Method-of-Moments Estimation for Exponential Losses

Why the Sample Mean Determines an Exponential Fit

An exponential model is a one-parameter model for a positive loss amount. Its simple relationship between the parameter and the mean makes method of moments especially direct: equating the model mean to the observed sample mean determines the fitted distribution. The calculation is short, but interpreting the parameter, keeping track of units, and distinguishing model variance from sample variance require care.

Assume $X_1,\ldots,X_n$ are independent, identically distributed, fully observed loss amounts from an exponential distribution. The sample is not truncated or censored, and the model applies to the losses being analyzed. For the rate form, the probability density is defined for $x\geq 0$ and has one positive parameter $\lambda$.

Definition: An exponential loss with rate $\lambda>0$ has density $f_X(x)=\lambda e^{-\lambda x}$ for $x\geq 0$, and $f_X(x)=0$ otherwise. Equivalently, its scale is $\theta=1/\lambda$, and its density is $f_X(x)=\frac{1}{\theta}e^{-x/\theta}$ for $x\geq 0$. The rate has units inverse to the loss amount; the scale has the same units as the loss.

The first raw sample moment is the sample mean, $\bar X=\frac{1}{n}\sum_{j=1}^{n}X_j$. Method of moments sets this observed moment equal to the corresponding model moment. For an exponential model, this gives one equation for the one unknown parameter.

Deriving the Model Moments and Estimators

For the rate parameterization, the model mean follows by integrating the density. Integration by parts gives the first moment; applying the same integral calculation to the second raw moment gives the variance.

E[X]=\int_0^\infty x\lambda e^{-\lambda x}\,dx=\frac{1}{\lambda}, \qquad E[X^2]=\int_0^\infty x^2\lambda e^{-\lambda x}\,dx=\frac{2}{\lambda^2}, \qquad \operatorname{Var}(X)=E[X^2]-(E[X])^2=\frac{1}{\lambda^2}.

For example, in the first integral take $u=x$ and $dv=\lambda e^{-\lambda x}dx$. Then $du=dx$ and $v=-e^{-\lambda x}$, so the boundary term $[-xe^{-\lambda x}]_0^\infty$ is zero and the remaining integral is $\int_0^\infty e^{-\lambda x}dx=1/\lambda$. The second-moment integral similarly reduces to $2\int_0^\infty xe^{-\lambda x}dx=2/\lambda^2$. Thus the variance is $2/\lambda^2-(1/\lambda)^2=1/\lambda^2$.

Theorem: For a complete independent sample from an exponential model, the method-of-moments rate estimate is $\hat\lambda_{\mathrm{MM}}=1/\bar X$, and the method-of-moments scale estimate is $\hat\theta_{\mathrm{MM}}=\bar X$. The fitted model mean is $\bar X$, and its fitted model variance is $\bar X^2$, provided $\bar X>0$.

The rate estimate follows by setting $E[X]=\bar X$: $1/\lambda=\bar X$, so $\lambda=1/\bar X$. In the scale parameterization $E[X]=\theta$, so $\theta=\bar X$. Substitution into the model variance yields $1/\hat\lambda_{\mathrm{MM}}^2=\hat\theta_{\mathrm{MM}}^2=\bar X^2$. The two parameterizations therefore describe the same fitted distribution, not two different fits.

Calculate the sample mean.
Add the complete observed losses and divide by the number of losses.

Identify the parameterization.
For a rate, take the reciprocal of the sample mean; for a scale, use the sample mean directly.

Calculate the fitted moments.
The fitted mean is $\bar X$, and the fitted variance is $\bar X^2$, in squared loss units.

Worked Calculations

Worked Example: Estimating a Rate From Dollar Losses

A hypothetical insurer observes five fully reported losses of $800, $1,200, $1,600, $2,400, and $4,000. Assume the losses are independent observations from one exponential distribution. Estimate the rate per dollar and report the fitted mean and variance.

