Bayesian Risk Models for Insurance Frequencies

Why Model a Risk’s Claim Rate?

Two insured risks can have different underlying claim frequencies even when their observed claim counts happen to match. A Bayesian frequency model represents that difference with a risk-specific parameter: the rate that is treated as fixed for a particular risk but uncertain before its experience is observed. The model then combines prior information about possible rates with the evidence in the risk’s claim history.

The distinction between a rate and a count is essential. A claim count is an outcome observed over a specified exposure; the rate is a parameter used to describe the distribution of possible outcomes. For example, observing one claim over two vehicle-years does not establish that the risk’s true rate is exactly one-half claim per vehicle-year. It provides evidence about that rate.

Definition: Let \(\lambda\) be the risk-specific claim rate per exposure unit. Let \(E\) be the risk’s exposure and \(N\) its observed claim count. In the Poisson frequency model, conditional on the rate, \(N\mid\lambda\sim\operatorname{Poisson}(E\lambda)\). The prior distribution describes uncertainty about \(\lambda\) before observing \(N\); the likelihood describes how probable the observed count is for each candidate value of \(\lambda\); and the posterior distribution describes uncertainty about \(\lambda\) after using both the prior and the observed count.

This model assumes that the rate is constant for the risk over the exposure being modeled, that claims follow a Poisson distribution conditional on that rate, and that exposure is measured consistently with the rate’s time or policy basis. In a Poisson model, the conditional mean and variance of the count are both \(E\lambda\). The model may be unsuitable if, for example, claim counts are materially overdispersed relative to Poisson variation or the risk’s rate changes over the observation period.

A prior can be discrete, assigning probabilities to a finite set of possible rates, or continuous, assigning probability across a range. The examples here use discrete priors so that each posterior probability and calculation can be seen directly. The prior is a model input; it should represent relevant information available before the particular risk’s observed count is used, not a probability chosen after seeing that count.

Bayes’ Rule for a Frequency Model

For a discrete prior with candidate rates \(\lambda_1,\ldots,\lambda_m\) and prior probabilities \(\pi_1,\ldots,\pi_m\), where \(\sum_{j=1}^m\pi_j=1\), the posterior probability of candidate rate \(\lambda_j\) after observing count \(n\) is

P(\lambda=\lambda_j\mid N=n) = \frac{P(N=n\mid\lambda_j)\pi_j} {\sum_{k=1}^{m}P(N=n\mid\lambda_k)\pi_k}.

Under the conditional Poisson model, the probability of observing \(n\) claims at rate \(\lambda_j\) and exposure \(E\) is

P(N=n\mid\lambda_j) = \frac{(E\lambda_j)^n e^{-E\lambda_j}}{n!}.

This probability, considered as a function of the candidate rate while the observed count is held fixed, is the likelihood. It is not itself a posterior probability: the prior probability must also be included, and the resulting weights must be normalized to sum to one.

Definition: For a discrete posterior over rates, the posterior mean claim rate is \(\sum_{j=1}^{m}\lambda_jP(\lambda=\lambda_j\mid N=n)\). It summarizes the posterior distribution as a single rate. Under squared-error loss for estimating \(\lambda\), the posterior mean minimizes the conditional expected squared error.

The mean’s squared-error property follows by writing the posterior expected loss for a proposed estimate \(a\). Let \(\mu=E[\lambda\mid N=n]\). Then

E[(\lambda-a)^2\mid N=n] = E[(\lambda-\mu)^2\mid N=n]+(a-\mu)^2.

To see this, expand \(\lambda-a=(\lambda-\mu)+(\mu-a)\). The cross term has expectation \(2(\mu-a)E[\lambda-\mu\mid N=n]=0\). Since \((a-\mu)^2\) is nonnegative, the expected squared error is minimized at \(a=\mu\). This does not mean that the posterior mean is the only useful estimate: a different decision criterion can call for a different summary.

Worked Bayesian Frequency Updates

Worked Example: One Claim and Two Candidate Rates

A hypothetical small-equipment portfolio assigns equal prior probabilities to a risk rate of \(0.08\) or \(0.24\) claims per equipment-year. The risk has \(E=1.5\) equipment-years and one observed claim. Assume no discounting or mortality; the frequency model is \(N\mid\lambda\sim\operatorname{Poisson}(E\lambda)\). Find the posterior probability of the higher rate and the posterior mean rate.

