Poisson Rate Estimation From Sample Moments

Why the Sample Mean Estimates a Poisson Rate

Count data arise in many actuarial settings: claims reported during a period, service requests received in an hour, or insured events observed over a fixed amount of exposure. When a Poisson model is appropriate, its parameter is also the expected count over the observation interval. That link makes the first raw sample moment a direct route to a method-of-moments estimate.

The key is to match like with like. For independent counts observed over equal intervals, the population first raw moment is the expected count per interval, and the sample first raw moment is the average observed count per interval. If intervals have different exposures, the unadjusted sample mean generally estimates an average count across those particular intervals, not the rate per unit of exposure.

Model and assumptions: For the basic method below, \(X_1,\ldots,X_n\) are independent and identically distributed counts with \(X_j\sim\operatorname{Poisson}(\lambda)\), where \(\lambda\geq 0\). Each count covers the same observation interval and uses the same count definition. The parameter \(\lambda\) is the expected number of events in one such interval. If the intended parameter space is restricted to \(\lambda>0\), an all-zero sample produces a boundary moment value that is not in that parameter space.

For \(\lambda>0\), the Poisson probability mass function is \(P(X=k)=e^{-\lambda}\lambda^k/k!\) for \(k=0,1,2,\ldots\). At \(\lambda=0\), the distribution is degenerate: \(P(X=0)=1\). The first raw population moment is \(E[X]\), while the first raw sample moment is \(m'_1=\frac{1}{n}\sum_{j=1}^{n}X_j=\bar X\).

Definition: In the method of moments, equate a sample raw moment to the corresponding population raw moment expressed as a function of the unknown parameter, then solve the resulting equation. For the one-parameter Poisson model with equal observation intervals, this gives \(\hat{\lambda}_{\mathrm{MM}}=\bar X\).

Deriving the First Raw Moment

The formula for the Poisson estimate follows from the Poisson mean. For \(\lambda>0\), start with the defining probability mass function and use \(k/k!=1/(k-1)!\) for positive integers \(k\):

\begin{aligned} E[X] &=\sum_{k=0}^{\infty} kP(X=k)\\ &=\sum_{k=1}^{\infty}k e^{-\lambda}\frac{\lambda^k}{k!}\\ &=e^{-\lambda}\lambda\sum_{k=1}^{\infty}\frac{\lambda^{k-1}}{(k-1)!}\\ &=e^{-\lambda}\lambda\sum_{r=0}^{\infty}\frac{\lambda^r}{r!}\\ &=e^{-\lambda}\lambda e^\lambda\\ &=\lambda. \end{aligned}

The substitution \(r=k-1\) makes the remaining sum the exponential series \(e^\lambda\), which cancels the factor \(e^{-\lambda}\). At \(\lambda=0\), \(X=0\) with probability one, so \(E[X]=0=\lambda\) there as well. Thus \(E[X]=\lambda\) throughout the parameter space \(\lambda\geq0\).

For the equal-interval model, the method-of-moments equation is therefore \(m'_1=E[X]\), or \(\bar X=\lambda\). Solving gives \(\hat{\lambda}_{\mathrm{MM}}=\bar X\). The estimate has the same units as the count per observation interval: if observations are daily counts, it is an estimated number of events per day; it is not automatically an estimate per hour or per year.

Identify the observation unit.
Confirm that every count covers the same kind of interval and that the model uses one common Poisson mean.

Calculate the first raw sample moment.
Add the observed counts and divide by the number of intervals: \(\bar X=\sum X_j/n\).

Match and interpret.
Set \(\bar X=\lambda\), then report the estimate in events per observation interval. Convert units only after identifying the interval represented by the data.

Worked Examples for Equal Observation Intervals

Worked Example: Estimating Events Per Day From Daily Counts

A hypothetical monitoring team records the number of covered equipment incidents reported on each of eight days. Assume the daily counts are independent and identically distributed as \(\operatorname{Poisson}(\lambda)\), with \(\lambda\geq0\). The recorded counts are \(1,0,2,1,3,0,1,2\). Estimate the expected daily count by the method of moments.

Method. Match the first raw sample moment to the Poisson first raw population moment, \(E[X]=\lambda\). Since the intervals are all one day, the estimate is the sample mean.

