Posterior Predictive Distributions for Future Claims

From Parameter Uncertainty to Future Claims

Bayesian updating describes what observed experience implies about an unknown claim parameter. Forecasting asks a different question: given that experience, what outcomes might occur in the future? The posterior distribution answers the first question; the posterior predictive distribution answers the second. The distinction matters because a claim outcome is random even if the parameter were known, and the parameter itself is uncertain after observing finite experience.

For example, a posterior probability of 0.7 assigned to a claim probability of 0.04 does not mean there is a 70% chance that a future exposure produces a claim. It means the parameter value 0.04 has posterior probability 0.7. To find the probability of a future claim, average the claim probability over all plausible parameter values, weighted by their posterior probabilities.

Definition: Let $\theta$ be an unknown model parameter, let $D$ denote observed data, and let $Y$ denote a future observation. Suppose the future observation has conditional probability mass function $p(y\mid\theta)$ or conditional density $f(y\mid\theta)$, and the posterior distribution of $\theta$ given $D$ is $p(\theta\mid D)$ or $\pi(\theta\mid D)$. The posterior predictive distribution is $$ p(y\mid D)=\int p(y\mid\theta)\,\pi(\theta\mid D)\,d\theta. $$ For a discrete parameter with posterior probabilities $P(\theta_j\mid D)$, the integral is replaced by the sum $$ P(Y=y\mid D)=\sum_j P(Y=y\mid\theta_j)\,P(\theta_j\mid D). $$

The formula assumes that, conditional on $\theta$, the future outcome follows the stated likelihood and is independent of the observed data. The posterior distribution summarizes how the observed data update beliefs about $\theta$; the conditional likelihood describes how future outcomes arise when $\theta$ is specified. The predictive distribution combines both sources of uncertainty.

This construction does not require a conjugate model. If the posterior and likelihood permit an analytic integral, it can be evaluated directly. Otherwise, numerical integration or simulation can approximate it. In either case, the object being calculated is a distribution over possible future outcomes, not a distribution over parameter values.

Why Averaging the Likelihood Gives the Predictive Distribution

The justification is the law of total probability. If $\theta$ is unknown, partition or average over its possible values. For a discrete parameter, the events $\{\theta=\theta_j\}$ exhaust the posterior possibilities, so

P(Y=y\mid D) =\sum_j P(Y=y,\theta=\theta_j\mid D) =\sum_j P(Y=y\mid\theta=\theta_j,D)P(\theta=\theta_j\mid D).

Under the model assumption that $Y$ is conditionally independent of $D$ given $\theta$, $P(Y=y\mid\theta,D)=P(Y=y\mid\theta)$. Substitution gives the discrete predictive formula. For a continuous parameter, the same averaging argument becomes integration with respect to the posterior distribution:

P(Y\in A\mid D) =\int P(Y\in A\mid\theta)\,\pi(\theta\mid D)\,d\theta

for any set of future outcomes $A$. Thus the formula is not an additional assumption about a new kind of likelihood. It follows by averaging the conditional future-outcome model over the posterior uncertainty about its parameter.

Worked Example: Predicting a Claim for One Future Exposure

A hypothetical portfolio has a binary claim indicator for each exposure. After observing historical experience, an actuary has a posterior distribution for the claim probability $\theta$: $P(\theta=0.02\mid D)=0.30$ and $P(\theta=0.08\mid D)=0.70$. Conditional on $\theta$, the future exposure has a Bernoulli claim indicator $Y$, with $P(Y=1\mid\theta)=\theta$. Find the predictive probabilities of a claim and no claim.

Method. Average the Bernoulli likelihood over the posterior probabilities. First, for a claim:

P(Y=1\mid D) =(0.02)(0.30)+(0.08)(0.70) =0.006+0.056 =0.062.

For no claim, the conditional probabilities are $1-0.02=0.98$ and $1-0.08=0.92$. Therefore,

P(Y=0\mid D) =(0.98)(0.30)+(0.92)(0.70) =0.294+0.644 =0.938.

The two predictive probabilities sum to $0.062+0.938=1.000$, as required. The posterior probabilities $0.30$ and $0.70$ are probabilities for parameter values, not probabilities of claim outcomes. The method avoids the common mistake of reporting $0.70$ as the claim probability merely because it is the larger posterior weight.

Predictive Probabilities for Count Outcomes

The same averaging principle applies when the future outcome is a claim count. Suppose the conditional count model is Poisson with parameter $\lambda$, so $P(N=n\mid\lambda)=e^{-\lambda}\lambda^n/n!$. If the posterior for $\lambda$ has several possible values, each value gives a conditional distribution for the future count. The predictive probability of a count is the posterior-weighted average of those conditional probabilities.

The count model and posterior must refer to compatible exposure units. For instance, if $\lambda$ is the expected count in one policy-year, the future Poisson likelihood must also describe one policy-year. The calculation below treats the posterior distribution as already obtained from historical data; it focuses on constructing the future distribution.

Worked Example: Forecasting a Future Count Under Rate Uncertainty

For a hypothetical block of policies, the posterior distribution for the Poisson rate per year is $P(\lambda=1\mid D)=0.25$ and $P(\lambda=2\mid D)=0.75$. Conditional on $\lambda$, the claim count $N$ in the next year is Poisson with rate $\lambda$. Calculate $P(N=0\mid D)$, $P(N=1\mid D)$, and $P(N\leq 1\mid D)$.

