Predictive Severity Under Exponential-Gamma Conjugacy

Why Predictive Severity Is Not Just One Exponential Distribution

Suppose individual claim severities are conditionally exponential, but the exponential rate is not known. After observing severities, uncertainty about that rate remains. A prediction for one new severity must account for both the randomness of the claim and the remaining uncertainty about the rate. Integrating over the posterior distribution of the rate produces a predictive distribution that is generally not exponential.

This distinction matters when calculating expected future severity or the probability that a new claim exceeds a threshold. The posterior mean rate is not, in general, enough to describe the predictive severity distribution. The predictive tail is heavier than the tail of an exponential distribution with rate equal to the posterior mean.

Model and notation: Let $\lambda>0$ be an unknown exponential rate, with conditional severity density $f_{X\mid\lambda}(x\mid\lambda)=\lambda e^{-\lambda x}$ for $x\ge 0$. Use a gamma distribution in the shape-rate parameterization: $\lambda\sim\operatorname{Gamma}(a,b)$ has density $b^a\lambda^{a-1}e^{-b\lambda}/\Gamma(a)$, where $a>0$ and $b>0$. Here $b$ has units of dollars when $X$ is measured in dollars, while $\lambda$ has units of inverse dollars.

Assume the observed severities $x_1,\ldots,x_n$ are independent and identically distributed conditional on $\lambda$, and that they are fully observed, uncensored exponential severities. With a prior $\operatorname{Gamma}(a_0,b_0)$, the likelihood is proportional to $\lambda^n e^{-\lambda\sum_{j=1}^n x_j}$. Therefore the posterior is $\operatorname{Gamma}(a,b)$, where

a=a_0+n,\qquad b=b_0+\sum_{j=1}^{n}x_j.

In the rest of this tutorial, $a$ and $b$ refer to these posterior parameters. A future severity $Y$ is conditionally exponential given $\lambda$, with the same rate $\lambda$. The predictive distribution is the distribution of $Y$ after averaging over the posterior uncertainty in $\lambda$.

Deriving the Predictive Distribution

The predictive density is the conditional density of $Y$ averaged over the posterior density of $\lambda$. This averaging is sometimes called mixing. It is important to integrate over the full posterior, rather than substitute one estimate of $\lambda$ into the conditional density.

Write the conditional density and posterior density.
For $y\ge 0$, multiply $\lambda e^{-\lambda y}$ by the $\operatorname{Gamma}(a,b)$ posterior density.

Integrate out the rate.
Integrate that product over all $\lambda>0$, using the gamma integral $\int_0^\infty \lambda^a e^{-c\lambda}\,d\lambda=\Gamma(a+1)/c^{a+1}$ for $c>0$.

Read off the predictive probabilities.
Integrate the resulting density, or use its survival function, to obtain probabilities for thresholds and intervals.

In detail, the posterior predictive density is

\begin{aligned} f_Y(y\mid x_1,\ldots,x_n) &=\int_0^\infty \lambda e^{-\lambda y} \frac{b^a}{\Gamma(a)}\lambda^{a-1}e^{-b\lambda}\,d\lambda\\ &=\frac{b^a}{\Gamma(a)} \int_0^\infty \lambda^a e^{-(b+y)\lambda}\,d\lambda\\ &=\frac{b^a\Gamma(a+1)}{\Gamma(a)(b+y)^{a+1}}\\ &=\frac{a b^a}{(b+y)^{a+1}},\qquad y\ge 0. \end{aligned}

The last equality uses $\Gamma(a+1)=a\Gamma(a)$. Integrating this density from zero to $y$ gives the cumulative distribution function; integrating from $y$ to infinity gives the survival function.

F_Y(y)=1-\left(\frac{b}{b+y}\right)^a,\qquad S_Y(y)=P(Y>y)=\left(\frac{b}{b+y}\right)^a,\qquad y\ge 0.

Key result: The posterior predictive severity has a Lomax (Pareto type II) distribution with shape $a$ and scale $b$. Its support is $y\ge 0$. For $0\le c<d$, the interval probability is $P(c<Y\le d)=\left(\frac{b}{b+c}\right)^a-\left(\frac{b}{b+d}\right)^a$. The predictive distribution depends on the posterior through both its shape and its rate parameter, not just through the posterior mean of $\lambda$.