Method. Use the first-moment equation for a complete exponential sample. This avoids confusing the reciprocal rate with the scale.

Step 1: Find the sample mean. The total is $800+1{,}200+1{,}600+2{,}400+4{,}000=10{,}000$ dollars, and $n=5$. Therefore $\bar X=10{,}000/5=\$2{,}000$.

Step 2: Estimate the rate. The rate is the reciprocal of the mean, so $\hat\lambda_{\mathrm{MM}}=1/2{,}000=0.0005$ per dollar.

Step 3: Calculate fitted moments. The fitted mean is $1/\hat\lambda_{\mathrm{MM}}=\$2{,}000$. The fitted variance is $1/(0.0005)^2=4{,}000{,}000$ dollars squared.

Check. The scale is $1/\hat\lambda_{\mathrm{MM}}=\$2{,}000$, agreeing with the sample mean. Squaring that scale gives the same fitted variance, $\$2{,}000^2=4{,}000{,}000$ dollars squared. The common mistake this calculation avoids is reporting 0.0005 dollars as the mean; 0.0005 is a rate measured per dollar.

Worked Example: Fitting the Scale in Thousands of Dollars

A hypothetical claim study records five complete losses, in thousands of dollars: 0.5, 1.0, 1.5, 2.0, and 3.0. Fit the exponential model using the scale parameterization. Report the scale, equivalent rate, fitted mean, and fitted variance.

Method. In the scale parameterization, the model mean equals $\theta$. Keep the calculations in thousands of dollars until converting the variance to dollars squared.

Step 1: Find the sample mean. The sum is $0.5+1.0+1.5+2.0+3.0=8.0$ thousand dollars. Thus $\bar X=8.0/5=1.6$ thousand dollars.

Step 2: Estimate both parameterizations. The scale estimate is $\hat\theta_{\mathrm{MM}}=1.6$ thousand dollars. The equivalent rate is $\hat\lambda_{\mathrm{MM}}=1/1.6=0.625$ per thousand dollars.

Step 3: Calculate the fitted moments. The fitted mean is 1.6 thousand dollars. The fitted variance is $(1.6)^2=2.56$ thousand-dollars-squared. Since one thousand dollars squared is $1{,}000{,}000$ dollars squared, this is $2{,}560{,}000$ dollars squared.

Check. The product of the estimated rate and scale is $0.625\times1.6=1$, as required by $\theta=1/\lambda$. The variance is in squared units, not thousands of dollars: reporting 2.56 dollars squared would be a unit-conversion error. The fitted standard deviation, if needed, is 1.6 thousand dollars.

Worked Example: Fitted Variance Is Not the Sample Variance

A hypothetical portfolio has three fully observed losses of $1,000, $2,000, and $3,000. Fit an exponential model by moments. Calculate the fitted model variance, then compare it with the unbiased sample variance to identify which quantity answers the model-fitting question.

Method. The method-of-moments fitted variance is obtained by substituting the fitted parameter into the exponential model variance. It is not calculated by taking deviations from the sample mean.

Step 1: Find the sample mean. The sum is $1{,}000+2{,}000+3{,}000=6{,}000$ dollars, so $\bar X=6{,}000/3=\$2{,}000$.

Step 2: Fit the model. The scale estimate is $\hat\theta_{\mathrm{MM}}=\$2{,}000$, and the rate estimate is $\hat\lambda_{\mathrm{MM}}=1/2{,}000=0.0005$ per dollar. The fitted model variance is $\hat\theta_{\mathrm{MM}}^2=\$2{,}000^2=4{,}000{,}000$ dollars squared.

Step 3: Calculate the sample variance for comparison. The deviations from the mean are $-\$1{,}000, \$0,$ and $+\$1{,}000$. Their squared deviations sum to $1{,}000{,}000+0+1{,}000{,}000=2{,}000{,}000$ dollars squared. The unbiased sample variance is $2{,}000{,}000/(3-1)=1{,}000{,}000$ dollars squared.

