Exposure and Prior Strength in Poisson Credibility

Why Exposure and Prior Strength Matter

A claim rate estimated from a small amount of exposure is usually less informative than the same rate estimated from a large amount. Bayesian Poisson credibility makes this distinction explicit: it combines a risk’s observed claim rate with a prior mean rate, assigning more weight to the observed rate as exposure grows. The Gamma prior also represents prior strength, but its shape parameter must be interpreted together with its rate parameter.

The model below applies when a risk has a constant underlying claim rate over the observation period and claims follow a Poisson count model conditional on that rate. Exposure and the claim rate must use compatible units. For example, if the rate is claims per vehicle-year, exposure must be measured in vehicle-years.

Definition: Let \(\lambda\) denote a risk’s unknown claim rate per exposure unit. Assume the observed claim count \(N\), conditional on \(\lambda\), has a Poisson distribution with mean \(e\lambda\), where \(e\) is the observed exposure. Write \(N\mid\lambda\sim\operatorname{Poisson}(e\lambda)\). Let the prior for \(\lambda\) be Gamma with shape \(\alpha>0\) and rate \(\beta>0\), denoted \(\lambda\sim\operatorname{Gamma}(\alpha,\beta)\). Under the shape-rate convention, the prior mean is \(\alpha/\beta\) and the prior variance is \(\alpha/\beta^2\).

The Gamma rate parameter \(\beta\) has exposure units, so it can be compared with observed exposure \(e\). The shape \(\alpha\) is dimensionless. In particular, \(\alpha\) is not itself the credibility weight, nor does it alone determine that weight. The model’s posterior mean can be expressed as a weighted average:

E[\lambda\mid N] = Z\frac{N}{e}+(1-Z)\frac{\alpha}{\beta}, \qquad Z=\frac{e}{e+\beta}.

Here \(N/e\) is the observed claim rate and \(\alpha/\beta\) is the prior mean rate. The credibility weight \(Z\) increases with exposure. For fixed \(\beta\), the weight does not directly depend on \(\alpha\); \(\alpha\) determines the prior mean when \(\beta\) is held fixed. To discuss a change in prior strength while keeping the prior mean unchanged, both Gamma parameters must be considered.

Deriving the Credibility Weight

The weight follows from the likelihood and prior under the stated model. The Poisson likelihood, as a function of \(\lambda\), is proportional to \(\lambda^N e^{-e\lambda}\). The Gamma prior density is proportional to \(\lambda^{\alpha-1}e^{-\beta\lambda}\). Multiplying them gives a posterior density proportional to \(\lambda^{\alpha+N-1}e^{-(\beta+e)\lambda}\), which is a Gamma density with shape \(\alpha+N\) and rate \(\beta+e\). Its mean gives the result.

\begin{aligned} f(\lambda\mid N) &\propto \lambda^N e^{-e\lambda} \lambda^{\alpha-1}e^{-\beta\lambda}\\ &= \lambda^{\alpha+N-1}e^{-(\beta+e)\lambda},\\ E[\lambda\mid N] &= \frac{\alpha+N}{\beta+e}. \end{aligned}

To identify the weights, split the numerator into the prior shape and observed count:

\frac{\alpha+N}{\beta+e} = \frac{e}{\beta+e}\frac{N}{e} + \frac{\beta}{\beta+e}\frac{\alpha}{\beta}.

The coefficients sum to one, since \(e/(\beta+e)+\beta/(\beta+e)=1\). Thus the posterior mean is exactly a credibility-weighted average of the observed rate and prior mean. This derivation relies on the Poisson likelihood with exposure as its mean multiplier and the Gamma shape-rate prior. If counts are not conditionally Poisson, the exposure is inconsistent with the rate’s units, or the risk rate changes during the period, this particular update and weight need not apply.

Interpretation: The prior contributes \(\beta\) exposure units to the weighting formula. The relative weight on observed experience is \(e/(e+\beta)\). If the prior mean \(\lambda_0\) is held fixed, then \(\alpha=\beta\lambda_0\); increasing \(\alpha\) while maintaining that mean also increases \(\beta\), and therefore increases prior strength relative to a fixed observed exposure.

Worked Applications

In each example, interest and mortality assumptions are not applicable: the claim count is not discounted, and no mortality model is used. The assumed probability model is the conditional Poisson count model with a Gamma shape-rate prior. Claim rates are reported per the stated exposure unit.

