Choosing Priors Through Prior Predictive Checks

Why Check a Prior Before Observing Experience?

A prior claim-rate distribution is a model for uncertainty about the risk’s underlying frequency, not merely a convenient starting value for a calculation. Before using it to update on experience, an actuary can ask what claim counts and aggregate losses it implies for a specified exposure. If those implications are incompatible with the hypothetical product, coverage period, or credible external knowledge, the prior needs reconsideration before any experience is observed.

A prior predictive check examines the distribution of data implied by a prior and the assumed data model, integrating over the uncertain parameter. It is a check on whether the prior and likelihood together generate reasonable hypothetical observations. It is not a comparison with the risk’s actual observed experience; that requires data and is a different diagnostic.

Definition: In a prior predictive check, specify a prior for an unknown parameter and a sampling model conditional on that parameter. The prior predictive distribution of a future observation is the sampling distribution averaged over the prior. For a claim-rate model, it describes the claim counts or totals the prior would generate before the risk’s experience is observed.

Poisson-Gamma Prior Predictive Counts

Suppose a risk has an unknown claim rate $\lambda$, measured per unit of exposure. Conditional on $\lambda$, let the claim count $N$ over exposure $e$ satisfy $N\mid\lambda\sim\operatorname{Poisson}(e\lambda)$. Assume the prior is $\lambda\sim\operatorname{Gamma}(\alpha,\beta)$, using the shape-rate parameterization: its density is proportional to $\lambda^{\alpha-1}e^{-\beta\lambda}$, its mean is $\alpha/\beta$, and its variance is $\alpha/\beta^2$. Here $\alpha>0$, $\beta>0$, and $e\geq0$.

This model assumes a constant rate over the stated exposure, conditional independence of claims under the Poisson model, and a rate prior that applies to the risk being considered. Exposure must be expressed in the same units used for $\lambda$. If the claim process is not reasonably modeled as Poisson conditional on its rate, or if the prior is not appropriate for the risk, the resulting predictive check does not validate that model.

P(N=n)=\int_0^\infty P(N=n\mid\lambda)f_\lambda(\lambda)\,d\lambda =\frac{\Gamma(\alpha+n)}{\Gamma(\alpha)n!} \left(\frac{\beta}{\beta+e}\right)^\alpha \left(\frac{e}{\beta+e}\right)^n,\qquad n=0,1,2,\ldots

This is a negative binomial predictive distribution under the stated shape-rate convention. Its probability of zero claims is $(\beta/(\beta+e))^\alpha$. The mean and variance can also be obtained without summing the probability mass function:

E[N]=e\frac{\alpha}{\beta},\qquad \operatorname{Var}(N)=e\frac{\alpha}{\beta} +e^2\frac{\alpha}{\beta^2}.

The first variance term is the average Poisson variation conditional on the rate. The second is the additional variation caused by uncertainty in the rate. In a prior predictive check, both terms matter: checking only the mean may miss a prior that assigns excessive probability to unusually high counts.

Why the Predictive Formula Holds

The predictive probability averages the Poisson likelihood over all possible rates, weighted by the prior density. This integration is the key step: using only the prior mean rate would give a Poisson plug-in distribution and would omit uncertainty about $\lambda$.

Multiply likelihood and prior.
The conditional count probability is $e^{-e\lambda}(e\lambda)^n/n!$. The Gamma prior density is $\beta^\alpha\lambda^{\alpha-1}e^{-\beta\lambda}/\Gamma(\alpha)$.

Combine terms involving the rate.
The product contains $\lambda^{\alpha+n-1}e^{-(\beta+e)\lambda}$, with the remaining factor $\beta^\alpha e^n/(\Gamma(\alpha)n!)$.

Integrate over all positive rates.
Using $\int_0^\infty \lambda^{k-1}e^{-c\lambda}\,d\lambda=\Gamma(k)/c^k$ for $k,c>0$, set $k=\alpha+n$ and $c=\beta+e$. This gives the predictive probability displayed above.

The mean follows from iterated expectation: $E[N]=E[E[N\mid\lambda]]=eE[\lambda]$. For the variance, the law of total variance gives $\operatorname{Var}(N)=E[\operatorname{Var}(N\mid\lambda)]+\operatorname{Var}(E[N\mid\lambda])$. Since a Poisson variable has conditional mean and variance $e\lambda$, the result is $eE[\lambda]+e^2\operatorname{Var}(\lambda)$. Substituting the Gamma mean and variance gives the stated formula.

Worked Example: Checking a Rate Prior Against a Count Range

A hypothetical fleet program is considering the prior $\lambda\sim\operatorname{Gamma}(2,4)$, where the rate is per vehicle-year. Check the prior predictive count distribution for a block of 2 vehicle-years, including its mean, variance, and probability of at least 3 claims. Assess the calculation by comparing the tail probability with a hypothetical planning limit of 15%.

