Posterior Predictive Means for Gamma Severities

Why the Posterior Predictive Mean Needs Care

When claim severities are modeled with a gamma distribution, the expected severity depends on its rate parameter. If that rate is uncertain, a future claim's expected severity must average over the rate's posterior distribution. Simply substituting the posterior mean rate into the severity mean generally gives a different answer. The distinction matters because the predictive mean uses the posterior expectation of the reciprocal rate.

This tutorial considers independent, complete observations from a gamma severity model with a known shape and an unknown rate. It derives the posterior predictive mean, states exactly when it is finite, and shows how to avoid confusing a gamma rate with a gamma scale. The results are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Definition: Use the rate parameterization $X_i\mid\lambda\sim\operatorname{Gamma}(k,\text{ rate }\lambda)$, where $k>0$ is the known shape and $\lambda>0$ is the unknown rate. Its density is $f_{X_i\mid\lambda}(x)=\lambda^k x^{k-1}e^{-\lambda x}/\Gamma(k)$ for $x>0$. Thus $E[X_i\mid\lambda]=k/\lambda$. Assign the rate a gamma prior $\lambda\sim\operatorname{Gamma}(\alpha,\text{ rate }\beta)$, with $\alpha>0,\beta>0$, and density $\beta^\alpha\lambda^{\alpha-1}e^{-\beta\lambda}/\Gamma(\alpha)$. Here $\beta$ has units of severity, while $\lambda$ has units of inverse severity.

Posterior Updating and the Predictive Mean

Suppose $n$ independent severities $x_1,\ldots,x_n$ are observed under the same $\lambda$, with $n\geq 0$. Define their total as $T=\sum_{i=1}^n x_i$, taking $T=0$ if no observations have been made. The likelihood terms that depend on $\lambda$ are $\lambda^{nk}e^{-\lambda T}$. Multiplying by the gamma prior therefore gives a gamma posterior.

\lambda\mid x_1,\ldots,x_n \sim \operatorname{Gamma}(\alpha+nk,\text{ rate }\beta+T).

Write $a=\alpha+nk$ and $b=\beta+T$. A future severity $X_{\mathrm{new}}$ is conditionally gamma with shape $k$ and rate $\lambda$, using the same parameter value that generated the observed severities. By the law of iterated expectation, $E[X_{\mathrm{new}}\mid x_1,\ldots,x_n]=E[E[X_{\mathrm{new}}\mid\lambda]\mid x_1,\ldots,x_n]=kE[1/\lambda\mid x_1,\ldots,x_n]$. For a gamma-distributed rate, this reciprocal expectation exists only if its shape exceeds one.

Theorem: With $a=\alpha+nk$ and $b=\beta+\sum_{i=1}^n x_i$, the posterior predictive mean of one future severity is $E[X_{\mathrm{new}}\mid x_1,\ldots,x_n]=kb/(a-1)$ when $a>1$. If $a\leq1$, this mean is infinite. The posterior predictive distribution still exists for $a\leq1$; it is its mean, not the distribution itself, that fails to be finite.

Why the Reciprocal-Rate Condition Appears

The key step is evaluating the posterior expectation of $1/\lambda$. If the posterior density is gamma with shape $a$ and rate $b$, then $E[1/\lambda\mid x_1,\ldots,x_n]$ is an integral over positive rates. Substitution of the gamma density reduces it to a gamma integral with shape $a-1$.

\begin{aligned} E\!\left[\frac{1}{\lambda}\mid x_1,\ldots,x_n\right] &=\int_0^\infty \frac{1}{\lambda} \frac{b^a}{\Gamma(a)}\lambda^{a-1}e^{-b\lambda}\,d\lambda\\ &=\frac{b^a}{\Gamma(a)} \int_0^\infty \lambda^{a-2}e^{-b\lambda}\,d\lambda\\ &=\frac{b^a}{\Gamma(a)} \frac{\Gamma(a-1)}{b^{a-1}}\\ &=\frac{b}{a-1},\qquad a>1. \end{aligned}

The integral in the second line behaves near zero like $\lambda^{a-2}$. An integral of this power from zero is finite only when $a-2>-1$, or $a>1$. At $a=1$, it diverges logarithmically; for $a<1$, it also diverges. Since $k>0$, multiplying the reciprocal expectation by $k$ preserves this finite-versus-infinite distinction.

