Simulating Posterior Predictive Claim Distributions

Why Simulate a Posterior Predictive Distribution?

A forecast of future claims is uncertain for two related reasons: future outcomes vary even if the model parameters were known, and the parameters themselves remain uncertain after observing data. A posterior predictive simulation represents both sources of uncertainty. It first draws plausible parameters from their posterior distribution, then draws future claims from the model conditional on those parameters.

This is different from simulating future claims at a single estimated parameter value. A plug-in simulation includes variation in future claims conditional on that value, but omits uncertainty about the value itself. The resulting forecast distribution can therefore be too concentrated. The simulation method here applies to frequency or aggregate-loss models, provided the posterior parameter draws and conditional claim model are specified.

Definition: Let $\theta$ denote model parameters, $D$ the observed data, and $Y$ a future claim quantity. The posterior predictive distribution is $P(Y\in A\mid D)=\int P(Y\in A\mid\theta,D)\,\pi(\theta\mid D)\,d\theta$, where $\pi(\theta\mid D)$ is the posterior distribution of the parameters. A simulation approximates this mixture by drawing $\theta$ from its posterior and then drawing $Y$ from its conditional predictive model.

The Two-Stage Simulation

Suppose the posterior distribution of the parameters can be sampled, and the conditional distribution of future claims can be simulated. For each simulation replicate, use one posterior parameter draw and generate future experience conditional on that draw. Repeating the pair of steps produces draws from the posterior predictive distribution.

Draw parameters.
For replicate $m$, draw $\theta^{(m)}$ from $\pi(\theta\mid D)$, or take a draw from a valid posterior simulation already produced.

Draw future claims conditionally.
Generate $Y^{(m)}$ from the future-data model given $\theta^{(m)}$ and the forecast exposure or period.

Summarize the predictive draws.
Use the $M$ values $Y^{(1)},\ldots,Y^{(M)}$ to estimate means, quantiles, intervals, or tail probabilities.

The pairing in the first two steps matters: each future outcome must be generated using the parameter draw from that same replicate. If several parameters are uncertain, draw their joint posterior distribution. Drawing parameters independently is appropriate only when the posterior actually factorizes or when the simulation method otherwise preserves their dependence.

Poisson Frequency With a Gamma Posterior

Consider a claim count $N$ over exposure $t$. Assume that, conditional on frequency rate $\lambda$, $N\mid\lambda\sim\operatorname{Poisson}(t\lambda)$. Suppose the posterior distribution after observing data is $\lambda\mid D\sim\operatorname{Gamma}(\alpha,\beta)$, using the shape-rate convention: the posterior mean is $\alpha/\beta$ and variance is $\alpha/\beta^2$. The two-stage simulation draws $\lambda$ from this Gamma posterior and then draws $N$ from the corresponding Poisson distribution.

The model presumes a constant rate over the forecast exposure, conditional independence of future events given $\lambda$, and a Poisson count likelihood. It does not automatically account for changes in exposure mix, time trends, or dependence between claims. Those features require an appropriate model before simulation.

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

The variance has a process component and a parameter component. To see why, apply the law of total variance:

\begin{aligned} \operatorname{Var}(N\mid D) &=E[\operatorname{Var}(N\mid\lambda,D)\mid D] +\operatorname{Var}(E[N\mid\lambda,D]\mid D)\\ &=E[t\lambda\mid D]+\operatorname{Var}(t\lambda\mid D)\\ &=t\frac{\alpha}{\beta}+t^2\frac{\alpha}{\beta^2}. \end{aligned}

The first term is the expected conditional Poisson variance. The second is the variance of the conditional mean across posterior draws. A plug-in forecast at $E[\lambda\mid D]$ retains the first term but not the second.

Worked Example: Updating Frequency and Checking Predictive Variance

A hypothetical portfolio uses a Poisson claim-count model with rate $\lambda$ per exposure unit. Before experience is observed, $\lambda\sim\operatorname{Gamma}(2,4)$, with shape 2 and rate 4. The portfolio then records 3 claims over 2 exposure units. Find the posterior distribution and the mean and variance of the count over 3 future exposure units.

Method and posterior update. For a Poisson rate with a Gamma shape-rate prior, observing $s$ claims over exposure $e$ gives posterior $\operatorname{Gamma}(2+s,4+e)$. Thus the posterior is $\operatorname{Gamma}(5,6)$. Its mean is $5/6$, and its variance is $5/6^2=5/36$.