Calculate the total. The sum is \(1+0+2+1+3+0+1+2=10\) incidents across eight days.

Calculate the sample moment. Thus \(\bar X=10/8=1.25\). The method-of-moments estimate is \(\hat{\lambda}_{\mathrm{MM}}=1.25\) incidents per day.

Check. Multiplying the reported mean by the number of days recovers the observed total: \(1.25(8)=10\). A common mistake is to report 10 as the estimated daily mean; 10 is the total across all eight days, not the average per day.

Worked Example: Converting a Per-Shift Estimate to an Hourly Rate

A hypothetical call centre records counts over twelve independent shifts, each lasting four hours. Assume each shift count has a common Poisson distribution with mean \(\lambda_{\text{shift}}\). The counts are \(0,1,2,1,0,3,1,2,1,0,2,1\). Estimate the expected count per shift and the corresponding rate per hour.

Method. First use the first raw sample moment to estimate the mean count per four-hour shift. Then divide by four to express the estimated average count per hour. This unit conversion is appropriate because every shift has the same four-hour exposure.

Sum the counts. Their total is \(0+1+2+1+0+3+1+2+1+0+2+1=14\) calls.

Estimate the shift mean. There are twelve shifts, so \(\hat{\lambda}_{\text{shift}}=14/12=7/6\), or approximately \(1.1667\) calls per shift.

Convert the unit. The total exposure is \(12(4)=48\) hours. The estimated hourly rate is \(14/48=7/24\), or approximately \(0.2917\) calls per hour. Equivalently, \((7/6)/4=7/24\).

Check. The two calculations agree, and \((7/24)(48)=14\) reproduces the total count. The intermediate rates are shown to four decimal places; the exact fractions are retained. Reporting \(7/6\) as calls per hour would be a time-unit error: it is the estimate per four-hour shift.

Worked Example: An All-Zero Sample and the Boundary Value

A hypothetical portfolio records covered claims over six equal, independent observation periods. Suppose each period count follows \(\operatorname{Poisson}(\lambda)\). The observed counts are \(0,0,0,0,0,0\). Find the method-of-moments result and explain what it means if the model requires \(\lambda>0\).

Method. Use the same first-moment equation, \(\bar X=\lambda\). The sample mean is zero because the sum of the six observations is zero: \(\bar X=(0+0+0+0+0+0)/6=0\).

Result with a closed parameter space. If \(\lambda\geq0\) is allowed, the method-of-moments estimate is \(\hat{\lambda}_{\mathrm{MM}}=0\). The fitted Poisson model then assigns probability one to a zero count in each period.

Result with a strictly positive parameter space. If the model stipulates \(\lambda>0\), the moment equation still yields the value zero, but zero is outside the stipulated parameter space. Therefore there is no admissible positive solution to the moment equation for this sample. Do not silently describe zero as a positive-rate estimate.

Check. An all-zero sample is possible even when the true rate is positive: under the stated independent model its probability is \((e^{-\lambda})^6=e^{-6\lambda}>0\) for every \(\lambda>0\). The boundary result is therefore not evidence that the event rate must truly be zero; it is the moment equation’s result for these data.

Unequal Exposure Requires an Exposure Adjustment

If observation periods have different known exposures, counts can have different means even when they share the same underlying rate. Let \(X_j\sim\operatorname{Poisson}(\lambda t_j)\) independently, where \(t_j>0\) is the exposure for period \(j\), measured in a common unit. Here \(\lambda\) is the expected count per exposure unit, and the expected count for period \(j\) is \(\lambda t_j\). In this setting, \(\bar X\) is not generally an estimate of \(\lambda\) per exposure unit.

One direct moment equation uses the total count \(T=\sum_{j=1}^{n}X_j\). By linearity of expectation, \(E[T]=\sum_j E[X_j]=\lambda\sum_j t_j\). Matching this expected total to the observed total gives the exposure-adjusted moment estimate \(\hat{\lambda}_{\mathrm{MM}}=T/\sum_j t_j\). This is a total count divided by total exposure, not an unadjusted mean of the period counts.

