Fisher Information for Exponential Loss Models

Why Fisher Information Matters for Exponential Losses

An exponential model is often used as a simple severity model when losses are positive and have a constant hazard rate. Estimating its rate parameter is straightforward, but assessing the precision of that estimate requires care: Fisher information depends on how the model is parameterized. In particular, the information for the rate is not the information for the mean.

This tutorial considers independent, identically distributed, complete observations from an exponential loss distribution. “Complete” means that every observed loss is available as an exact value, rather than only being known to exceed a deductible or to have been censored at a threshold. Truncated or censored likelihoods have different likelihood contributions and are not covered by the formulas below.

Definition: Let $X_1,\ldots,X_n$ be independent observations from an exponential distribution with rate parameter $\lambda>0$. Its density and mean are $$ f_X(x;\lambda)=\lambda e^{-\lambda x},\quad x\geq 0, \qquad E[X]=\frac{1}{\lambda}. $$ Here $\lambda$ is measured in inverse loss units. For example, if losses are measured in dollars, $\lambda$ is in inverse dollars. The interest rate and mortality assumptions do not apply to this severity-estimation model.

Likelihood, Score, and the Rate Estimate

The likelihood is the joint density of the observations, viewed as a function of $\lambda$. Independence allows the individual densities to be multiplied. Taking logarithms turns that product into a sum, which makes differentiation simpler.

L(\lambda)=\prod_{j=1}^{n}\lambda e^{-\lambda X_j} =\lambda^n e^{-\lambda\sum_{j=1}^{n}X_j}, \qquad \ell(\lambda)=n\log\lambda-\lambda\sum_{j=1}^{n}X_j.

The score is the derivative of the log-likelihood with respect to the parameter. It indicates the direction in which the log-likelihood is locally increasing. Setting the score equal to zero gives the maximum-likelihood estimate (MLE), provided the observed losses are positive and their sum is positive.

U(\lambda)=\frac{\partial\ell}{\partial\lambda} =\frac{n}{\lambda}-\sum_{j=1}^{n}X_j, \qquad \widehat{\lambda}_{\mathrm{MLE}} =\frac{n}{\sum_{j=1}^{n}X_j} =\frac{1}{\overline{X}}.

The second derivative is $-n/\lambda^2$, which is negative for every $\lambda>0$. Thus the log-likelihood is strictly concave, and the score’s zero is the unique maximum. The result also shows why the MLE is the reciprocal of the sample mean: for an exponential distribution, the population mean is $1/\lambda$.

Worked Example: Estimating a Rate From Four Complete Losses

A hypothetical portfolio study records four complete losses, in dollars: $20, 30, 50,$ and $60$. Assume these are independent observations from the same exponential distribution. No interest or mortality assumption is involved. Estimate the rate and calculate the expected Fisher information at the estimate.

Step 1: Sum the observations and find the sample mean. $$ \sum_{j=1}^{4}X_j=20+30+50+60=160, \qquad \overline{X}=\frac{160}{4}=40\text{ dollars}. $$

Step 2: Calculate the MLE. $$ \widehat{\lambda}=\frac{4}{160}=\frac{1}{40}=0.025 \text{ per dollar}. $$

Step 3: Evaluate the sample expected information. For $n$ independent complete observations, the expected information is $n/\lambda^2$. Substituting the estimate gives $$ \widehat{I}(\widehat{\lambda}) =\frac{4}{(0.025)^2} =\frac{4}{0.000625} =6{,}400\text{ dollars}^2. $$ The corresponding information-based variance estimate is $1/6{,}400=0.00015625$ per dollar squared, and the standard-error estimate is $\sqrt{0.00015625}=0.0125$ per dollar.

Check. The estimated mean loss is $1/\widehat{\lambda}=40$ dollars, equal to the sample mean. The reported information is positive, and its reciprocal gives the stated variance estimate. The method is the plug-in Fisher-information method; it avoids treating the information itself as a variance.

Fisher Information and Its Derivation

For a scalar parameter, the expected Fisher information is the expected negative second derivative of the log-likelihood. Under the regularity conditions for the information identity, it also equals the expected square of the score. Here those quantities can both be evaluated directly.

-\frac{\partial^2\ell}{\partial\lambda^2} =\frac{n}{\lambda^2}, \qquad I_n(\lambda) =E\!\left[-\frac{\partial^2\ell}{\partial\lambda^2}\right] =\frac{n}{\lambda^2}.