Predictive mean. With future exposure $t=3$, the expected count is $3(5/6)=15/6=2.5$ claims.

Predictive variance. The conditional process component is $tE[\lambda\mid D]=3(5/6)=2.5$. The parameter component is $t^2\operatorname{Var}(\lambda\mid D)=9(5/36)=45/36=1.25$. Therefore the predictive variance is $2.5+1.25=3.75$ claims squared.

Check. The variance exceeds the mean by $3.75-2.5=1.25$, exactly the parameter-uncertainty component. The common mistake is to report 2.5 as the variance by treating the posterior mean rate as known; that is only the conditional Poisson variance at the plug-in rate.

Worked Simulation: From Posterior Draws to Count Summaries

A direct simulation is useful when the predictive distribution is not easy to summarize analytically. The next example uses only four replicates so every calculation can be inspected. Such a small simulation demonstrates the mechanics, not a reliable precision level for a real forecast.

Worked Example: Simulating a Future Claim Count

Suppose a hypothetical claim-count model has posterior $\lambda\mid D\sim\operatorname{Gamma}(1,1)$, in shape-rate form, and the forecast exposure is $t=1$. Use the four posterior inverse-transform uniforms $0.3935,\ 0.6321,\ 0.7769,\ 0.8647$. For a shape-1, rate-1 Gamma distribution, the inverse transform is $\lambda=-\ln(1-U)$; the resulting rate draws are $0.5,\ 1,\ 1.5,\ 2$, respectively. For the conditional Poisson draws, use uniforms $0.8,\ 0.4,\ 0.9,\ 0.2$, in that order.

Replicate 1. At $\lambda=0.5$, the Poisson probability of zero claims is $e^{-0.5}=0.6065$, and the cumulative probability through one claim is $e^{-0.5}(1+0.5)=0.9098$. Since $0.6065<0.8\leq0.9098$, draw $N=1$.

Replicate 2. At $\lambda=1$, the cumulative probability through zero is $e^{-1}=0.3679$, and through one is $e^{-1}(1+1)=0.7358$. Since $0.3679<0.4\leq0.7358$, draw $N=1$.

Replicate 3. At $\lambda=1.5$, the cumulative probability through two is $e^{-1.5}(1+1.5+1.5^2/2)=0.8088$, and through three is $0.9344$. Since $0.8088<0.9\leq0.9344$, draw $N=3$.

Replicate 4. At $\lambda=2$, the cumulative probability through zero is $e^{-2}=0.1353$, and through one is $e^{-2}(1+2)=0.4060$. Since $0.1353<0.2\leq0.4060$, draw $N=1$.

The simulated counts are $1,1,3,1$. Their simulated mean is $(1+1+3+1)/4=1.5$. The empirical probability of at least 2 claims is $1/4=0.2500$. Sorted counts are $1,1,1,3$; using the empirical inverse-CDF quantile convention, the simulated 95% interval with endpoints at the 2.5th and 97.5th percentiles is $[1,3]$. The mean of 1.5 is not expected to equal the exact predictive mean of 1 for a four-draw simulation; the discrepancy is sampling noise, not a change in the model.

Check. Every future count was generated conditional on its own rate draw. The common mistake is to use the average rate for all four Poisson draws, which would remove between-draw parameter uncertainty from the simulation.

Simulating Aggregate Losses

For aggregate loss, the conditional future-data step may itself involve several stages. For example, conditional on frequency rate $\lambda$, draw $N\sim\operatorname{Poisson}(t\lambda)$; conditional on a severity parameter $\theta$, draw claim amounts $X_1,\ldots,X_N$; and calculate $S=\sum_{j=1}^{N}X_j$, with $S=0$ when $N=0$. If both parameters are uncertain, draw the pair $(\lambda,\theta)$ jointly from its posterior at the start of each replicate.

The following example assumes exponential severities with mean $\theta$, conditional independence of severities given $\theta$, and conditional independence of frequency and severity given the paired posterior draw. A real application should use the dependence structure supported by its model rather than impose independence for convenience.

