Posterior Predictive Intervals for Future Losses

Why Predictive Intervals Are Different

A posterior estimate can summarize what is currently known about a risk, but it does not describe the full range of outcomes that might occur next. A future claim depends both on the unknown risk parameter and on random variation even if that parameter were known. A posterior predictive interval accounts for both sources of uncertainty.

This distinction matters when the decision concerns a future claim count, a future severity, or another future observation. An interval for an unknown parameter answers a different question: where the parameter may lie given the observed data. The posterior predictive interval instead describes where a future observation may lie, conditional on the observed data and the model. These methods are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Let $\theta$ be an unknown model parameter, $D$ the observed data, and $Y_{\mathrm{new}}$ a future observation. The posterior predictive distribution is the conditional distribution of $Y_{\mathrm{new}}$ given $D$, obtained by averaging the future-observation distribution over the posterior distribution of $\theta$: $$ P(Y_{\mathrm{new}}\in A\mid D) = \int P(Y_{\mathrm{new}}\in A\mid \theta,D)\,\pi(\theta\mid D)\,d\theta. $$ An equal-tail $100(1-\alpha)\%$ posterior predictive interval has endpoints $q_{\alpha/2}$ and $q_{1-\alpha/2}$, the corresponding quantiles of this predictive distribution. For a discrete outcome, define $q_p=\inf\{y:F(y\mid D)\ge p\}$; the interval's actual probability can exceed its nominal level because the distribution has jumps.

The data $D$ update uncertainty about $\theta$, giving the posterior distribution $\pi(\theta\mid D)$. The predictive calculation then averages over that posterior. Even after the data are observed, the future outcome is not generally distributed as though $\theta$ were known to equal a single estimate.

Why Parameter Uncertainty Must Be Included

The law of total variance shows how the two sources of predictive uncertainty combine. Conditional on the data, first consider the future outcome's variance if $\theta$ were known; then add the variation in its conditional mean as $\theta$ varies over its posterior.

\operatorname{Var}(Y_{\mathrm{new}}\mid D) = E[\operatorname{Var}(Y_{\mathrm{new}}\mid \theta,D)\mid D] + \operatorname{Var}(E[Y_{\mathrm{new}}\mid \theta,D]\mid D).

The first term is expected process variance. The second is variance in the conditional mean caused by parameter uncertainty. A plug-in calculation that replaces $\theta$ by its posterior mean usually omits the second term. It may therefore produce an interval that is too narrow.

Key distinction: A posterior credible interval for $\theta$ summarizes uncertainty about the parameter. A posterior predictive interval for $Y_{\mathrm{new}}$ summarizes uncertainty about a future outcome. The latter includes both posterior parameter uncertainty and future process variation; it is not obtained merely by relabeling a parameter interval.

Worked Example: Predicting a Future Claim Count

Worked Example: A Gamma-Poisson Predictive Interval

A hypothetical portfolio has an unknown claim rate $\lambda$ per unit of exposure. Assume the prior is $\lambda\sim\operatorname{Gamma}(2,2)$, using the shape-rate parameterization, and that counts conditional on $\lambda$ are Poisson. The portfolio records 3 claims over exposure 2. Find an equal-tail 80% posterior predictive interval for the count $K$ in a future exposure of 1 unit, and state the interval's actual predictive probability.

Method and posterior. For a Gamma(shape $\alpha$, rate $\beta$) prior and a Poisson count over exposure $e$, observing $n$ claims gives a Gamma posterior with shape $\alpha+n$ and rate $\beta+e$. Thus

\lambda\mid D\sim\operatorname{Gamma}(2+3,2+2) =\operatorname{Gamma}(5,4).

For future exposure $h=1$, the conditional model is $K\mid\lambda\sim\operatorname{Poisson}(\lambda)$. Integrating its probability mass function over the posterior gives

P(K=k\mid D) = \frac{\Gamma(5+k)}{\Gamma(5)k!} \left(\frac{4}{5}\right)^5 \left(\frac{1}{5}\right)^k,\qquad k=0,1,2,\ldots.