Method. Apply the Poisson likelihood at each parameter value, then weight by its posterior probability. For zero claims:

P(N=0\mid D) =(0.25)e^{-1}+(0.75)e^{-2} =0.25(0.367879)+0.75(0.135335) =0.193471.

For exactly one claim, the conditional Poisson probability is $e^{-\lambda}\lambda$:

P(N=1\mid D) =(0.25)e^{-1}(1)+(0.75)e^{-2}(2) =0.091970+0.203003 =0.294973.

The event $N\leq 1$ consists of the disjoint outcomes $N=0$ and $N=1$, so

P(N\leq 1\mid D)=0.193471+0.294973=0.488444.

To four decimal places, the requested probabilities are $0.1935$, $0.2950$, and $0.4884$. As a check, the last probability is the sum of the first two and is between zero and one. The method avoids treating either posterior rate value as certain and reporting only its conditional Poisson probabilities.

Several Future Claims Share the Same Uncertain Parameter

When predicting multiple future observations for the same risk or portfolio, it is often appropriate to use one shared parameter value for all of them. Conditional on that parameter, the observations may be independent. After averaging over the posterior, however, their predictive outcomes need not be independent: observing one future outcome would provide information about the shared parameter and therefore affect predictions for the others.

For a future vector $Y_1,\ldots,Y_m$, the joint posterior predictive distribution is

p(y_1,\ldots,y_m\mid D) =\int \left\{\prod_{k=1}^{m}p(y_k\mid\theta)\right\} \pi(\theta\mid D)\,d\theta,

when the future observations are conditionally independent given $\theta$. The product is formed first, using one common $\theta$, and then averaged over the posterior. In general, this is not the same as multiplying the separate marginal predictive probabilities. The distinction is important for probabilities about totals or joint events.

Worked Example: Two Future Exposures With a Shared Claim Probability

A hypothetical class has two future exposures. Its claim probability is uncertain, with $P(\theta=0.10\mid D)=0.40$ and $P(\theta=0.30\mid D)=0.60$. Conditional on the common $\theta$, the two claim indicators are independent Bernoulli variables. Find the predictive probabilities of zero, exactly one, and two claims.

Method. For each parameter value, first calculate the binomial probabilities for two exposures; then average each probability over the posterior. For zero claims:

P(N=0\mid D) =(0.40)(0.90^2)+(0.60)(0.70^2) =(0.40)(0.81)+(0.60)(0.49) =0.324+0.294=0.618.

For exactly one claim, the conditional probability is $2\theta(1-\theta)$:

P(N=1\mid D) =(0.40)\{2(0.10)(0.90)\}+(0.60)\{2(0.30)(0.70)\} =(0.40)(0.18)+(0.60)(0.42) =0.072+0.252=0.324.

For two claims:

P(N=2\mid D) =(0.40)(0.10^2)+(0.60)(0.30^2) =(0.40)(0.01)+(0.60)(0.09) =0.004+0.054=0.058.

The probabilities sum to $0.618+0.324+0.058=1.000$. In particular, the probability of at least one claim is $1-0.618=0.382$. The method avoids treating the two exposures as if each had a newly and independently drawn parameter. The model specifies one common unknown $\theta$, so the joint likelihood for both exposures is averaged over that shared parameter.

Posterior Distribution and Predictive Distribution Are Different Objects

A posterior distribution assigns probabilities to parameter values after observing data. A posterior predictive distribution assigns probabilities to future outcomes after observing the same data. They generally have different possible values and different interpretations. In the one-exposure example, the parameter can be $0.02$ or $0.08$, while the future outcome is either no claim or a claim. The posterior probabilities of the parameter values are not themselves probabilities of those future outcomes.

A conditional forecast evaluated at one selected parameter value is also not automatically the posterior predictive distribution. Such a forecast may be useful for a specified scenario, but it does not average over the posterior uncertainty unless that posterior is concentrated at the selected value. When the question asks for a predictive probability, use the full posterior-weighted likelihood unless the problem explicitly conditions on a particular parameter value.

Key takeaway: To predict a future claim outcome, calculate its likelihood for each possible parameter value and average those likelihoods using posterior probabilities or posterior density. Keep the parameter posterior and the future-outcome distribution separate, and use a shared parameter in a joint forecast when the model specifies one.

Check Your Understanding

For each problem, average the conditional future-outcome likelihood over the stated posterior. Give probabilities to four decimal places when needed.

A future exposure has a Bernoulli claim indicator with unknown probability $\theta$. Given historical data, $P(\theta=0.04\mid D)=0.25$ and $P(\theta=0.12\mid D)=0.75$. Find the predictive probability of a claim. (Answer: 0.1000)
A future annual count is conditionally Poisson with rate $\lambda$. The posterior probabilities are $P(\lambda=0.5\mid D)=0.60$ and $P(\lambda=1.5\mid D)=0.40$. Find the predictive probability of zero claims. (Answer: 0.4532)
Two future exposures share a claim probability $\theta$, and are conditionally independent given $\theta$. If $P(\theta=0.05\mid D)=0.20$ and $P(\theta=0.25\mid D)=0.80$, find the predictive probability that neither exposure has a claim. (Answer: 0.6305)
A future count is conditionally Poisson. The posterior has $P(\lambda=1\mid D)=0.30$ and $P(\lambda=3\mid D)=0.70$. Find the predictive probability of exactly one claim. (Answer: 0.2148)
A future exposure has a Bernoulli claim indicator. The posterior probabilities are $P(\theta=0.20\mid D)=0.70$ and $P(\theta=0.50\mid D)=0.30$. Find the predictive probability of no claim. (Answer: 0.7100)

What You'll Learn