Posterior Means as Credibility-Weighted Claim Rates

Why the Posterior Mean Has a Credibility Form

A rate estimate based only on a risk’s observed claims can be unstable when its exposure is limited. A Bayesian estimate combines that observed rate with a prior rate representing information about the population of comparable risks. In the Gamma-Poisson model, this combination has a particularly useful form: the posterior mean is a weighted average of the observed claim rate and the prior mean.

The weights are determined by the exposure and the prior parameters. Deriving them algebraically makes the result easier to interpret and helps prevent errors such as averaging the prior and observed rates equally, or using the prior weight for the observed rate. The calculations below use hypothetical insurance portfolios and claim counts.

Definition: Suppose the unknown claim rate is \(\lambda\), measured as expected claims per unit of exposure. Conditional on \(\lambda\), a claim count \(N\) over fixed exposure \(e\) follows a Poisson distribution with mean \(e\lambda\): \(N\mid\lambda\sim\operatorname{Poisson}(e\lambda)\). A Gamma prior with shape \(\alpha>0\) and rate parameter \(\beta>0\) has density proportional to \(\lambda^{\alpha-1}e^{-\beta\lambda}\), and prior mean \(\alpha/\beta\). With this shape-rate convention, the posterior mean is \((\alpha+N)/(\beta+e)\).

Exposure must be measured in units consistent with the claim rate. For example, if \(\lambda\) is claims per policy-year, \(e\) is in policy-years. For multiple periods with the same underlying rate, the model assumes counts are conditionally independent given \(\lambda\), and each period’s count has a Poisson mean equal to its exposure times \(\lambda\). The total count and total exposure then determine the update.

These are frequency-model calculations: claim severity is not being modeled. There is no interest-rate or mortality assumption in these examples because neither discounting nor survival contingencies enter the Poisson count model.

Deriving the Weighted Average

Let the prior be Gamma with shape \(\alpha\) and rate \(\beta\), and observe \(N=n\) claims over exposure \(e>0\). The Poisson likelihood, as a function of \(\lambda\), is proportional to \(\lambda^n e^{-e\lambda}\). Multiplying it by the prior density gives the posterior density up to a constant:

\lambda^{\alpha-1}e^{-\beta\lambda}\lambda^n e^{-e\lambda} = \lambda^{\alpha+n-1}e^{-(\beta+e)\lambda}.

This is the kernel of a Gamma distribution with shape \(\alpha+n\) and rate \(\beta+e\). Its mean is therefore \((\alpha+n)/(\beta+e)\). To reveal the credibility weights, split the numerator into the prior-shape term and the observed-count term:

\begin{aligned} E[\lambda\mid N=n] &=\frac{\alpha+n}{\beta+e}\\ &=\frac{\beta}{\beta+e}\frac{\alpha}{\beta} +\frac{e}{\beta+e}\frac{n}{e}. \end{aligned}

The equality follows because the first product on the right is \(\alpha/(\beta+e)\), and the second is \(n/(\beta+e)\). Their sum is the posterior mean. Thus the prior mean is multiplied by \(\beta/(\beta+e)\), while the observed claim rate \(n/e\) is multiplied by \(e/(\beta+e)\).

Theorem: For \(e>0\), the Gamma-Poisson posterior mean is \(Z(n/e)+(1-Z)(\alpha/\beta)\), where \(Z=e/(\beta+e)\). The observed-rate weight is \(Z\), the prior-mean weight is \(1-Z=\beta/(\beta+e)\), and the weights are nonnegative and sum to one. This weighted-rate expression does not apply when \(e=0\), because \(n/e\) is undefined; with zero exposure and no observed claims, the posterior remains the prior.

This is a convex combination: the posterior mean lies between the prior mean and observed rate when both weights are positive. It does not generally equal their simple average. Nor is the weight on the observed rate \(1-Z\); matching each rate to its coefficient is essential. The result concerns the posterior mean, not every feature of the posterior distribution.

Specify the rate and exposure units.
For example, define the unknown rate as claims per policy-year and express all exposure in policy-years.

