Gamma Priors for Exponential Claim Severity

Why Update a Severity Rate?

A hypothetical insurer may model individual claim amounts as exponentially distributed, while allowing the rate of that distribution to differ across portfolios or risk groups. A Gamma prior provides a convenient way to represent uncertainty about that rate. After observing complete claim amounts, the prior combines with their likelihood to give a Gamma posterior.

The update depends on two data summaries: the number of observed claims and their total amount. It is important to keep the measurement unit fixed throughout the calculation. If amounts are recorded in thousands of dollars, the rate is measured per thousand dollars; if amounts are in dollars, it is measured per dollar.

Model and assumptions: Conditional on an unknown rate $\lambda>0$, the complete claim amounts $X_1,\ldots,X_n$ are independent and identically distributed with exponential density $f_X(x\mid\lambda)=\lambda e^{-\lambda x}$, for $x>0$. Before observing the amounts, $\lambda$ has a Gamma prior with shape $a>0$ and rate $b>0$, written $\lambda\sim\operatorname{Gamma}(a,b)$. The observations are assumed to be fully observed, uncensored, and expressed in one consistent amount unit.

Here the Gamma distribution uses the shape-rate convention, not shape-scale. Its density is proportional to $\lambda^{a-1}e^{-b\lambda}$, and its mean and variance are $a/b$ and $a/b^2$. The rate $b$ has the same units as a claim amount, so that $b\lambda$ is dimensionless. The exponential rate $\lambda$ has units inverse to the claim amount.

Deriving the Posterior

Write the observed total claim amount as $T=\sum_{j=1}^{n}x_j$. Conditional on $\lambda$, the likelihood of the complete observations is

L(\lambda\mid x_1,\ldots,x_n) =\prod_{j=1}^{n}\lambda e^{-\lambda x_j} =\lambda^n e^{-\lambda T}.

Multiplying by the prior density shows why the posterior remains Gamma. The prior contributes a power of $\lambda$ and an exponential term; the likelihood adds $n$ to the power and $T$ to the exponential rate.

Write the prior kernel.
For $\lambda\sim\operatorname{Gamma}(a,b)$, the prior density as a function of $\lambda$ is proportional to $\lambda^{a-1}e^{-b\lambda}$.

Multiply by the likelihood.
The posterior kernel is proportional to $\lambda^{a-1}e^{-b\lambda}\lambda^n e^{-T\lambda}=\lambda^{a+n-1}e^{-(b+T)\lambda}$.

Identify and normalize the distribution.
This is the kernel of a Gamma density with shape $a+n$ and rate $b+T$. Its normalizing constant is $(b+T)^{a+n}/\Gamma(a+n)$.

Theorem: Under the stated model and with complete observations $x_1,\ldots,x_n$, the posterior distribution is

\lambda\mid x_1,\ldots,x_n \sim \operatorname{Gamma}\left(a+n,\ b+\sum_{j=1}^{n}x_j\right).

Thus the observed sample affects the posterior through $n$, the number of claims, and $T=\sum_j x_j$, the total claim amount. Given these summaries, the individual arrangement of the observations does not change this posterior update.

The shape is increased by the number of complete observations, while the rate is increased by the total observed amount. The two updates have distinct roles: $n$ counts how many exponential density contributions were observed, and $T$ records their total size. This is not an update based on the sample mean alone unless the sample size is also known, because $T=n\bar{x}$.

Worked Updates

Worked Example: Updating From Four Complete Amounts

A hypothetical portfolio records four complete claim amounts, in thousands of dollars: $1,2,3,$ and $4$. Assume conditional independence and an exponential severity distribution with unknown rate $\lambda$ per thousand dollars. Before seeing these amounts, let $\lambda\sim\operatorname{Gamma}(3,2)$ in the shape-rate convention. Find the posterior distribution, mean, and mode of $\lambda$.

Method and data summary. Use the complete-data Gamma-exponential update. The count is $n=4$, and the total is

$$ T=1+2+3+4=10 $$

in thousands of dollars. The data are treated as complete, not censored; no frequency model is being applied.

Posterior parameters. Add the count to the prior shape and the total to the prior rate:

a_{\text{post}}=3+4=7,\qquad b_{\text{post}}=2+10=12.

Therefore $\lambda\mid x_1,\ldots,x_4\sim\operatorname{Gamma}(7,12)$. Its posterior mean is $7/12=0.5833$ per thousand dollars. Since the shape is greater than 1, its mode is $(7-1)/12=0.5000$ per thousand dollars.

Check and trap avoided. The sample mean is $10/4=2.5$ thousand dollars, and the exponential maximum-likelihood rate is $4/10=0.4$ per thousand dollars. Neither number is the posterior shape or rate: the posterior parameters must be updated separately. In particular, adding the sample mean $2.5$ to the prior rate would be incorrect; the rate update uses the total $T=10$.

Worked Example: Keeping Dollar Units and Rate Units Consistent

Suppose a hypothetical insurer observes five complete claims with average amount $\$2{,}000$. Conditional on the common unknown rate, assume independent exponential claim amounts measured in dollars. The prior is $\lambda\sim\operatorname{Gamma}(4,8000)$, where the second parameter is the rate parameter in dollar units. Find the posterior and its mean and mode.

Method and total. Recover the total amount from the count and average:

n=5,\qquad T=5(2000)=\$10{,}000.