Worked Example: Estimating a Rate With Unequal Exposures

A hypothetical insurer observes independent claim counts for four groups of policies. Their exposures are \(2,1,3,2\) policy-years, and the corresponding counts are \(3,1,4,2\). Assume \(X_j\sim\operatorname{Poisson}(\lambda t_j)\), with one common rate \(\lambda\) per policy-year. Estimate that rate by matching the first moment of the total count.

Method. Use \(E[T]=\lambda\sum_j t_j\), then equate the observed total count to its model expectation.

Calculate total count and exposure. The observed total is \(T=3+1+4+2=10\) claims. The total exposure is \(2+1+3+2=8\) policy-years.

Solve the moment equation. Since \(E[T]=8\lambda\), matching \(T=10\) to this expectation gives \(10=8\lambda\). Hence \(\hat{\lambda}_{\mathrm{MM}}=10/8=1.25\) claims per policy-year.

Check and avoid the trap. The unadjusted average count is \(10/4=2.5\) claims per group, but groups have unequal exposures, so that is not the rate per policy-year. The adjusted estimate satisfies \(1.25(8)=10\), recovering the observed total.

Common Mistakes and Interpretation

The method is short, but a correct answer depends on matching the data structure and units to the model. Before reporting an estimate, check the assumptions and the arithmetic rather than applying “the estimate is the mean” without qualification.

Using the total instead of the mean. For \(n\) equal intervals, the first raw sample moment is \(\sum_j X_j/n\), not \(\sum_j X_j\). The total is the numerator; divide by the number of intervals.
Confusing raw and central moments. The first raw sample moment is \(\bar X\). The first central sample moment is zero by construction and cannot estimate the Poisson mean.
Ignoring unequal exposure. When \(E[X_j]=\lambda t_j\), divide the total count by total exposure to estimate a rate per exposure unit. Do not treat the unadjusted group average as that rate.
Changing time units incorrectly. A mean count per four-hour shift must be divided by four to express a count per hour. State the interval or exposure unit beside the estimate.
Overlooking the parameter boundary. If every count is zero, the moment equation yields zero. This is admissible when \(\lambda\geq0\), but not when the stipulated parameter space is \(\lambda>0\).

These checks are especially useful in actuarial count applications, where “rate” might mean a count per policy-year, per month, or per reporting period. The numerical value alone is incomplete unless its exposure unit is clear. The estimate is a method-of-moments result under the specified Poisson model; the model assumptions and the observation design determine whether the formula applies.

Key takeaway: For independent, identically distributed Poisson counts over equal intervals, \(E[X]=\lambda\) makes the method-of-moments estimate the sample mean. For counts with means \(\lambda t_j\) over unequal known exposures, match the total count to its expected total and estimate the rate as total count divided by total exposure. An all-zero sample gives the boundary value zero.

Check Your Understanding

For each problem, identify the relevant count or exposure unit before calculating.

A hypothetical sensor records independent counts over five equal nights: \(0,2,1,4,3\). Under a common Poisson mean per night, what is the method-of-moments estimate per night? (Answer: \(2\) events per night.)
A hypothetical dispatch desk records \(1,0,2,2,1,3,0,1,2\) calls over nine equal days. Estimate the expected number of calls per day. (Answer: \(12/9=4/3\), or approximately \(1.3333\) calls per day.)
A hypothetical study records zero covered events in each of four equal, independent periods. What does the moment equation give, and what qualification applies if the parameter space requires \(\lambda>0\)? (Answer: The equation gives \(\hat{\lambda}_{\mathrm{MM}}=0\); it is a boundary value and is not an admissible estimate in a strictly positive parameter space.)
Independent hypothetical policy groups have exposures \(1,2,4\) policy-years and observed counts \(2,3,7\). Under \(X_j\sim\operatorname{Poisson}(\lambda t_j)\), estimate the rate per policy-year. (Answer: \(12/7\approx1.7143\) claims per policy-year.)
A hypothetical system records \(2,1,4,0,3,2,5,1,1,5\) events over ten equal days. Assuming a constant Poisson mean per day, use the moment estimate to calculate the corresponding estimated expected count over 30 days. (Answer: The total is \(24\), so the estimate is \(24/10=2.4\) per day and \(30(2.4)=72\) events over 30 days.)

What You'll Learn