,\beta)\) in the shape-rate convention, with density proportional to \(\lambda^{\alpha-1}e^{-\beta\lambda}\). Thus \(E[\lambda]=\alpha/\beta\). For periods \(j=1,\ldots,m\), let the exposure \(e_j\) be known and let \(N_j\mid\lambda\sim\operatorname{Poisson}(e_j\lambda)\), independently across periods conditional on \(\lambda\). Then the posterior is \(\operatorname{Gamma}\!\left(\alpha+\sum_{j=1}^m N_j,\ \beta+\sum_{j=1}^m e_j\right)\).

Exposure must be measured in the same units used to define \(\lambda\). For example, if the rate is per policy-year, then all period exposures must be in policy-years. The model also treats exposures as fixed and known; uncertainty in exposure is not included in this likelihood.

Why Total Claims and Total Exposure Are Sufficient

For a single period, the Poisson probability of observing \(n_j\) claims at exposure \(e_j\), conditional on \(\lambda\), is \(e^{-e_j\lambda}(e_j\lambda)^{n_j}/n_j!\). Conditional independence means that the joint likelihood is the product of these period-specific probabilities. Grouping terms that involve \(\lambda\) gives:

\begin{aligned} L(\lambda) &=\prod_{j=1}^{m}\frac{e^{-e_j\lambda}(e_j\lambda)^{n_j}}{n_j!}\\ &=\left(\prod_{j=1}^{m}\frac{e_j^{n_j}}{n_j!}\right) \lambda^{\sum_j n_j}e^{-\lambda\sum_j e_j}. \end{aligned}

The factor in parentheses does not depend on \(\lambda\), so it does not affect the posterior distribution of \(\lambda\). The likelihood’s dependence on \(\lambda\) is entirely through the total count \(\sum_j n_j\) and total exposure \(\sum_j e_j\). Multiplying by the Gamma prior kernel, \(\lambda^{\alpha-1}e^{-\beta\lambda}\), yields a posterior kernel proportional to:

\lambda^{\alpha+\sum_j n_j-1} e^{-(\beta+\sum_j e_j)\lambda}.

This is the kernel of a Gamma distribution with shape \(\alpha+\sum_j n_j\) and rate \(\beta+\sum_j e_j\), which establishes the stated update. In this model, the individual period labels do not add further information about \(\lambda\) once total claims and total exposure are known. That conclusion relies on the shared-rate and conditional-independence assumptions; it does not mean that period patterns are always irrelevant to actuarial analysis.

Theorem: Under the stated Gamma-Poisson assumptions, the posterior mean claim rate is \(E[\lambda\mid N_1,\ldots,N_m]=(\alpha+\sum_j N_j)/(\beta+\sum_j e_j)\). Equivalently, it is a credibility-weighted average of the prior mean rate and the pooled observed frequency:

E[\lambda\mid\mathbf{N}] = \frac{\beta}{\beta+E}\frac{\alpha}{\beta} + \frac{E}{\beta+E}\frac{N}{E}, \qquad N=\sum_{j=1}^m N_j,\quad E=\sum_{j=1}^m e_j,\quad E>0.

To verify the identity, the right side is \(\alpha/(\beta+E)+N/(\beta+E)=(\alpha+N)/(\beta+E)\). The credibility weight on the pooled observed frequency is \(E/(\beta+E)\), while the prior mean receives weight \(\beta/(\beta+E)\). The prior rate parameter \(\beta\) therefore acts like prior exposure in this mean formula. If total exposure is zero, there is no observed frequency to use; the posterior remains the prior.

Worked Updates Across Periods

In the examples, all counts are hypothetical. The Gamma distribution is parameterized by shape and rate, so the second parameter is exposure-like and the mean is shape divided by rate. Period exposures and the prior rate are expressed in policy-years, unless otherwise stated.

Worked Example: Two Periods With Equal Prior and Experience Weights

A hypothetical policyholder has prior \(\lambda\sim\operatorname{Gamma}(3,120)\) claims per policy-year. In the first period, the policyholder has 2 claims in 90 policy-years; in the second, there is 1 claim in 30 policy-years. Find the posterior distribution and mean, then verify the credibility-weighted form.

Method and assumptions. Treat the two period counts as conditionally independent Poisson variables with the same underlying \(\lambda\). Add the counts and exposures: \(N=2+1=3\) claims and \(E=90+30=120\) policy-years.

Update the Gamma shape by the total claims and the Gamma rate by total exposure: \(\alpha_{\text{post}}=3+3=6\) and \(\beta_{\text{post}}=120+120=240\). Thus the posterior is \(\operatorname{Gamma}(6,240)\), and its mean is \(6/240=0.0250\) claims per policy-year.

The prior mean is \(3/120=0.0250\), and the pooled observed frequency is \(3/120=0.0250\). The experience weight is \(120/(120+120)=1/2\), and the prior weight is also \(1/2\). Therefore the credibility-weighted mean is \((1/2)(0.0250)+(1/2)(3/120)=0.0125+0.0125=0.0250\).

Check. The weighted mean agrees with the posterior mean. A common mistake is to use \(3/30\) as the pooled frequency, which uses only the second period’s exposure. The correct denominator for the combined count is \(90+30=120\) policy-years.

Worked Example: Unequal Period Exposures and Unequal Weights

A hypothetical fleet has a prior claim rate \(\lambda\sim\operatorname{Gamma}(2,80)\) per vehicle-year. It reports 3 claims in a period with 40 vehicle-years of exposure, followed by 1 claim in a period with 60 vehicle-years. Find the posterior mean and show the corresponding credibility weights.