Check and interpretation. The fitted model variance is $4{,}000{,}000$ dollars squared, while the unbiased sample variance is $1{,}000{,}000$ dollars squared. These need not be equal: they are different calculations. The model-fitting question calls for the former. Substituting the estimated scale into the exponential variance formula confirms the fitted result.

Interpretation, Assumptions, and Common Traps

The method is appropriate only when the sample mean is an estimate of the mean of the losses to which the exponential model is being fitted. It presumes a common exponential distribution for independent, completely observed losses. If observations are censored, the recorded values may not be the actual losses; if sampling is truncated, the observed values come from a conditional distribution. In either case, directly equating the raw observed mean to the unconditioned exponential mean is not justified. The observation mechanism must be addressed.

A second distinction is between the variance of the fitted loss distribution and the sampling variance of the estimator. The fitted loss variance is $\bar X^2$. Under the stated independent exponential model, the sample mean itself has sampling variance $\operatorname{Var}(\bar X)=\operatorname{Var}(X)/n$. These quantities answer different questions: one describes variation among losses under the fitted model, while the other describes variation in the sample mean across repeated samples.

Key takeaway: For a complete exponential sample with positive sample mean, the method-of-moments scale is $\bar X$, the rate is $1/\bar X$, the fitted mean is $\bar X$, and the fitted variance is $\bar X^2$. Check the parameterization and units before reporting results.

Using the rate as the mean. The rate is the reciprocal of the mean, not the mean itself.
Squaring the rate to get variance. The rate-form variance is $1/\lambda^2$; equivalently, square the scale.
Using sample variance as fitted variance. The model variance comes from the fitted exponential parameter. It is not generally equal to the sample variance.
Mixing units. If losses are measured in thousands of dollars, the rate is per thousand dollars and the variance is in thousand-dollars-squared. Convert squared units by squaring the scale factor.
Ignoring the observation design. A mean of censored or truncated observations is not automatically the unconditioned exponential mean.

Check Your Understanding

For each problem, use the exponential model and the stated measurement units. Assume complete independent observations unless the problem says otherwise.

A hypothetical sample contains losses of $500, $1,000, $1,500, and $3,000. Find the rate per dollar, fitted mean, and fitted variance in dollars squared. (Answer: $\bar X=\$1{,}500$, $\hat\lambda_{\mathrm{MM}}=1/1{,}500\approx0.0006667$ per dollar, fitted mean $\$1{,}500$, fitted variance $2{,}250{,}000$ dollars squared.)
Five losses, recorded in thousands of dollars, are 0.4, 0.8, 1.2, 1.6, and 1.0. Find the scale, equivalent rate, and fitted variance. (Answer: $\bar X=\hat\theta_{\mathrm{MM}}=1.0$ thousand dollars, $\hat\lambda_{\mathrm{MM}}=1$ per thousand dollars, fitted variance $1.0$ thousand-dollars-squared.)
An exponential model has fitted rate $0.002$ per dollar. Calculate its fitted mean and variance, with units. (Answer: fitted mean $\$500$; fitted variance $250{,}000$ dollars squared.)
A complete sample has mean 2.4 thousand dollars. Give the method-of-moments scale, rate, and fitted variance in both thousand-dollars-squared and dollars-squared. (Answer: scale $2.4$ thousand dollars; rate $1/2.4\approx0.4167$ per thousand dollars; variance $5.76$ thousand-dollars-squared $=5{,}760{,}000$ dollars squared.)
A study records the payment amount only when a loss exceeds a reporting threshold; the observed sample mean is 4,000 dollars. Can $1/4{,}000$ be asserted as the method-of-moments estimate of the unconditioned exponential rate using the complete-sample formula? State why. (Answer: No. The observations are truncated at the threshold, so their raw mean is a conditional mean, not automatically the unconditioned exponential mean.)

What You'll Learn