Worked Example: Two-Stage Simulation of Aggregate Loss

A hypothetical portfolio has one unit of future exposure. Its posterior draws for frequency rate $\lambda$ and exponential severity mean $\theta$, in dollars, are the four pairs $(0.5,100), (1,150), (1.5,100), (2,200)$. These are supplied draws from the joint posterior. Use conditional Poisson uniforms $0.8,0.4,0.9,0.2$ to generate counts. For each severity, use a Uniform(0,1) survival-scale draw $V$, so $X=-\theta\ln(V)$. Use the listed severity draws: replicate 1 has $V=e^{-1}$; replicate 2 has $V=e^{-1}$; replicate 3 has $V=e^{-1},e^{-0.5},e^{-2}$; replicate 4 has $V=e^{-0.5}$. Estimate the mean aggregate loss and the probability aggregate loss exceeds \$200.

Frequency stage. The same Poisson calculations as in the previous count example give counts $1,1,3,1$ for rates $0.5,1,1.5,2$. The count in replicate 3 is 3 because its conditional cumulative probability through two is $0.8088$, below its uniform $0.9$, while through three it is $0.9344$.

Severity and aggregate stage. Since $-\ln(e^{-c})=c$, the severity amounts are computed directly. Replicate 1 has one claim of $-100\ln(e^{-1})=\$100$, giving $S^{(1)}=\$100$. Replicate 2 has one claim of $-150\ln(e^{-1})=\$150$, giving $S^{(2)}=\$150$. Replicate 3 has severities $-100\ln(e^{-1})=\$100$, $-100\ln(e^{-0.5})=\$50$, and $-100\ln(e^{-2})=\$200$, giving $S^{(3)}=\$350$. Replicate 4 has one severity $-200\ln(e^{-0.5})=\$100$, giving $S^{(4)}=\$100$.

Summaries and check. The simulated mean is $(100+150+350+100)/4=\$175.00$. Only one of the four totals exceeds \$200, so the estimated probability is $1/4=0.2500$. The total of \$350 is correctly above the threshold, while the total of \$200 is not: the question asks for strictly greater than \$200. The common mistake is to compare an individual claim with the threshold rather than compare the replicate’s aggregate loss.

Reading Simulation Summaries Carefully

For $M$ simulated predictive outcomes $Y^{(1)},\ldots,Y^{(M)}$, the sample mean estimates the posterior predictive mean. The empirical probability of exceeding a threshold $c$ is the number of outcomes above $c$, divided by $M$. Predictive quantiles are estimated from the sorted outcomes. State the quantile convention when the number of draws is small, because interpolation and order-statistic conventions can produce different endpoints.

A simulated probability also has Monte Carlo error: increasing $M$ reduces simulation noise but does not repair a misspecified model or unrepresentative posterior draws. For an estimated event probability $\hat p$, a common approximate Monte Carlo standard error is $\sqrt{\hat p(1-\hat p)/M}$. For example, with $\hat p=0.25$ and $M=10{,}000$, this is $\sqrt{0.25(0.75)/10{,}000}=0.0043$, to four decimal places. This quantifies simulation precision for that estimated probability, not the uncertainty in the future claims themselves; the latter is what the predictive distribution represents.

Key takeaway: A posterior predictive replicate consists of a posterior parameter draw followed by a future claim draw conditional on that parameter. Preserve joint posterior dependence, use the correct forecast exposure and conditional likelihood, and distinguish Monte Carlo error from predictive uncertainty.

Check Your Understanding

For each problem, keep posterior parameter uncertainty separate from conditional claim variation.

A Poisson rate has prior $\operatorname{Gamma}(2,3)$ in shape-rate form. The observed experience is 4 claims over 2 exposure units. Find the posterior distribution and the predictive mean and variance for 1.5 future exposure units. (Answer: posterior $\operatorname{Gamma}(6,5)$; predictive mean $1.8000$, variance $2.3400$.)
For a future count with exposure 1, the posterior rate is $\operatorname{Gamma}(3,2)$, shape-rate. Find the predictive mean and variance, and identify the parameter-uncertainty part of the variance. (Answer: mean $1.5000$, variance $2.2500$, parameter component $0.7500$.)
Five simulated future counts are $0,1,1,2,4$. Estimate the probability of at least 2 claims using the empirical proportion. (Answer: $2/5=0.4000$.)
Four simulated aggregate losses, in dollars, are $\$80,\ \$200,\ \$240,\ \$500$. Estimate the probability of loss strictly greater than $\$200$, and the simulated mean loss. (Answer: probability $2/4=0.5000$; mean $\$255.00$.)
An estimated predictive tail probability is $0.3000$ from $M=1{,}000$ independent simulation replicates. Calculate its approximate Monte Carlo standard error to four decimal places. (Answer: $\sqrt{0.3(0.7)/1{,}000}=0.0145$.)

What You'll Learn