Worked Example: Combining a Claim Rate With a Prior Mean

A hypothetical portfolio segment has 80 vehicle-years of exposure and 20 claims. Assume \(N\mid\lambda\sim\operatorname{Poisson}(80\lambda)\) and \(\lambda\sim\operatorname{Gamma}(12,60)\), using the shape-rate convention. Calculate the posterior mean rate and the credibility weight on the segment’s observed rate. Interest rate: not applicable. Mortality assumption: not applicable.

Step 1: Find the prior mean and observed rate. The prior mean is \(12/60=0.20\) claims per vehicle-year. The observed rate is \(20/80=0.25\) claims per vehicle-year.

Step 2: Calculate the weight. The exposure is \(e=80\) and the prior rate parameter is \(\beta=60\), both expressed in compatible exposure units. Therefore, \(Z=80/(80+60)=80/140=4/7\), or approximately \(0.5714\). The prior receives weight \(1-Z=3/7\).

Step 3: Calculate the posterior mean. Using the credibility-weighted form gives \( (4/7)(0.25)+(3/7)(0.20)=1/7+0.6/7=1.6/7=0.228571\ldots \). Equivalently, the posterior is \(\operatorname{Gamma}(12+20,60+80)=\operatorname{Gamma}(32,140)\), with mean \(32/140=0.228571\ldots\). The posterior mean rate is 0.2286 claims per vehicle-year, rounded to four decimal places.

Check and common mistake avoided. The weights \(4/7\) and \(3/7\) sum to one, and the answer lies between the observed rate \(0.25\) and prior mean \(0.20\). A common mistake is to use \(N\) directly in the weighted average; the quantity being averaged with the prior mean is the rate \(N/e\), not the raw count.

Worked Example: Comparing Two Exposure Levels

Two hypothetical risks share the prior \(\operatorname{Gamma}(9,60)\), with rates measured per policy-year. Risk A has 30 policy-years and 9 claims; Risk B has 120 policy-years and 30 claims. Assume each count, conditional on its rate, is Poisson with mean equal to exposure times rate, and use the stated Gamma shape-rate prior for each risk. Find each posterior mean. Interest rate: not applicable. Mortality assumption: not applicable.

Step 1: Establish the shared prior mean. It is \(9/60=0.15\) claims per policy-year.

Step 2: Calculate Risk A. Its observed rate is \(9/30=0.30\). Its credibility weight is \(30/(30+60)=1/3\). The posterior mean is \( (1/3)(0.30)+(2/3)(0.15)=0.10+0.10=0.20 \) claims per policy-year. Checking by the posterior parameters gives \((9+9)/(60+30)=18/90=0.20\).

Step 3: Calculate Risk B. Its observed rate is \(30/120=0.25\). Its weight is \(120/(120+60)=2/3\). The posterior mean is \( (2/3)(0.25)+(1/3)(0.15)=0.166666\ldots+0.05=0.216666\ldots \) claims per policy-year. The posterior-parameter check is \((9+30)/(60+120)=39/180=0.216666\ldots\).

Risk A’s posterior mean is 0.2000, and Risk B’s is 0.2167 claims per policy-year, rounded to four decimal places. Risk B has more exposure, so its rate receives more weight, even though its observed rate is lower than Risk A’s. Each posterior mean lies between its own observed rate and the common prior mean.

Common mistake avoided. Giving both risks the same credibility weight because they share a prior ignores their different exposures. The weight depends on \(e\), so it is \(1/3\) for Risk A and \(2/3\) for Risk B.

Worked Example: Increasing Prior Strength at a Fixed Mean

An analyst wants to compare two hypothetical Gamma priors with the same mean rate of \(0.12\) claims per inspection-year but different strengths. The first is \(\operatorname{Gamma}(6,50)\); the second is \(\operatorname{Gamma}(18,150)\). A risk has 100 inspection-years and 25 claims. Assume \(N\mid\lambda\sim\operatorname{Poisson}(100\lambda)\), with the specified prior used for each separate calculation. Find both credibility weights and posterior means. Interest rate: not applicable. Mortality assumption: not applicable.

Step 1: Verify that the prior means match. For the first prior, \(6/50=0.12\). For the second, \(18/150=0.12\). The comparison therefore holds the prior mean fixed while increasing the shape and rate together.

Step 2: Calculate the first estimate. The observed rate is \(25/100=0.25\). With \(\beta=50\), the weight is \(100/(100+50)=2/3\). The posterior mean is \( (2/3)(0.25)+(1/3)(0.12)=0.166666\ldots+0.04=0.206666\ldots \). The posterior-parameter check is \((6+25)/(50+100)=31/150=0.206666\ldots\).