Update. The posterior shape is $4+5=9$, and the posterior rate is $8000+10{,}000=18{,}000$ dollars. Thus

\lambda\mid x_1,\ldots,x_5\sim\operatorname{Gamma}(9,18000).

The posterior mean rate is $9/18{,}000=0.0005$ per dollar. Because the shape is greater than 1, the mode is $8/18{,}000=0.0004444$ per dollar.

Check and trap avoided. The observed average is $\$2{,}000$, so the exponential maximum-likelihood rate is $1/2000=0.0005$ per dollar, matching the posterior mean rate in this particular example. The prior rate and data total are both dollar amounts, so adding them is dimensionally valid. Adding a total stated in thousands to a prior rate stated in dollars would not be valid until the units were converted.

Worked Example: Using the Posterior to Assess a Rate Threshold

A hypothetical specialty portfolio has complete claim amounts measured in thousands of dollars. Assume the amounts are conditionally independent and exponential given $\lambda$, with prior $\lambda\sim\operatorname{Gamma}(2,4)$. Six observed claims have average amount $3.5$ thousand dollars. Find the posterior and calculate $P(\lambda>0.4\mid\text{data})$, with $\lambda$ measured per thousand dollars.

Method and data summary. The count is $n=6$, so the total is $T=6(3.5)=21$ thousand dollars. The posterior parameters are

a_{\text{post}}=2+6=8,\qquad b_{\text{post}}=4+21=25.

Hence the posterior is $\operatorname{Gamma}(8,25)$. To calculate the threshold probability, scale the posterior rate variable: if $\lambda\sim\operatorname{Gamma}(8,25)$, then $25\lambda\sim\operatorname{Gamma}(8,1)$. The event $\lambda>0.4$ is equivalent to $25\lambda>10$.

For an integer shape of 8, the survival probability of a unit-rate Gamma variable at 10 is $e^{-10}\sum_{k=0}^{7}10^k/k!$. Substitution gives

P(\lambda>0.4\mid\text{data}) =e^{-10}\left(1+10+50+166.6667+416.6667+833.3333+1388.8889+1984.1270\right) \approx 0.2202.

Check and trap avoided. The requested event concerns the rate, not the mean severity $1/\lambda$. Also, the threshold must be multiplied by the posterior rate $25$ before using the unit-rate Gamma survival expression. The result, rounded to four decimal places, is $0.2202$.

Interpretation and Scope

The posterior describes uncertainty about the exponential severity rate after observing the complete claim amounts. For example, its mean is the posterior expectation of $\lambda$, whereas its mode is the most likely value of $\lambda$ under the posterior density. These summaries are not interchangeable, and neither should be confused with the reciprocal of a rate summary.

Although the focus here is the parameter update, the stated model and posterior also determine the distribution of a future claim amount under the same exponential severity assumption. Integrating over a $\operatorname{Gamma}(\alpha,\beta)$ posterior for $\lambda$ gives predictive density $f(x)=\alpha\beta^\alpha/(\beta+x)^{\alpha+1}$ for $x>0$. Thus the severity update does determine future severity prediction under this model. It does not model how many claims occur, or by itself determine the distribution of aggregate losses; those require a frequency model as well.

Common exam trap: In the shape-rate convention, the posterior parameters are $a+n$ and $b+T$. They are not $a+T$ and $b+n$. If a source instead uses a Gamma shape-scale convention, its second parameter has a different meaning and cannot be substituted into this update without conversion.

A final check is to verify the units and data interpretation before calculating. The observed amounts must be complete claim severities from the modeled population; a claim count alone is not severity data. If amounts are censored, truncated, or measured under differing scales, the likelihood is not simply $\lambda^n e^{-\lambda T}$, so this posterior formula does not apply as written.

Check Your Understanding

For each problem, use the shape-rate convention and assume complete, conditionally independent exponential claim amounts unless the problem says otherwise.

A prior is $\operatorname{Gamma}(2,3)$ for a rate measured per thousand dollars. Three complete claims total $6$ thousand dollars. Give the posterior parameters. (Answer: $\operatorname{Gamma}(5,9)$.)
A dollar-denominated rate has prior $\operatorname{Gamma}(5,10)$. Four complete claims have average amount $\$2{,}500$. Give the posterior distribution. (Answer: $\operatorname{Gamma}(9,10010)$, with rate parameter in dollars.)
For a rate per thousand dollars, the prior is $\operatorname{Gamma}(3,2)$. Two complete claims have average amount $4$ thousand dollars. Find the posterior mean rate. (Answer: Posterior $\operatorname{Gamma}(5,10)$; mean rate $5/10=0.5$ per thousand dollars.)
A dollar-denominated rate has prior $\operatorname{Gamma}(4,5000)$. Six complete claims total $\$24{,}000$. Find the posterior shape, rate, and mode. (Answer: Posterior $\operatorname{Gamma}(10,29000)$; mode $9/29000\approx0.0003103$ per dollar.)
Two complete datasets have the same total amount of $12$ thousand dollars, but one has three claims and the other has six. With prior $\operatorname{Gamma}(2,5)$, give both posteriors and explain why they differ. (Answer: The posteriors are $\operatorname{Gamma}(5,17)$ and $\operatorname{Gamma}(8,17)$, respectively; the total gives the same rate update, but the different claim counts give different shape updates.)

What You'll Learn