Combine counts and exposure.
For multiple periods, calculate \(n=\sum_j n_j\) and \(e=\sum_j e_j\). Do not average period claim rates without accounting for their exposures.

Calculate the prior and observed rates.
The prior mean is \(\alpha/\beta\); for positive exposure, the observed rate is \(n/e\).

Apply the matching weights and check.
Use \(e/(\beta+e)\) on the observed rate and \(\beta/(\beta+e)\) on the prior mean. Confirm the result equals \((\alpha+n)/(\beta+e)\).

Worked Applications

Worked Example: Updating a Claim Rate for One Portfolio

A hypothetical insurer models claims for a class of small equipment policies. Let \(\lambda\) be annual claims per policy-year. The prior is Gamma with shape \(\alpha=12\) and rate \(\beta=4\) policy-years. The portfolio produces \(n=15\) claims over \(e=6\) policy-years. Find the posterior mean, showing the weighted-average form. The count model is Poisson with mean \(e\lambda\); no interest or mortality assumption applies.

Method and prior rate. The Gamma-Poisson conjugate update gives prior mean \(\alpha/\beta=12/4=3\) claims per policy-year.

Observed rate and weights. The observed rate is \(n/e=15/6=2.5\). The observed-rate weight is \(Z=e/(\beta+e)=6/(4+6)=0.6\). The prior weight is \(1-Z=4/10=0.4\).

Weighted estimate. The posterior mean is \(0.6(2.5)+0.4(3)=1.5+1.2=2.7\) claims per policy-year.

Direct check. The posterior is Gamma with shape \(\alpha+n=12+15=27\) and rate \(\beta+e=4+6=10\). Its mean is \(27/10=2.7\), agreeing with the weighted calculation.

Common mistake avoided. Giving the two rates equal weights would produce \((2.5+3)/2=2.75\), which is not the posterior mean. The weights are \(0.6\) and \(0.4\), not \(0.5\) and \(0.5\).

Worked Example: Combining Claim Counts Across Periods

For a hypothetical class of seasonal commercial vehicles, let \(\lambda\) be claims per vehicle-year, with a Gamma prior having shape \(\alpha=10\) and rate \(\beta=5\) vehicle-years. Period 1 has 9 claims over 3 vehicle-years, and period 2 has 4 claims over 5 vehicle-years. Assume the same underlying rate in both periods and conditional independence given \(\lambda\). Find the posterior mean. The period counts are Poisson with means equal to exposure times \(\lambda\); no interest or mortality assumption applies.

Method and aggregate data. Add the counts and exposures: \(n=9+4=13\) claims and \(e=3+5=8\) vehicle-years. The prior mean is \(\alpha/\beta=10/5=2\) claims per vehicle-year.

Observed rate and weights. The aggregate observed rate is \(n/e=13/8=1.625\). The weight on that rate is \(Z=8/(5+8)=8/13\), and the weight on the prior mean is \(5/13\).

Weighted estimate. The posterior mean is \((8/13)(13/8)+(5/13)(2)=1+10/13=23/13=1.769230\ldots\) claims per vehicle-year. Rounded only at the final step, to four decimal places, this is \(1.7692\).

Direct check. The posterior shape is \(10+13=23\), and the posterior rate is \(5+8=13\). The posterior mean \(23/13=1.769230\ldots\) confirms the calculation.

Common mistake avoided. Averaging the two period rates equally gives \((9/3+4/5)/2=(3+0.8)/2=1.9\). That is incorrect because the periods have different exposures. The combined rate is total claims divided by total exposure, \(13/8\).

Worked Example: When the Observed Rate Matches the Prior

A hypothetical portfolio of inspection-service policies has a claim rate \(\lambda\) measured per 100 policy-years. Its prior is Gamma with shape \(\alpha=6\) and rate \(\beta=3\) in units consistent with that exposure. It records \(n=24\) claims over \(e=12\) exposure units of 100 policy-years. Find the posterior mean and verify it both ways. Conditional on \(\lambda\), the count is Poisson with mean \(e\lambda\); no interest or mortality assumption applies.