Method. Calculate the Poisson likelihood at each candidate rate, multiply each likelihood by its prior probability, and normalize. This method avoids treating the observed frequency as the risk’s known rate.

At \(\lambda=0.08\), \(E\lambda=1.5(0.08)=0.12\), so \(P(N=1\mid\lambda=0.08)=0.12e^{-0.12}=0.106430\). The prior-weighted value is \(0.5(0.106430)=0.053215\).

At \(\lambda=0.24\), \(E\lambda=1.5(0.24)=0.36\), so \(P(N=1\mid\lambda=0.24)=0.36e^{-0.36}=0.251164\). The prior-weighted value is \(0.5(0.251164)=0.125582\). The total is \(0.053215+0.125582=0.178797\).

Therefore, the posterior probability of the higher rate is \(0.125582/0.178797=0.7024\), to four decimals. The posterior probability of the lower rate is \(1-0.7024=0.2976\), using unrounded values for the calculation. The posterior mean is \(0.08(0.2976)+0.24(0.7024)=0.1924\) claims per equipment-year, to four decimals. Using unrounded posterior probabilities gives the same rounded result.

Check. The two posterior probabilities sum to \(1.0000\), and the posterior mean, \(0.1924\), lies between the two possible rates.

Worked Example: No Claims Do Not Imply a Zero Rate

A hypothetical specialty contractor has a prior distribution assigning probabilities \(0.20\), \(0.50\), and \(0.30\) to rates of \(0.05\), \(0.10\), and \(0.20\) claims per contract-year, respectively. The contractor has five contract-years of exposure and no observed claims. Assume no discounting or mortality and use \(N\mid\lambda\sim\operatorname{Poisson}(E\lambda)\). Find the posterior probabilities and posterior mean.

Method. With zero observed claims, the Poisson likelihood is \(e^{-E\lambda}\). Weight each candidate rate’s likelihood by its prior probability, then normalize. The common mistake is to assign a zero rate merely because the observed count is zero.

For \(\lambda=0.05\), \(E\lambda=5(0.05)=0.25\), the likelihood is \(e^{-0.25}=0.778801\), and the prior-weighted value is \(0.20(0.778801)=0.155760\).

For \(\lambda=0.10\), \(E\lambda=0.50\), the likelihood is \(e^{-0.50}=0.606531\), and the prior-weighted value is \(0.50(0.606531)=0.303265\). For \(\lambda=0.20\), \(E\lambda=1.00\), the likelihood is \(e^{-1}=0.367879\), and the prior-weighted value is \(0.30(0.367879)=0.110364\).

The sum of the weighted values is \(0.155760+0.303265+0.110364=0.569389\). Dividing each value by this total gives posterior probabilities \(0.2735\), \(0.5326\), and \(0.1939\), respectively, to four decimals. They sum to \(1.0000\) after rounding.

The posterior mean rate is \(0.05(0.2735)+0.10(0.5326)+0.20(0.1939)=0.1057\) claims per contract-year, to four decimals. The zero count makes lower rates more plausible, but the posterior mean remains positive because the prior assigned probability to positive rates and zero claims can occur at those rates.

Check. The posterior mean lies between \(0.05\) and \(0.20\), and the middle rate retains the largest posterior probability. Both are consistent with the observed zero count and the prior.

Worked Example: Prior Probabilities Change the Update

A hypothetical mobile-equipment risk has rate \(0.10\) with prior probability \(0.8005\) or rate \(0.30\) with prior probability \(0.1995\), in claims per equipment-year. It has two equipment-years of exposure and one claim. Assume no discounting or mortality and \(N\mid\lambda\sim\operatorname{Poisson}(E\lambda)\). Find the posterior probability of the higher rate and the posterior mean.

Method. Apply Bayes’ rule using the stated unequal prior probabilities. This example highlights the trap of normalizing likelihoods alone, as if the candidate rates had equal prior probabilities.