The power form of the survival function explains why the predictive tail is heavier than an exponential tail: $S_Y(y)$ decreases as a power of $b+y$, rather than exponentially in $y$. Consequently, moment existence must be checked rather than assumed.

Predictive Moments and Their Conditions

For a positive real number $k$, the $k$th predictive moment exists if and only if $k<a$. When it exists, it is

E[Y^k\mid x_1,\ldots,x_n] =b^k\frac{\Gamma(k+1)\Gamma(a-k)}{\Gamma(a)},\qquad k<a.

To justify both the expression and its condition, start from the predictive density and make the substitution $z=y/b$. The resulting integral is proportional to $\int_0^\infty z^k(1+z)^{-(a+1)}\,dz$. Its tail behaves like $z^{k-a-1}$; this integral converges exactly when $k-a-1<-1$, or $k<a$. When it converges, the beta integral gives the stated gamma-function expression. Thus the condition is part of the result: outside it, the corresponding moment is infinite.

In particular, the predictive mean exists only when $a>1$, and the variance exists only when $a>2$. For these cases,

E[Y\mid x_1,\ldots,x_n]=\frac{b}{a-1}\quad(a>1), \qquad \operatorname{Var}(Y\mid x_1,\ldots,x_n) =\frac{a b^2}{(a-1)^2(a-2)}\quad(a>2).

The mean also has a direct mixture interpretation. Conditional on the rate, $E[Y\mid\lambda]=1/\lambda$, so the predictive mean is the posterior expectation of $1/\lambda$. For a $\operatorname{Gamma}(a,b)$ distribution, that expectation is $b/(a-1)$ only when $a>1$. It is not $1/E[\lambda]$, which equals $b/a$. When $a\le 1$, $E[\lambda]=a/b$ is still finite, but the predictive severity mean is infinite.

Worked Example: Updating the Rate and Calculating a Tail Probability

A hypothetical insurer models individual claim severities in dollars as exponential with unknown rate $\lambda$. The prior is $\operatorname{Gamma}(2,1000)$ in the shape-rate convention. Two fully observed severities are \$1,000 and \$2,000. Find the predictive mean for one new severity and the probability it exceeds \$2,000. Give the probability to four decimal places and the mean to the nearest cent.

Method and posterior update. The sample size is $n=2$, and the severity sum is $1{,}000+2{,}000=3{,}000$ dollars. Therefore,

a=2+2=4,\qquad b=1{,}000+3{,}000=4{,}000.

The predictive model is Lomax with shape $4$ and scale $4{,}000$ dollars. Since $a=4>1$, the mean exists:

E[Y]=\frac{4{,}000}{4-1} =\frac{4{,}000}{3} =\$1{,}333.33.

For the tail probability, use the predictive survival function:

P(Y>2{,}000) =\left(\frac{4{,}000}{4{,}000+2{,}000}\right)^4 =\left(\frac{2}{3}\right)^4 =\frac{16}{81} =0.1975.

Check. The probability is between zero and one, and $P(Y\le 2{,}000)=1-0.1975=0.8025$, to four decimal places. Common mistake avoided: substituting the posterior mean rate $a/b=4/4{,}000$ into an exponential model would give tail probability $e^{-2}$, not the posterior predictive tail. That substitution ignores the remaining uncertainty in $\lambda$.

Using the Predictive Distribution in Exam Calculations

For a threshold question, first identify whether the event is a survival event, a cumulative event, or an interval event. Then use the matching expression. A deductible-like threshold in a question about whether a severity exceeds a specified amount calls for $S_Y(y)$; it does not change the underlying severity model unless the question explicitly defines a modified payment.

Worked Example: Probability in a Severity Band

For a hypothetical portfolio, the posterior rate for exponential severity is $\operatorname{Gamma}(3,2{,}000)$, with rate measured per dollar. A new claim severity $Y$ has the posterior predictive distribution under this model. Calculate $P(2{,}000<Y\le 6{,}000)$, to four decimal places.