Method and rates. The prior mean is \(6/3=2\) claims per 100 policy-years. The observed rate is \(24/12=2\) in the same units.

Weights. The observed-rate weight is \(Z=12/(3+12)=12/15=0.8\), and the prior weight is \(3/15=0.2\).

Weighted estimate. The posterior mean is \(0.8(2)+0.2(2)=1.6+0.4=2\) claims per 100 policy-years.

Direct check. The posterior shape is \(6+24=30\), and the posterior rate is \(3+12=15\). Thus the posterior mean is \(30/15=2\), as calculated.

Common mistake avoided. A candidate may assume that a higher observed-rate weight must make the posterior rate higher than the prior mean. Here both rates are equal, so any convex weighting leaves the estimate at 2.

Interpretation and Exam Traps

The algebra gives a direct way to check both the calculation and its interpretation. The weight on observed experience is the exposure divided by prior rate parameter plus exposure. The coefficient \(\beta\) is part of the specified Gamma prior; it is not itself a claim count or the prior mean. The rate \(\alpha/\beta\) must be calculated before combining it with \(n/e\).

The weighted expression requires positive exposure for the observed claim rate. If \(e=0\), the model has no Poisson exposure and, in the ordinary count setup, \(n=0\). The posterior parameters remain \(\alpha\) and \(\beta\), so the posterior mean remains \(\alpha/\beta\). It would be invalid to calculate \(n/e\) or to use the weighted-rate expression by substituting \(e=0\) into a formula that includes that undefined rate.

Keep the parameterization consistent. A Gamma distribution can also be expressed using a scale parameter, but the \(\beta\) used here is a rate parameter: its prior mean is \(\alpha/\beta\), and the posterior rate is \(\beta+e\). Mixing the rate and scale conventions changes the mean and the update. A reliable check is to derive the posterior kernel and confirm the exponent has coefficient \(\beta+e\).

Key takeaway: For positive exposure, calculate the posterior mean either directly as \((\alpha+n)/(\beta+e)\) or as the weighted average of \(n/e\) and \(\alpha/\beta\). In the weighted version, \(e/(\beta+e)\) belongs to the observed rate, \(\beta/(\beta+e)\) belongs to the prior mean, and the two weights sum to one.

Check Your Understanding

For each hypothetical claim-count model, identify the posterior mean using the stated Gamma shape-rate convention. Answers retain the claim-rate units in the problem.

A risk has a Gamma prior with \(\alpha=8\) and \(\beta=4\) exposure units. It has 9 claims over 6 exposure units. Find the prior mean, observed-rate weight, and posterior mean. (Answer: prior mean \(2\), observed-rate weight \(0.6\), posterior mean \(1.7\) claims per exposure unit)
A class has a Gamma prior with \(\alpha=6\) and \(\beta=3\) vehicle-years. It records 5 claims over 2 vehicle-years and then 7 claims over 4 vehicle-years, under one constant rate. Find the aggregate observed rate and posterior mean. (Answer: aggregate observed rate \(2\), posterior mean \(2\) claims per vehicle-year)
A portfolio has \(\alpha=9\), \(\beta=3\) policy-years, and 10 claims over 5 policy-years. Calculate both weights and the posterior mean. (Answer: prior weight \(3/8\), observed-rate weight \(5/8\), posterior mean \(2.375\) claims per policy-year)
A risk has prior parameters \(\alpha=4\), \(\beta=2\) exposure units, with 15 claims over 10 exposure units. Calculate the posterior mean from the weighted form and verify using the posterior parameters. (Answer: posterior mean \(19/12\), or \(1.5833\) claims per exposure unit)
A risk has a Gamma prior with \(\alpha=7\) and \(\beta=2\) exposure units, but has zero exposure and no claims. State the posterior mean and explain whether an observed claim rate can be used. (Answer: posterior mean \(3.5\) claims per exposure unit; no observed rate is defined because exposure is zero)

What You'll Learn