Exponential Severity Estimation By Mean

Why Estimate Exponential Severity by Its Mean?

An exponential severity model can be written in terms of either a rate or a mean. The mean parameter is often convenient in actuarial work because it is directly interpretable as the model's expected loss. The rate parameter is its reciprocal, so the two parameterizations describe the same family of distributions, but their estimates must be translated carefully.

This tutorial derives the maximum likelihood estimate (MLE) of the mean from independent, complete severity observations. “Complete” means each loss amount is observed in full: there is no censoring, truncation, or deductible-based restriction on which losses enter the data. Those features change the likelihood and generally require a different estimation calculation.

Definition: An exponential severity $X$ with mean parameter $\theta>0$ has density $$ f_X(x;\theta)=\frac{1}{\theta}e^{-x/\theta}, \qquad x>0, $$ and density zero otherwise. Its expected value is $E[X]=\theta$. The equivalent rate parameter is $\lambda=1/\theta$, giving density $f_X(x;\lambda)=\lambda e^{-\lambda x}$ for $x>0$.

Likelihood and Mean Estimate

Let $X_1,\ldots,X_n$ be independent, identically distributed complete observations from the exponential model with mean $\theta$. Independence means the joint density is the product of the individual densities. For observed positive values $x_1,\ldots,x_n$, the likelihood as a function of $\theta$ is

L(\theta)=\prod_{j=1}^{n}\frac{1}{\theta}e^{-x_j/\theta} =\theta^{-n}\exp\left(-\frac{\sum_{j=1}^{n}x_j}{\theta}\right), \qquad \theta>0.

The log-likelihood is easier to maximize because the logarithm is strictly increasing: maximizing $L(\theta)$ is equivalent to maximizing its logarithm. Writing $\bar{x}=(1/n)\sum_{j=1}^{n}x_j$, differentiate the log-likelihood with respect to $\theta$.

\ell(\theta)=\log L(\theta) =-n\log\theta-\frac{\sum_{j=1}^{n}x_j}{\theta}, \qquad \ell'(\theta)=-\frac{n}{\theta}+\frac{\sum_{j=1}^{n}x_j}{\theta^2}.

Setting the score, the first derivative of the log-likelihood, equal to zero gives $-n/\theta+\sum x_j/\theta^2=0$. Since $\theta>0$, multiplication by $\theta^2$ gives $-n\theta+\sum x_j=0$, and hence $\theta=\bar{x}$. The second derivative is $\ell''(\theta)=n/\theta^2-2\sum x_j/\theta^3$. At $\theta=\bar{x}$, this is $-n/\bar{x}^{\,2}<0$, confirming a local maximum. In fact, the score is positive when $\theta<\bar{x}$ and negative when $\theta>\bar{x}$, so the log-likelihood increases up to $\bar{x}$ and decreases after it. This makes $\bar{x}$ the unique global maximum over $\theta>0$.

Theorem: For independent, identically distributed, complete exponential severity observations $x_1,\ldots,x_n>0$, the maximum likelihood estimate of the mean parameter is $$ \widehat{\theta}_{\mathrm{MLE}}=\bar{x}=\frac{1}{n}\sum_{j=1}^{n}x_j. $$ This result assumes the sample consists of full, uncensored and untruncated losses from the stated exponential model.

Why the Rate-Based Estimate Agrees

The same observations can be modeled using the rate $\lambda=1/\theta$. In that parameterization, the likelihood is $L(\lambda)=\lambda^n\exp(-\lambda\sum x_j)$, for $\lambda>0$. Its log-likelihood is $n\log\lambda-\lambda\sum x_j$, whose derivative is $n/\lambda-\sum x_j$. Setting the derivative to zero yields $\widehat{\lambda}=n/\sum x_j=1/\bar{x}$. Thus, the rate estimate and mean estimate are reciprocals.

This is also an instance of the invariance property of maximum likelihood: if a parameter's MLE is $\widehat{\lambda}$, then the MLE of a one-to-one transformation $g(\lambda)$ is $g(\widehat{\lambda})$. Here $\theta=1/\lambda$, so the mean estimate obtained from the rate estimate is $1/\widehat{\lambda}=\bar{x}$. This confirms that both parameterizations lead to the same fitted exponential distribution.

Worked Example: Estimating the Mean from Losses in Thousands

A hypothetical insurer observes four complete, independent claim severities, measured in thousands of dollars: $12, 18, 24,$ and $30$. Assume the severities follow an exponential distribution. Estimate the mean parameter and rate, and calculate the fitted probability that a loss exceeds 42 thousand dollars.

Method. Use the complete-data mean MLE, then transform it to the rate parameter. The observed sum is $12+18+24+30=84$ thousand dollars, and the sample size is $n=4$.

Mean estimate. The sample mean is $84/4=21$ thousand dollars, so $\widehat{\theta}_{\mathrm{MLE}}=21$ thousand dollars.

Rate estimate. The reciprocal is $\widehat{\lambda}=1/21=0.047619$ per thousand dollars. Equivalently, the rate-based formula gives $n/\sum x_j=4/84=1/21$ per thousand dollars.

Exceedance probability. For an exponential severity with mean $\theta$, $P(X>x)=e^{-x/\theta}$ for $x\geq0$. Therefore the fitted probability is $e^{-42/21}=e^{-2}=0.1353$, to four decimal places.

Check and trap avoided. The estimated rate is exactly the reciprocal of the estimated mean, and the units agree: thousand dollars for the mean and per thousand dollars for the rate. A common error is to report $1/21$ thousand dollars as the mean estimate. That is the rate estimate, not the mean; the mean estimate is $21$ thousand dollars.