Method and assumptions. Use the Poisson-Gamma prior predictive model, with $\alpha=2$, $\beta=4$, and $e=2$. The exposure is in vehicle-years, matching the units of the rate. The model is conditional Poisson and has no observed claims incorporated.

The predictive probability parameters are $\beta/(\beta+e)=4/6=2/3$ and $e/(\beta+e)=2/6=1/3$. The expected count is $e\alpha/\beta=2(2)/4=1$. The variance is

\operatorname{Var}(N)=2\frac{2}{4}+2^2\frac{2}{4^2} =1+\frac{1}{2}=1.5.

To calculate $P(N\geq3)$, first calculate the probabilities for 0, 1, and 2 claims. For $n=0$, $P(N=0)=(2/3)^2=4/9$. For $n=1$, the mass is $2(1/3)(2/3)^2=8/27$. For $n=2$, it is $\frac{2(3)}{2}(1/3)^2(2/3)^2=4/27$. Thus,

P(N\geq3)=1-\frac{4}{9}-\frac{8}{27}-\frac{4}{27} =\frac{3}{27}=\frac{1}{9}=0.1111.

The tail probability is about 11.11%, below the hypothetical 15% planning limit. That does not prove the prior is suitable; it says only that this particular threshold check does not flag it. Common mistake avoided: a Poisson model with the prior mean rate plugged in would have mean 1 but variance 1. It would ignore the additional $0.5$ of variance from rate uncertainty and therefore give a different tail probability.

Compare Predictive Implications, Not Just Prior Means

Two rate priors can have the same mean and still imply very different plausible ranges of claim counts. A prior predictive check therefore examines quantities suited to the decision: probabilities of exceeding operational thresholds, probabilities of no claims, or upper predictive quantiles. A threshold should be defined in context before comparing priors; it is not a universal pass-or-fail rule.

Worked Example: Same Mean Rate, Different Count Tails

A hypothetical equipment portfolio has 4 exposure-years. Compare candidate priors $\operatorname{Gamma}(1,2)$ and $\operatorname{Gamma}(4,8)$, with rates per exposure-year. Both have prior mean rate $0.5$. For each prior, calculate the predictive mean and variance and the probability of at least 5 claims.

Method and assumptions. Apply the Poisson-Gamma predictive formulas separately, using $e=4$. Both priors have expected count $4(0.5)=2$, so the comparison focuses on predictive uncertainty and the upper tail rather than the mean.

For $\operatorname{Gamma}(1,2)$, the predictive probability parameters are $p=2/(2+4)=1/3$ and $q=2/3$. Its mean is 2 and its variance is

4\frac{1}{2}+4^2\frac{1}{2^2}=2+4=6.

For shape $\alpha=1$, the predictive mass is geometric, so $P(N\geq5)=q^5=(2/3)^5=32/243=0.1317$.

For $\operatorname{Gamma}(4,8)$, $p=8/(8+4)=2/3$ and $q=1/3$. Its mean is also 2, while its variance is

4\frac{4}{8}+4^2\frac{4}{8^2}=2+1=3.

For this prior, $P(N=0)=p^4=16/81$. Successive probabilities follow the ratio $P(N=n+1)/P(N=n)=((4+n)/(n+1))q$. Therefore $P(N=1)=64/243$, $P(N=2)=160/729$, $P(N=3)=320/2187$, and $P(N=4)=560/6561$. Converting the first five probabilities to denominator 6561 gives $1296/6561$, $1728/6561$, $1440/6561$, $960/6561$, and $560/6561$. Their sum is $5984/6561$, so

P(N\geq5)=1-\frac{5984}{6561} =\frac{577}{6561}=0.0879.

Both priors imply an expected count of 2, but the first assigns more probability to 5 or more claims: 13.17% compared with 8.79%. Which is plausible depends on the hypothetical portfolio and the intended use. Common mistake avoided: matching prior means does not establish that two priors have equivalent predictive implications.

From Predictive Counts to Aggregate Totals

A prior predictive check can also examine aggregate loss, provided the severity assumptions are stated. A simple diagnostic assumes every claim has a fixed payment $c$, so total payment is $S=cN$. Then $E[S]=cE[N]$, $\operatorname{Var}(S)=c^2\operatorname{Var}(N)$, and count thresholds translate directly into total thresholds. This fixed-payment assumption is useful for a transparent check, but it is not a general model of insurance severity.

If individual payments $Y_1,Y_2,\ldots$ are instead independent and identically distributed, independent of $N$ and $\lambda$, and have finite mean and variance, then $S=\sum_{j=1}^N Y_j$ satisfies

E[S]=E[N]E[Y_1],\qquad \operatorname{Var}(S)=E[N]\operatorname{Var}(Y_1) +\operatorname{Var}(N)\big(E[Y_1]\big)^2.