Method. Under a common rate and conditional independence, the total count is \(N=3+1=4\), and the total exposure is \(E=40+60=100\) vehicle-years. The posterior shape is \(2+4=6\), and the posterior rate is \(80+100=180\). Hence the posterior distribution is \(\operatorname{Gamma}(6,180)\).

The posterior mean is \(6/180=0.033333\ldots\) claims per vehicle-year. The prior mean is \(2/80=0.025\), and the pooled observed frequency is \(4/100=0.04\). The experience weight is \(100/180=5/9\), so the prior weight is \(80/180=4/9\). The weighted calculation is: \((4/9)(0.025)+(5/9)(0.04)=0.011111\ldots+0.022222\ldots=0.033333\ldots\).

Check. The weighted expression matches \(6/180\). The 60 vehicle-year period contributes more exposure than the 40 vehicle-year period, but the update does not average their two frequencies equally. Pooling counts and exposure gives the appropriate observed frequency.

Worked Example: A Zero-Claim Period in a Multi-Period Update

A hypothetical commercial account has prior \(\lambda\sim\operatorname{Gamma}(4,100)\) claims per policy-year. Its experience is 0 claims in 25 policy-years, 4 claims in 75 policy-years, and 1 claim in 50 policy-years. Calculate the posterior distribution and mean. Explain the role of the zero-claim period.

Method. The total number of claims is \(N=0+4+1=5\), and the total exposure is \(E=25+75+50=150\) policy-years. The posterior shape is \(4+5=9\), and the posterior rate is \(100+150=250\). Therefore the posterior is \(\operatorname{Gamma}(9,250)\), with mean \(9/250=0.036\) claims per policy-year.

The prior mean is \(4/100=0.04\), and the pooled observed frequency is \(5/150=1/30=0.033333\ldots\). The experience weight is \(150/250=0.6\), and the prior weight is \(100/250=0.4\). The credibility calculation is \((0.4)(0.04)+(0.6)(5/150)=0.016+0.020=0.036\).

Check. The result equals the posterior mean \(9/250\). The zero claims in the first period add no units to the posterior shape, but its 25 policy-years are included in total exposure and therefore increase the posterior rate. Omitting that exposure would incorrectly overstate the pooled frequency.

Conditions for Combining Periods

The total-statistics update is concise, but its assumptions should be checked before applying it. In particular, the model regards the period-specific count differences as random variation around one persistent underlying claim rate. It does not automatically accommodate a systematic shift in the risk or the measurement process.

Check the rate definition.
Confirm that \(\lambda\) is a rate per clearly defined exposure unit and that every period’s exposure uses that same unit.

Check the common-rate assumption.
Assess whether the periods plausibly represent the same underlying risk level. A known coverage, territory, or operational change may call for a different model or separate rates.

Pool both components.
Compute total claims \(N=\sum_jN_j\) and total exposure \(E=\sum_je_j\); do not average period frequencies without exposure weights.

Update and interpret.
Use posterior shape \(\alpha+N\) and rate \(\beta+E\). Interpret the posterior mean in the original rate units.

If the true rate changes by period, a single common \(\lambda\) is not a suitable representation of those changing rates. Pooling may conceal a trend or a structural break. Likewise, conditional independence rules out additional dependence between period counts after conditioning on \(\lambda\). The Gamma-Poisson update is not a general correction for temporal dependence or unmodeled changes.

A frequent exam trap is to add the period frequencies instead of pooling their numerators and denominators. For instance, if one period has 2 claims in 20 exposure units and another has 3 claims in 80, the pooled frequency is \(5/100=0.05\), not the unweighted average of \(2/20=0.10\) and \(3/80=0.0375\). The unweighted average gives the short period the same influence as the period with four times its exposure.

Key takeaway: With one persistent Poisson rate and conditionally independent period counts, all periods update the prior through two totals: the total number of claims and the total exposure. Keep the exposure units consistent, and do not pool periods if the common-rate assumption is not appropriate.

Check Your Understanding

For each problem, use the shape-rate Gamma convention and assume a common rate across periods unless the problem asks you to assess that assumption.

A prior is \(\operatorname{Gamma}(2,50)\) claims per policy-year. Experience is 1 claim in 20 policy-years and 2 claims in 30 policy-years. Find the posterior distribution and posterior mean. (Answer: \(\operatorname{Gamma}(5,100)\); mean \(0.0500\) claims per policy-year.)
A prior is \(\operatorname{Gamma}(3,90)\). Three periods have respective counts 0, 2, and 1, with exposures 20, 40, and 40 policy-years. Find the pooled observed frequency and posterior mean. (Answer: pooled frequency \(3/100=0.0300\); posterior \(\operatorname{Gamma}(6,190)\), mean \(0.03158\) claims per policy-year.)
A prior is \(\operatorname{Gamma}(5,200)\). Experience consists of 4 claims in 80 policy-years and 0 claims in 20 policy-years. Find the posterior distribution and the credibility weight on pooled experience. (Answer: \(\operatorname{Gamma}(9,300)\); experience weight \(100/300=1/3\).)
A prior is \(\operatorname{Gamma}(2,40)\) per vehicle-year. An account has 1 claim in 20 vehicle-years and 3 claims in 60 vehicle-years. Calculate the posterior mean using both the posterior parameters and the credibility-weighted form. (Answer: posterior \(\operatorname{Gamma}(6,120)\); mean \((6/120)=(1/3)(2/40)+(2/3)(4/80)=0.0500\) claims per vehicle-year.)
A hypothetical insured has a material midyear change in vehicle use, and the analyst believes the underlying claim rate increased after the change. Is the one-rate pooled update automatically justified? State the model issue. (Answer: No. A single common \(\lambda\) assumes a persistent rate across periods; a systematic rate change requires a model that represents the change rather than blindly pooling.)

What You'll Learn