Method. The posterior shape is $a=3$, and the rate parameter is $b=2{,}000$ dollars. Subtract the survival probability at the upper endpoint from the survival probability at the lower endpoint:

\begin{aligned} P(2{,}000<Y\le6{,}000) &=S_Y(2{,}000)-S_Y(6{,}000)\\ &=\left(\frac{2{,}000}{2{,}000+2{,}000}\right)^3 -\left(\frac{2{,}000}{2{,}000+6{,}000}\right)^3\\ &=\left(\frac12\right)^3-\left(\frac14\right)^3\\ &=\frac18-\frac1{64} =\frac7{64} =0.1094. \end{aligned}

Check. The interval probability is positive and is smaller than $S_Y(2{,}000)=0.1250$, as it must be because the interval excludes outcomes above \$6,000. Common mistake avoided: using $S_Y(6{,}000)-S_Y(2{,}000)$ reverses the subtraction and produces an impossible negative probability.

Worked Example: A Finite Mean but Infinite Variance

A hypothetical severity model begins with a $\operatorname{Gamma}(0.5,1000)$ prior for the exponential rate. One fully observed severity is \$2,000. Find the predictive mean and $P(Y>3{,}000)$. Also determine whether the predictive variance is finite. Report the probability to four decimal places and the mean to the nearest cent.

Method and posterior update. There is one observation, so the posterior shape and rate are

a=0.5+1=1.5,\qquad b=1{,}000+2{,}000=3{,}000.

Since $1.5>1$, the predictive mean exists:

E[Y]=\frac{3{,}000}{1.5-1} =\frac{3{,}000}{0.5} =\$6{,}000.00.

The tail probability is

P(Y>3{,}000) =\left(\frac{3{,}000}{3{,}000+3{,}000}\right)^{1.5} =(0.5)^{1.5} =\frac{1}{2\sqrt{2}} =0.3536.

The variance is not finite, because its existence requires $a>2$, whereas $a=1.5$. Check. The mean exists because the first-moment threshold is $a>1$, while the second-moment threshold is stricter. Common mistake avoided: assuming that a finite predictive mean implies a finite predictive variance.

Common Traps and Practical Checks

Confusing shape-rate with shape-scale. Here $b$ is the gamma rate parameter, so the prior density contains $e^{-b\lambda}$, and the posterior update is $b=b_0+\sum x_j$. Using a shape-scale convention without changing the formulas leads to incorrect updates and units.
Using the posterior mean rate as if the rate were known. The predictive density averages over the entire posterior. In general, $E[1/\lambda\mid\text{data}]\ne 1/E[\lambda\mid\text{data}]$, and the predictive tail is not an exponential tail at the posterior mean rate.
Reporting a moment without checking its condition. The predictive mean is infinite for $a\le1$; the variance is infinite for $a\le2$. A formula with a zero or negative denominator is a warning that the moment formula does not apply, not a usable numerical answer.
Mixing dollar and inverse-dollar units. The posterior rate parameter $b$ has the same units as severity. A threshold $y$ in the survival formula must use the same currency unit as $b$.
Applying the model outside its assumptions. The update above assumes fully observed individual severities that are conditionally independent exponential observations. Truncation, censoring, or a different conditional severity model requires a different likelihood and may produce a different posterior.

Key takeaway: With a $\operatorname{Gamma}(a,b)$ posterior for an exponential rate, integrate over $\lambda$ to obtain a Lomax predictive severity: $S_Y(y)=(b/(b+y))^a$. Its mean is $b/(a-1)$ only for $a>1$, and its variance is finite only for $a>2$. This distinction is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, use the gamma shape-rate convention and the posterior predictive distribution derived above.

A hypothetical model has prior rate distribution $\operatorname{Gamma}(2,500)$. Two fully observed severities are \$500 and \$1,500. Find the predictive mean and the probability that a new severity exceeds \$2,500. (Answer: posterior $\operatorname{Gamma}(4,2{,}500)$; mean \$833.33; probability $0.0625$.)
A hypothetical model has posterior rate distribution $\operatorname{Gamma}(4,3{,}000)$. What is the probability that a new severity is at most \$3,000? (Answer: $0.9375$.)
A new severity has predictive parameters $a=2.5$ and $b=\$5{,}000$. Find its mean and its probability of exceeding \$5,000. (Answer: mean \$3,333.33; probability $0.1768$.)
A posterior rate distribution is $\operatorname{Gamma}(1.5,2{,}000)$. Determine the predictive mean and whether the predictive variance is finite. (Answer: mean \$4,000; variance is infinite.)
A new severity has predictive parameters $a=4$ and $b=\$1{,}000$. Find the probability that it is greater than \$1,000 but no more than \$3,000. (Answer: $0.0586$.)

What You'll Learn