For the lower rate, \(E\lambda=2(0.10)=0.20\), and the likelihood is \(0.20e^{-0.20}=0.163746\). Its prior-weighted value is \(0.8005(0.163746)=0.131079\). For the higher rate, \(E\lambda=2(0.30)=0.60\), and the likelihood is \(0.60e^{-0.60}=0.329287\). Its prior-weighted value is \(0.1995(0.329287)=0.065693\).

The total weight is \(0.131079+0.065693=0.196772\). Thus the posterior probability of the higher rate is \(0.065693/0.196772=0.3339\), to four decimals. The posterior probability of the lower rate is \(1-0.3339=0.6661\). The posterior mean is \(0.10(0.6661)+0.30(0.3339)=0.1668\) claims per equipment-year, to four decimals, with unrounded probabilities used before rounding.

Common mistake and correction. Normalizing the likelihoods alone gives \(0.329287/(0.163746+0.329287)=0.6679\) for the higher rate. That calculation omits the unequal prior probabilities and is not the posterior for this problem. Multiplying by the prior probabilities first gives the correct posterior probability, \(0.3339\). As a check, the posterior mean \(0.1668\) lies between the two candidate rates.

Interpreting the Model and Its Limits

The posterior is a probability distribution over the unknown rate, not a distribution of future claims. The observed count \(N\) is an outcome; \(\lambda\) is the risk-specific parameter. A posterior mean rate can be used as a summary for estimating the risk’s frequency, but a forecast of future claims also depends on the future exposure and the conditional frequency model.

Exposure must use the same basis as the rate. If a rate is stated per policy-year, a half-year of coverage corresponds to \(E=0.5\), not \(E=1\). Changing the exposure unit without changing the numerical rate changes the model. Likewise, a claim count by itself is not a frequency unless its associated exposure is understood.

Bayesian updating is useful because it makes the influence of prior information explicit, particularly when a risk has little experience. It does not remove the need for judgment: the prior, exposure definition, and conditional count model must be appropriate for the application. If risks differ in ways not represented by the model, or the rate is not stable over the modeled period, the resulting posterior can be precise-looking but poorly calibrated.

Key takeaway: A Bayesian frequency update starts with a risk-specific rate and a prior distribution, evaluates the observed count under each candidate rate, weights by the prior, and normalizes to obtain the posterior. Keep the claim count, exposure, prior, likelihood, and posterior distinct at every step.

Check Your Understanding

For each question, use a conditional Poisson frequency model and state posterior probabilities to four decimals and rates to four decimals.

A hypothetical risk has prior probability \(0.5\) at each rate \(0.04\) and \(0.12\) claims per unit of exposure. It has \(E=2\) and \(N=0\). Find the posterior probability of the higher rate and the posterior mean. (Answer: \(0.4601\); \(0.0768\) claims per exposure unit.)
A hypothetical service fleet has prior probabilities \(0.75\) and \(0.25\) at annual rates \(0.10\) and \(0.20\). It has three fleet-years of exposure and one claim. Find the posterior probability of the higher rate and the posterior mean. (Answer: \(0.3306\); \(0.1331\) claims per fleet-year.)
A hypothetical risk has prior probabilities \(0.60\) and \(0.40\) at rates \(0.05\) and \(0.15\) claims per unit of exposure. It has \(E=4\) and \(N=0\). Find both posterior probabilities and the posterior mean. (Answer: \(0.6911\) at rate \(0.05\), \(0.3089\) at rate \(0.15\); posterior mean \(0.0809\).)
Rates are annual. A hypothetical insured has equal prior probabilities at rates \(0.20\) and \(0.60\), covers six months, and has no claims. Convert the exposure to years, then find the posterior probability of the higher rate and posterior mean. (Answer: \(E=0.5\) years; \(0.4502\); \(0.3801\) claims per year.)
A hypothetical risk has equal prior probabilities at rates \(0.10\) and \(0.30\) claims per exposure unit. It has \(E=1\) and \(N=2\). Find the posterior probability of the higher rate and posterior mean. (Answer: \(0.8805\); \(0.2761\) claims per exposure unit.)

What You'll Learn