Find the quantiles. The first probabilities are

\begin{aligned} P(K=0\mid D)&=(4/5)^5=0.32768,\\ P(K=1\mid D)&=0.32768(5)(1/5)=0.32768,\\ P(K=2\mid D)&=P(K=1\mid D)(6/2)(1/5)=0.196608,\\ P(K=3\mid D)&=P(K=2\mid D)(7/3)(1/5)=0.0917504. \end{aligned}

Therefore $F(0\mid D)=0.32768$, $F(1\mid D)=0.65536$, $F(2\mid D)=0.851968$, and $F(3\mid D)=0.9437184$. The 10th percentile is 0, and the 90th percentile is 3. The equal-tail 80% interval is $[0,3]$. Its actual predictive probability is $F(3\mid D)=0.9437184$, or 94.37184%, since no count can fall below zero and the upper endpoint must include the full probability at 3.

Check and common mistake. The posterior mean rate is $5/4=1.25$, so the predictive mean count is 1.25. The predictive variance is $1.25+5/16=1.5625$, where $5/16$ is the extra variation from uncertainty about $\lambda$. Using only a Poisson distribution with rate 1.25 would give variance 1.25 and ignore that extra uncertainty. That plug-in model is not the posterior predictive distribution.

Deriving the Gamma-Poisson Predictive Distribution

The preceding count formula follows directly from the conditional Poisson model and the Gamma posterior. Write the posterior as $\operatorname{Gamma}(r,b)$, with density $b^r\lambda^{r-1}e^{-b\lambda}/\Gamma(r)$. For future exposure $h$, conditionally $K\mid\lambda\sim\operatorname{Poisson}(h\lambda)$. Averaging the Poisson mass over the posterior gives

\begin{aligned} P(K=k\mid D) &=\int_0^\infty \frac{e^{-h\lambda}(h\lambda)^k}{k!} \frac{b^r\lambda^{r-1}e^{-b\lambda}}{\Gamma(r)} \,d\lambda\\ &=\frac{b^r h^k}{k!\Gamma(r)} \int_0^\infty \lambda^{r+k-1}e^{-(b+h)\lambda}\,d\lambda\\ &=\frac{\Gamma(r+k)}{\Gamma(r)k!} \left(\frac{b}{b+h}\right)^r \left(\frac{h}{b+h}\right)^k. \end{aligned}

The last step uses the Gamma integral: for $c>0$, $\int_0^\infty \lambda^{a-1}e^{-c\lambda}\,d\lambda=\Gamma(a)/c^a$. This is a negative binomial predictive distribution. Its mean and variance are

E[K\mid D]=\frac{rh}{b}, \qquad \operatorname{Var}(K\mid D)=\frac{rh}{b}+\frac{rh^2}{b^2}.

The first variance term is the expected Poisson process variance. The second is the contribution from uncertainty in the rate. The derivation assumes a Gamma posterior for a Poisson rate and a future count conditionally Poisson over exposure $h$. It does not apply unchanged if the count model or posterior distribution has a different form.

Worked Example: A Normal-Normal Future Severity

Worked Example: Predicting One Future Claim Cost

For a hypothetical coverage, claim costs $X_i$ are conditionally independent $N(\theta,100)$ amounts in dollars, with known variance $100$. The unknown mean has prior $\theta\sim N(100,100)$. Four observed claim costs have sample mean $\bar{x}=110$. Find a 95% posterior predictive interval for one future claim cost $X_{\mathrm{new}}$.

Method and posterior. Under the Normal-Normal model, posterior precision is prior precision plus data precision. Hence

\begin{aligned} v_n&=\left(\frac{1}{100}+\frac{4}{100}\right)^{-1}=20,\\ m_n&=v_n\left(\frac{100}{100}+\frac{4(110)}{100}\right) =20(0.5+4.4)=98. \end{aligned}

Correction: the prior contribution in the weighted mean is $100/100=1$, not $0.5$. Therefore the posterior mean is

m_n=20\left(1+\frac{440}{100}\right)=20(5.4)=108, \qquad \theta\mid D\sim N(108,20).

Conditional on $\theta$, the future cost has mean $\theta$ and variance 100. Averaging over the posterior gives $X_{\mathrm{new}}\mid D\sim N(108,100+20)=N(108,120)$. Using the 97.5th standard normal percentile $1.96$, the 95% interval is

108\pm1.96\sqrt{120} =108\pm21.4707 =[86.5293,129.4707].

Rounded to the nearest cent, the interval is $[$86.53,$129.47]$. The calculation adds the future claim variance to posterior uncertainty about the mean. An interval for $\theta$ alone would instead use variance 20 and would not describe the spread of an individual future claim. The initial precision expression also illustrates a common arithmetic trap: the prior mean contributes $100/100=1$ to the weighted sum.