Worked Examples

Worked Example: Checking Both Parameterizations and the Likelihood

A hypothetical portfolio has four complete, independent claim severities of $\$300,\$500,\$700,$ and $\$900$. Assume an exponential severity model. Find the MLE of the mean, transform the rate-based MLE, and write the maximized likelihood in the mean and rate parameterizations.

Method. First compute the sample mean, then use the likelihood at that estimate. The sum is $300+500+700+900=\$2{,}400$, and $n=4$.

Mean estimate. $\widehat{\theta}_{\mathrm{MLE}}=\$2{,}400/4=\$600$.

Rate estimate. From the rate likelihood, $\widehat{\lambda}=4/2{,}400=1/600$ per dollar. Transforming back gives $1/\widehat{\lambda}=\$600$, agreeing with the sample-mean estimate.

Likelihood at the maximum. In mean form, the likelihood at $\widehat{\theta}=\$600$ is $600^{-4}\exp[-(2{,}400/600)]=600^{-4}e^{-4}$, approximately $1.4132\times10^{-13}$ per dollar to the fourth power. In rate form it is $(1/600)^4\exp[-(1/600)(2{,}400)]=(1/600)^4e^{-4}$, the same value.

Check and trap avoided. The two expressions agree because $1/\widehat{\theta}=\widehat{\lambda}$, and both use the same total loss. Do not substitute the mean parameter where a rate belongs in the rate-form density, or omit the exponent involving the sum of the observed losses.

Worked Example: Interpreting the Estimate in Thousands

A hypothetical product has three complete, independent claim severities recorded in thousands of dollars: $5,10,$ and $15$. Assume they follow an exponential distribution. Estimate its mean and rate, and find the fitted probability that a claim exceeds 20 thousand dollars.

Method. Use the sample mean as the mean-parameter MLE and check it against the rate MLE. The observed total is $5+10+15=30$ thousand dollars, with $n=3$.

Mean estimate. The sample mean is $30/3=10$ thousand dollars, so $\widehat{\theta}_{\mathrm{MLE}}=10$ thousand dollars.

Rate estimate. The rate MLE is $n/\sum x_j=3/30=0.1$ per thousand dollars. Its reciprocal is $10$ thousand dollars, which matches the mean estimate.

Fitted probability. Using the mean form of the survival probability gives $P(X>20)=e^{-20/10}=e^{-2}=0.1353$, to four decimal places.

Check and trap avoided. The threshold and mean are both measured in thousands of dollars, so their ratio is dimensionless. If the threshold were entered as $\$20$ while the mean remained $10$ thousand dollars, the calculation would mix units and give an incorrect result. Equivalently, one could use $\$20{,}000$ and a mean of $\$10{,}000$.

Assumptions, Units, and Common Pitfalls

The estimator $\bar{x}$ follows from a specific likelihood, not simply from the fact that the severity model is exponential. It requires independent observations from the same exponential distribution, with every sampled loss observed in full. If records are truncated, censored, or restricted by a deductible, the likelihood must reflect how the observations entered the data. Applying the complete-data sample-mean formula without accounting for that observation process is generally not justified.

The parameterization also determines the meaning and units of every estimate. The mean $\theta$ has the same units as a loss. The rate $\lambda$ has inverse-loss units. Consequently, an estimate of $500$ dollars for $\theta$ corresponds to an estimate of $1/500$ per dollar for $\lambda$, not $500$ per dollar. When calculating probabilities such as $P(X>x)$, $x$ and $\theta$ must use the same units so that $x/\theta$ is dimensionless.

The fitted mean is an estimate of the model's expected severity, not a claim that each future loss will equal the sample mean. Individual exponential losses remain variable. The estimate identifies the parameter that maximizes the likelihood of the observed complete sample under the assumed model.

Key takeaway: For complete independent exponential severity observations, the MLE of the mean is the sample mean. The corresponding rate MLE is its reciprocal. Check both the parameterization and the units before interpreting or using either estimate.

Check Your Understanding

For each problem, assume the stated complete losses are independent observations from an exponential severity model. Answers involving probabilities are rounded to four decimal places.

A hypothetical coverage produces complete losses of $4,8,$ and $12$ thousand dollars. Find the mean-parameter MLE and the rate-parameter MLE, including their units. (Answer: $\widehat{\theta}=8$ thousand dollars; $\widehat{\lambda}=0.125$ per thousand dollars.)
Four complete hypothetical severities are $\$250,\$350,\$650,$ and $\$750$. Estimate the mean and verify it by transforming the rate estimate. (Answer: $\widehat{\theta}=\$500$; $\widehat{\lambda}=0.002$ per dollar, whose reciprocal is $\$500$.)
Six complete observations have total severity $\$9{,}600$. Find the mean and rate MLEs. (Answer: $\widehat{\theta}=\$1{,}600$; $\widehat{\lambda}=0.000625$ per dollar.)
Three complete severities, in thousands of dollars, are $2,2,$ and $8$. Find the fitted mean and calculate the probability that a future loss exceeds 6 thousand dollars. (Answer: $\widehat{\theta}=4$ thousand dollars; $P(X>6)=e^{-1.5}=0.2231$.)
A complete-sample rate-based MLE is $0.04$ per dollar. State the corresponding mean estimate and explain what happens to the units under the transformation. (Answer: $\widehat{\theta}=25$ dollars; taking the reciprocal changes per-dollar units to dollars.)

What You'll Learn