The first variance term reflects variation in payment amounts conditional on the count. The second reflects count variation, including uncertainty in the rate. These formulas require the stated independence and identical-distribution assumptions; dependence between frequency and severity or a different payment structure requires an appropriate model.

Worked Example: Checking a Prior Predictive Loss Total

For a hypothetical repair program, let $\lambda\sim\operatorname{Gamma}(2,3)$ per service-unit-year and let the future exposure be 3 service-unit-years. For this check only, assume every claim produces a fixed payment of $1,200. Find the predictive mean and standard deviation of total payments and the probability that total payments exceed $3,600.

Method and assumptions. First find the Poisson-Gamma predictive count distribution with $\alpha=2$, $\beta=3$, and $e=3$. Then use $S=1200N$. This assumes a fixed payment per claim, no payment variation, and no other expenses.

The mean count is $3(2)/3=2$, and its variance is

\operatorname{Var}(N)=3\frac{2}{3}+3^2\frac{2}{3^2}=2+2=4.

Therefore $E[S]=1200(2)=\$2{,}400$. The variance of $S$ is $1200^2(4)=\$5{,}760{,}000$, so its standard deviation is $1200\sqrt{4}=\$2{,}400$.

Since $S=1200N$, the condition $S>\$3{,}600$ is equivalent to $N\geq4$, not $N\geq3$: three claims produce exactly $3,600. Here $p=\beta/(\beta+e)=1/2$. The probabilities for 0, 1, 2, and 3 claims are respectively $1/4$, $1/4$, $3/16$, and $1/8$. Hence

P(S>\$3{,}600)=P(N\geq4) =1-\left(\frac{1}{4}+\frac{1}{4}+\frac{3}{16}+\frac{1}{8}\right) =\frac{3}{16}=0.1875.

The probability is 18.75%, and the calculation checks against the fact that the first qualifying count is four. Common mistake avoided: treating “exceeds $3,600” as “at least $3,600” would incorrectly include $N=3$.

Practical Checks and Exam Traps

Prior predictive checks are most useful when the quantities being checked are connected to the intended application. A compact review can include the following:

Check units and exposure. A rate per vehicle-year cannot be multiplied by calendar years alone unless the exposure has been converted to vehicle-years.
Check the expected scale. Compare the implied mean count or total with a reasonable range for the hypothetical risk and coverage period.
Check dispersion and tails. Calculate probabilities of zero claims and of counts above a meaningful threshold, not just the expected count.
State the threshold precisely. Distinguish “at least” from “greater than,” especially when totals are discrete multiples of a fixed payment.
Keep the timing straight. Prior predictive checks integrate over the prior before using the risk’s observed experience; they do not update the prior.

A prior predictive result is a diagnostic, not a mechanical selection rule. A candidate prior may pass one threshold check and still imply implausible values elsewhere. Conversely, an unusual outcome is not automatically impossible: its relevance depends on the product, exposure, and decision being modeled. State the chosen checks and the hypothetical benchmark clearly. This approach is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Key takeaway: Combine the rate prior with the conditional claim model to obtain the distribution of hypothetical data before observing experience. Check its mean, dispersion, and decision-relevant tail probabilities, while keeping exposure units, severity assumptions, and threshold wording explicit.

Check Your Understanding

For each problem, use the shape-rate Gamma convention and the Poisson-Gamma prior predictive model unless the question asks for a conceptual explanation.

A hypothetical account has $\lambda\sim\operatorname{Gamma}(3,6)$ per exposure-year and future exposure $e=2$. Find the predictive mean and variance of its count, and its probability of zero claims. (Answer: $E[N]=1$, $\operatorname{Var}(N)=4/3$, $P(N=0)=0.4219$.)
For $\lambda\sim\operatorname{Gamma}(2,5)$ and $e=5$, calculate the predictive variance and $P(N\geq2)$. (Answer: $\operatorname{Var}(N)=4$, $P(N\geq2)=0.5000$.)
For $\lambda\sim\operatorname{Gamma}(1,2)$ and $e=2$, calculate $P(N\geq3)$. State why a Poisson distribution using only the prior mean rate is not the same predictive model. (Answer: $P(N\geq3)=0.1250$; the Poisson plug-in omits uncertainty in $\lambda$.)
A hypothetical program has $\lambda\sim\operatorname{Gamma}(2,4)$ per exposure-year, exposure $e=4$, and a fixed payment of $750 per claim. Find $E[S]$ and $P(S>\$2{,}250)$. (Answer: $E[S]=\$1{,}500$, $P(S>\$2{,}250)=0.1875$.)
Two candidate priors produce the same prior mean claim rate. Name two additional prior predictive quantities that can reveal meaningful differences between them, and explain why. (Answer: For example, the predictive variance and $P(N\geq k)$ for a relevant threshold $k$; they reflect dispersion and tail behavior that the mean alone does not show.)

What You'll Learn