The expression does not depend on the observed losses, so in this model the observed negative second derivative and expected information happen to be identical. The information for one observation is $I_1(\lambda)=1/\lambda^2$. The sample expression follows directly from the log-likelihood for all $n$ observations; no incomplete-data likelihood is being used.

The score-based identity gives the same result. For one observation, $U_1(\lambda)=1/\lambda-X_1$. Since $E[X_1]=1/\lambda$ and $\operatorname{Var}(X_1)=1/\lambda^2$, the score has mean zero and $$ E[U_1(\lambda)^2] =\operatorname{Var}\!\left(\frac{1}{\lambda}-X_1\right) =\operatorname{Var}(X_1) =\frac{1}{\lambda^2}. $$ For the full sample, the score is a sum of independent mean-zero individual scores, so its variance is $n/\lambda^2$, agreeing with the expected negative second derivative.

This is a direct justification of the information formula: the curvature of the log-likelihood and the variability of the score both quantify how strongly the likelihood distinguishes nearby values of $\lambda$. Larger information corresponds to a smaller information-based variance estimate.

Theorem: For $n$ independent complete observations from an exponential distribution with rate $\lambda>0$, the expected Fisher information is $I_n(\lambda)=n/\lambda^2$. Replacing the unknown rate by its MLE gives the estimated variance $\widehat{\operatorname{Var}}(\widehat{\lambda})=\widehat{\lambda}^{\,2}/n$ and estimated standard error $\widehat{\lambda}/\sqrt{n}$. These are information-based, large-sample uncertainty estimates; they are not a claim that the finite-sample MLE is exactly normally distributed.

Rate Versus Mean Parameterization

A useful check on calculations is to distinguish the rate $\lambda$ from the mean $\theta=1/\lambda$. The density can instead be written $f_X(x;\theta)=\theta^{-1}e^{-x/\theta}$, with $\theta>0$. The sample information for this mean parameter is $n/\theta^2$, not $n/\lambda^2$. The units also differ: information for $\theta$ has units of inverse dollars squared, while information for $\lambda$ has units of dollars squared.

The parameters are related nonlinearly, so their information values are not interchangeable. In this one-parameter case, the chain rule provides a consistency check: $d\theta/d\lambda=-1/\lambda^2$. Information for a reparameterized scalar parameter is multiplied by the square of the derivative of the original parameter with respect to the new parameter. Thus $I_\lambda=I_\theta(d\theta/d\lambda)^2$, which yields $n/\lambda^2$ when $\theta=1/\lambda$.

Worked Example: Checking Both Parameterizations

A hypothetical analysis observes three independent, complete losses of $20, 40,$ and $60$ dollars. Assume an exponential model. There is no interest rate or mortality assumption. Find the MLE of the rate, compute information for both the mean and rate parameterizations at their estimates, and obtain the rate’s information-based standard error.

Step 1: Calculate the sample mean and rate estimate. $$ \sum X_j=20+40+60=120, \qquad \overline{X}=\frac{120}{3}=40, \qquad \widehat{\lambda}=\frac{3}{120}=\frac{1}{40}=0.025. $$

Step 2: Calculate information for the mean parameter. Here $\widehat{\theta}=\overline{X}=40$ dollars, so $$ \widehat{I}(\widehat{\theta}) =\frac{3}{40^2} =\frac{3}{1{,}600} =0.001875\text{ per dollar}^2. $$ This is information for the mean parameter, not for the rate. Its reciprocal, $1/0.001875=533.3333$ dollars squared, is the information-based variance estimate for $\widehat{\theta}$.

Step 3: Calculate information and uncertainty for the rate. $$ \widehat{I}(\widehat{\lambda}) =\frac{3}{(0.025)^2} =\frac{3}{0.000625} =4{,}800\text{ dollars}^2, $$ $$ \widehat{\operatorname{Var}}(\widehat{\lambda}) =\frac{1}{4{,}800} =0.0002083333\text{ per dollar}^2, \qquad \widehat{\operatorname{SE}}(\widehat{\lambda}) =\sqrt{0.0002083333} \approx0.01443\text{ per dollar}. $$

Step 4: Check the transformation. The derivative of $\lambda=1/\theta$ with respect to $\theta$, evaluated at $40$, is $-1/40^2=-1/1{,}600$. Transforming the mean-parameter variance gives $$ \left(\frac{1}{1{,}600}\right)^2(533.3333) \approx0.0002083333, $$ the same rate variance estimate as the inverse rate information. The information values $0.001875$ and $4{,}800$ differ because they refer to different parameters; neither should be substituted for the other.