Step 3: Calculate the second estimate. With \(\beta=150\), the weight is \(100/(100+150)=0.40\). The posterior mean is \( (0.40)(0.25)+(0.60)(0.12)=0.10+0.072=0.172 \). The posterior-parameter check is \((18+25)/(150+100)=43/250=0.172\).

The first prior gives weight \(0.6667\) to the observed rate and a posterior mean of 0.2067; the stronger second prior gives weight \(0.4000\) and a posterior mean of 0.1720 claims per inspection-year. Rates are rounded to four decimal places. Both estimates lie between \(0.12\) and \(0.25\), and the stronger prior pulls the estimate closer to its mean.

Common mistake avoided. It would be incorrect to say that changing \(\alpha\) alone directly changes \(Z\). Here the prior mean is fixed, so increasing \(\alpha\) also increases \(\beta\), which increases the prior strength and lowers \(Z\). If \(\beta\) were held fixed instead, changing \(\alpha\) would also change the prior mean.

Reading Prior Strength Correctly

There are two useful ways to describe the prior’s influence. In the shape-rate parameterization, \(\beta\) determines the prior’s contribution to the exposure-based credibility weight, while \(\alpha/\beta\) determines the prior mean. If the mean is fixed at \(\lambda_0\), then \(\beta=\alpha/\lambda_0\), and the weight can also be written as

Z=\frac{e}{e+\alpha/\lambda_0} =\frac{e\lambda_0}{e\lambda_0+\alpha}.

This expression shows how shape represents strength when the prior mean is held constant: a larger \(\alpha\) corresponds to more prior exposure-equivalent weight. It is also consistent with the Gamma prior variance. At fixed mean \(\lambda_0\), the variance is \(\lambda_0^2/\alpha\), so increasing \(\alpha\) makes the prior more concentrated around that mean.

The formula also provides useful boundary checks. When \(e\) is very small relative to \(\beta\), \(Z\) is near zero and the estimate is near the prior mean. When \(e\) is very large relative to \(\beta\), \(Z\) is near one and the estimate is near the observed rate. These are limiting interpretations, not a reason to ignore model assumptions: the Poisson rate must remain meaningful over the exposure being summarized.

Exam trap: Do not use \(\alpha/(\alpha+e)\) as the observed-experience weight under the shape-rate convention. Exposure \(e\) and shape \(\alpha\) are not generally comparable quantities. Use \(e/(e+\beta)\); only after holding the prior mean fixed and substituting \(\beta=\alpha/\lambda_0\) may the equivalent expression involving \(\alpha\) be used.

Check Your Understanding

For each problem, use the Gamma shape-rate convention and express rates per the exposure unit given. Answers are shown at the end of each question for checking.

A hypothetical risk has 60 machine-years of exposure and 18 claims. Its prior is \(\operatorname{Gamma}(8,40)\). Under the conditional Poisson model, find the observed-rate credibility weight and posterior mean. (Answer: \(Z=0.6000\); posterior mean \(=0.2600\) claims per machine-year.)
Two priors have the same mean of \(0.10\) claims per unit of exposure: \(\operatorname{Gamma}(5,50)\) and \(\operatorname{Gamma}(20,200)\). A risk has exposure \(50\) and \(10\) claims. Find the posterior mean under each prior and state which gives more weight to the observed rate. (Answer: \(0.1500\) under the first prior and \(0.1200\) under the second; the first prior gives more observed-rate weight, \(0.5000\) versus \(0.2000\).)
A hypothetical portfolio has prior \(\operatorname{Gamma}(12,80)\), exposure \(40\), and \(4\) claims. Find the prior mean, observed rate, credibility weight, and posterior mean. (Answer: prior mean \(=0.1500\), observed rate \(=0.1000\), \(Z=1/3\), posterior mean \(=0.1333\) claims per exposure unit.)
A Gamma prior has shape \(15\) and mean rate \(0.25\) claims per exposure unit. A risk has exposure \(60\) and \(18\) claims. Determine the prior rate parameter, the observed-rate weight, and the posterior mean. (Answer: \(\beta=60\), \(Z=0.5000\), posterior mean \(=0.2750\) claims per exposure unit.)
Compare priors \(\operatorname{Gamma}(10,50)\) and \(\operatorname{Gamma}(20,100)\) for a risk with exposure \(75\) and \(24\) claims. Calculate each prior mean, credibility weight, and posterior mean, and explain the difference. (Answer: both prior means are \(0.2000\); weights are \(0.6000\) and \(0.4286\); posterior means are \(0.2720\) and \(0.2514\) claims per exposure unit, respectively. The second prior is stronger at the same mean and therefore gives less weight to the observed rate.)

What You'll Learn