Worked Example: A Predictive Interval for Exponential Severity

Worked Example: Updating an Unknown Exponential Rate

Assume hypothetical positive claim severities are conditionally exponential with rate $\lambda$, so $f_X(x\mid\lambda)=\lambda e^{-\lambda x}$ for $x>0$. Let the prior be $\lambda\sim\operatorname{Gamma}(2,1000)$, with rate measured per dollar. Three observed severities are $100, $200, and $300. Find the central 90% posterior predictive interval for one future severity.

Update the rate distribution. The observed total is $100+200+300=600$, and there are three observations. The Gamma posterior is

\lambda\mid D\sim\operatorname{Gamma}(2+3,1000+600) =\operatorname{Gamma}(5,1600).

For a future severity $X_{\mathrm{new}}$, its conditional survival function is $e^{-\lambda x}$. Integrating over the posterior gives

\begin{aligned} S_{X_{\mathrm{new}}\mid D}(x) &=E[e^{-\lambda x}\mid D]\\ &=\left(\frac{1600}{1600+x}\right)^5,\qquad x\ge 0. \end{aligned}

Invert the predictive distribution. For quantile level $q$, set the predictive cumulative probability equal to $q$:

1-\left(\frac{1600}{1600+x_q}\right)^5=q, \qquad x_q=1600\left((1-q)^{-1/5}-1\right).

For the lower endpoint, $q=0.05$:

x_{0.05}=1600(0.95^{-1/5}-1)\approx \$16.50.

For the upper endpoint, $q=0.95$:

x_{0.95}=1600(0.05^{-1/5}-1)\approx \$1{,}312.90.

Thus the central 90% posterior predictive interval is [$16.50,$1,312.90]. The interval is wide and asymmetric because the predictive distribution is right-skewed. It is a range for a future severity, not a range for the unknown rate $\lambda$. This calculation assumes independent exponential severities conditional on a single Gamma-distributed rate; a different severity likelihood would require a different predictive distribution.

Interpreting and Reporting Predictive Intervals

A predictive interval should identify both the future quantity and the conditioning information. For example, “the future count over one exposure unit, given the observed data and the Gamma-Poisson model” is more precise than “the 80% interval.” State whether an interval is central equal-tail or uses another rule. For discrete counts, report the endpoints and, when useful, the actual probability between them; the nominal level need not be attained exactly.

Do not confuse a posterior predictive interval with a confidence interval for a parameter. Nor should a predictive interval be interpreted as a guarantee that a specified proportion of future individual outcomes will fall inside it under every possible model. Its probability statement is conditional on the stated model and observed data. Model assumptions, including the choice of prior and likelihood, remain important.

Key takeaway: To predict a future claim, average its conditional distribution over the posterior distribution of the unknown parameter, then find quantiles of that predictive distribution. This includes both process variation and parameter uncertainty.

Check Your Understanding

For each problem, identify the requested predictive quantity before calculating.

A claim count has a Poisson likelihood with unknown rate $\lambda$ per exposure unit. The prior is $\operatorname{Gamma}(1,3)$, shape-rate; 2 claims are observed over exposure 1. What is the posterior predictive probability of zero claims over a future exposure of 1? (Answer: 0.5120)
Claim costs are conditionally $N(\theta,64)$. The prior is $\theta\sim N(50,100)$, and 4 costs have mean 56. Give the posterior predictive mean and variance for one future cost. (Answer: Mean $55.1724$; variance $77.7931$)
Severities are conditionally exponential with unknown rate $\lambda$, with prior $\operatorname{Gamma}(2,500)$, shape-rate. Two observed severities are $100 and $300. Find the posterior predictive probability that the next severity exceeds $900. (Answer: 0.0625)
For the Normal-Normal model in question 2, is an interval based only on the posterior variance of $\theta$ a predictive interval for one future claim cost? State the missing variance component. (Answer: No; it omits the future process variance $64$)
A discrete posterior predictive distribution has $F(2\mid D)=0.84$ and $F(3\mid D)=0.93$. What is its 90th percentile under the generalized-inverse definition, and why might the probability of an interval ending there exceed its nominal level? (Answer: 3; discrete probability masses make the cumulative distribution jump past 0.90)

What You'll Learn