Using the Plug-In Estimate and Avoiding Traps

Because $\lambda$ is unknown, a practical information-based variance estimate substitutes $\widehat{\lambda}$ into $I_n(\lambda)$ and takes the reciprocal: $$ \widehat{\operatorname{Var}}(\widehat{\lambda}) =\frac{1}{I_n(\widehat{\lambda})} =\frac{\widehat{\lambda}^{\,2}}{n} =\frac{1}{n\overline{X}^{\,2}}. $$ The square root is the estimated standard error. This calculation is useful for summarizing the approximate precision of the rate estimate. It does not alter the MLE itself, and it does not mean the variance is the information; information is inverted to obtain the variance estimate.

Worked Example: Applying the Plug-In Variance Formula

A hypothetical loss study has five independent, complete observations from an exponential distribution, with observed total loss $125$ dollars. Find the rate MLE and its information-based variance and standard-error estimates. No interest or mortality assumption is relevant.

Step 1: Find the mean and rate estimate. $$ \overline{X}=\frac{125}{5}=25\text{ dollars}, \qquad \widehat{\lambda}=\frac{5}{125}=\frac{1}{25}=0.04 \text{ per dollar}. $$

Step 2: Find the estimated information. $$ \widehat{I}(\widehat{\lambda}) =\frac{5}{(0.04)^2} =\frac{5}{0.0016} =3{,}125\text{ dollars}^2. $$

Step 3: Invert for variance, then take the square root. $$ \widehat{\operatorname{Var}}(\widehat{\lambda}) =\frac{1}{3{,}125} =0.00032\text{ per dollar}^2, \qquad \widehat{\operatorname{SE}}(\widehat{\lambda}) =\sqrt{0.00032} \approx0.01789\text{ per dollar}. $$

Check. The direct formula gives $(0.04)^2/5=0.0016/5=0.00032$, matching the reciprocal-information calculation. The method is the plug-in information estimate; the common mistake it avoids is reporting $3{,}125$ as a variance rather than inverting it.

Keep the following distinctions clear when solving an estimation problem:

Rate versus mean: $\lambda$ is the reciprocal of the exponential mean. Use the information formula for the parameter actually being estimated.
Information versus variance: variance is estimated by the reciprocal of information. The estimated standard error is the square root of that variance.
Complete versus incomplete observations: the likelihood above assumes exact, uncensored observations. A censored or truncated sample requires the corresponding likelihood, score, and information.
Sample size: the sample information is $n/\lambda^2$, whereas the information for one observation is $1/\lambda^2$. Do not omit $n$ when calculating sample uncertainty.

The exponential model’s simple formulas make it a useful setting for understanding how likelihood curvature becomes an uncertainty estimate. The same discipline applies more broadly: state the parameterization, write the likelihood for the actual observation process, derive the information for that parameter, and only then invert and take a square root if a standard error is required. This treatment is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, assume independent complete observations from the stated exponential model. No interest or mortality assumption applies. Check that each information value and variance estimate refers to the parameter named.

A hypothetical sample contains losses of $10, 30,$ and $50$ dollars. Find $\widehat{\lambda}$, the sample information at the estimate, and the information-based variance estimate for the rate. (Answer: $\widehat{\lambda}=1/30=0.03333$ per dollar; $\widehat{I}=2{,}700$ dollars squared; estimated variance $=1/2{,}700\approx0.0003704$ per dollar squared.)
For $n=4$ complete exponential losses, the sample mean is $20$ dollars. Find the rate MLE and estimated standard error. (Answer: $\widehat{\lambda}=0.05$ per dollar; estimated standard error $=0.025$ per dollar.)
A sample has $n=5$ and $\widehat{\lambda}=0.02$ per dollar. Find the estimated expected information and the information-based variance estimate. (Answer: $\widehat{I}=12{,}500$ dollars squared; estimated variance $=0.00008$ per dollar squared.)
For an exponential mean parameter $\theta$, a sample of $n=3$ has sample mean $30$ dollars. Find information for $\theta$, then give the rate estimate and rate information. (Answer: $\widehat{I}(\theta)=3/30^2=0.003333$ per dollar squared; $\widehat{\lambda}=1/30$ per dollar; $\widehat{I}(\lambda)=2{,}700$ dollars squared.)
A researcher observes five complete losses with total $200$ dollars. Calculate the rate MLE, information-based rate variance estimate, and standard error. (Answer: $\widehat{\lambda}=5/200=0.025$ per dollar; estimated variance $=0.000125$ per dollar squared; estimated standard error $\approx0.01118$ per dollar.)

What You'll Learn