This also explains why the answer is not $k/E[\lambda\mid x_1,\ldots,x_n]$. In general, the expectation of a reciprocal is not the reciprocal of an expectation. The posterior mean rate is $a/b$, but the relevant quantity for severity is $E[1/\lambda\mid x_1,\ldots,x_n]=b/(a-1)$, when it exists.

Worked Calculations

Worked Example: Updating from Two Observed Severities

A hypothetical portfolio has severities $X_i\mid\lambda\sim\operatorname{Gamma}(2,\text{ rate }\lambda)$. The prior for the rate is $\operatorname{Gamma}(3,\text{ rate }300)$. Two complete severities, $\$100$ and $\$200$, are observed. Find the posterior predictive mean of one future severity, and compare it with the plug-in calculation using the posterior mean rate.

Method. Update the gamma rate prior using $n=2$, known shape $k=2$, and observed total severity $T=100+200=\$300$.

The posterior shape is $a=\alpha+nk=3+(2)(2)=7$. The posterior rate parameter is $b=\beta+T=300+300=\$600$. Thus $\lambda\mid\text{data}\sim\operatorname{Gamma}(7,\text{ rate }600)$.

Because $a=7>1$, the predictive mean is finite: $E[X_{\mathrm{new}}\mid\text{data}]=kb/(a-1)=(2)(600)/(7-1)=1200/6=\$200.00$.

For comparison, the posterior mean rate is $E[\lambda\mid\text{data}]=a/b=7/600$ per dollar. Substituting that rate into the conditional severity mean gives $2/(7/600)=1200/7=\$171.43$, which is not the posterior predictive mean. Check: the observed total is $\$300$, and the posterior rate parameter is $\$300+\$300=\$600$; both updates use the observed severities in dollars.

Common mistake avoided: using $k/E[\lambda\mid\text{data}]$ instead of $kE[1/\lambda\mid\text{data}]$. The former is a plug-in estimate, not the predictive expectation.

Worked Example: Data Make an Initially Infinite Mean Finite

A hypothetical severity model has known shape $k=1.5$ and rate prior $\lambda\sim\operatorname{Gamma}(0.8,\text{ rate }100)$. One complete severity of $\$50$ is observed. Determine whether the prior predictive mean and posterior predictive mean are finite, and calculate the latter.

Method. Check the gamma shape governing the reciprocal-rate expectation before and after observing the severity.

Before data, the rate-prior shape is $\alpha=0.8\leq1$. Therefore $E[1/\lambda]$ under the prior is infinite, so the prior predictive severity mean is infinite. This conclusion does not mean that a severity must be infinite; it means the distribution's expected value does not converge.

After one observation, the posterior shape is $a=0.8+(1)(1.5)=2.3$. The posterior rate is $b=100+50=\$150$. Since $a=2.3>1$, the posterior predictive mean is finite:

E[X_{\mathrm{new}}\mid X_1=50] =\frac{(1.5)(150)}{2.3-1} =\frac{225}{1.3} =\$173.08.

Check: the denominator is $1.3$, not $2.3$, because the reciprocal expectation uses $a-1$. The data increase the rate posterior shape enough to make its reciprocal expectation finite.

Common mistake avoided: assuming the prior's infinite predictive mean guarantees an infinite posterior predictive mean. The relevant condition must be checked using the posterior shape after updating.

Worked Example: Using the Total Severity as the Data Summary

A hypothetical policy group has severity shape $k=2.5$, known across the group, and prior rate distribution $\operatorname{Gamma}(2.5,\text{ rate }160)$. Three complete severities are $\$40,\$80,\$120$. Find the posterior predictive mean for one further severity and state which data summary enters the rate update.

Method. Apply conjugate updating; with fixed shape and common rate, the likelihood contribution to the rate depends on the number of observations and their total.

The sample size is $n=3$, and the total severity is $T=40+80+120=\$240$. The posterior shape is $a=2.5+(3)(2.5)=10$. The posterior rate parameter is $b=160+240=\$400$. Consequently, $\lambda\mid\text{data}\sim\operatorname{Gamma}(10,\text{ rate }400)$, and the finite predictive mean is

E[X_{\mathrm{new}}\mid\text{data}] =\frac{(2.5)(400)}{10-1} =\frac{1000}{9} =\$111.11.

The posterior mean rate is $10/400=0.025$ per dollar. The plug-in severity mean would be $2.5/0.025=\$100.00$, whereas the predictive mean is $\$111.11$. Check: the posterior rate parameter is $160+240=400$, and the posterior shape exceeds one, so the predictive mean is finite.

Common mistake avoided: replacing the total severity $T$ with the sample mean in the posterior rate update. The likelihood contributes $\exp(-\lambda T)$, so the rate parameter update adds the total, not the average.

Assumptions, Interpretation, and Exam Traps

The update and predictive mean require the observations to be conditionally independent given a common rate $\lambda$, with the same known severity shape $k$. The observations must be complete severity values from the stated model. A different shape for each observation, a different rate for each observation, truncation, censoring, or a changing claim environment requires a likelihood appropriate to that situation; the update above should not be applied unchanged.

The posterior predictive mean is an average over parameter uncertainty and future severity variability. It is the expected value of a new single severity conditional on the data, not an estimate of the realized next claim and not a frequency or aggregate-loss calculation. The model's rate $\lambda$ is not the prior rate parameter $\beta$: $\lambda$ is random, while $\beta$ is a fixed parameter of its prior distribution.

Identify the parameterization.
Confirm that the severity distribution uses rate $\lambda$, so its conditional mean is $k/\lambda$, and that the prior uses gamma shape $\alpha$ and rate $\beta$.

Summarize the observed severities.
Record the number $n$ and total $T=\sum x_i$. Keep the total in the same monetary units as $\beta$.

Update both gamma parameters.
Use posterior shape $a=\alpha+nk$ and rate $b=\beta+T$.

Check finiteness before calculating.
If $a\leq1$, the posterior predictive mean is infinite. If $a>1$, calculate $kb/(a-1)$.

A frequent parameterization trap is to use the scale-prior formula while the problem specifies a rate prior, or to treat $\beta$ as the mean of $\lambda$. Under the conventions here, $E[\lambda]=\alpha/\beta$, whereas the severity conditional mean is $k/\lambda$. Another trap is to report a finite numerical result without checking $a>1$. The posterior distribution is valid for every $a>0$, but its reciprocal-rate expectation is not finite for all such shapes.

Key takeaway: For a known gamma severity shape $k$, a gamma prior on the common severity rate updates to shape $\alpha+nk$ and rate $\beta+\sum x_i$. The posterior predictive mean of one new severity is $k(\beta+\sum x_i)/(\alpha+nk-1)$ only when $\alpha+nk>1$; otherwise it is infinite.

Check Your Understanding

For each problem, use the gamma rate parameterization stated above. Report monetary answers to the nearest cent.

A severity model has known shape $k=2$, with prior rate $\operatorname{Gamma}(4,\text{ rate }200)$. Two complete severities total $\$300$. Find the posterior predictive mean of one new severity. (Answer: $\$142.86$)
Let $k=0.75$, and let the rate prior be $\operatorname{Gamma}(1.4,\text{ rate }90)$. Two observations total $\$50$. Is the posterior predictive mean finite, and if so, what is it? (Answer: Finite; $\$55.26$)
A severity has known shape $k=1.2$, with rate prior $\operatorname{Gamma}(1.2,\text{ rate }75)$. One complete severity of $\$20$ is observed. Find the posterior predictive mean. (Answer: $\$81.43$)
A hypothetical portfolio has $k=3$, rate prior $\operatorname{Gamma}(0.5,\text{ rate }40)$, and one observed severity of $\$30$. Is the posterior predictive mean finite? Calculate it if it is. (Answer: Finite; $\$84.00$)
For known shape $k=1.5$, the rate prior is $\operatorname{Gamma}(2,\text{ rate }120)$, and observed severities are $\$100$ and $\$140$. Find the posterior rate distribution and the predictive mean for one new severity. (Answer: $\operatorname{Gamma}(5,\text{ rate }360)$; $\$135.00$)

What You'll Learn