Why Poisson Claim Counts Have Equal Mean and Variance

Why the Mean and Variance Matter

A limited-fluctuation analysis asks whether observed experience is stable enough to meet a specified precision requirement. For a claim count, the answer depends on how much the count is expected to vary around its mean. Under a Poisson model, the expected count and its variance are equal. That simple relationship lets an actuary obtain the count’s standard deviation directly from its expected value, and it is the key input to a count-based fluctuation standard.

The equality is a model result, not a universal property of claim data. It applies when the count follows a Poisson distribution with a fixed parameter. If risks have materially different rates, experience is dependent, or the underlying rate changes over time, the observed variance may not match the Poisson assumption. The model should therefore be stated before using the equality.

Deriving the Mean and Variance

The probability function gives a direct derivation. For \(\lambda>0\), multiply each possible count \(n\) by its probability to obtain the mean. Re-indexing the sum with \(m=n-1\) leaves the probability-series sum for a Poisson distribution.

To find the variance, first calculate the factorial moment \(E[N(N-1)]\). This moment is useful because \(N(N-1)\) cancels the first two factors in \(n!\), leaving a sum that can again be re-indexed.

Since \(N^2=N(N-1)+N\), it follows that \(E[N^2]=\lambda^2+\lambda\). Subtract the square of the mean to obtain the variance:

Thus the standard deviation is \(\sqrt{\lambda}\). When \(\lambda>0\), the standard deviation divided by the mean—the relative standard deviation—is \(1/\sqrt{\lambda}\). Increasing the expected count increases the absolute standard deviation, but decreases its size relative to the mean.

Combining Independent Poisson Counts

Suppose counts \(N_1,\ldots,N_k\) are independent, with \(N_j\sim\operatorname{Poisson}(\lambda_j)\). Their total is also Poisson, with parameter equal to the sum of the parameters. One justification uses probability-generating functions: the generating function of \(N_j\) is \(\exp\{\lambda_j(z-1)\}\). Independence makes the generating function of the sum the product of these functions, which is \(\exp\{(\sum_j\lambda_j)(z-1)\}\), the generating function of a Poisson count with parameter \(\sum_j\lambda_j\). Consequently, the total has mean and variance both equal to \(\sum_j\lambda_j\).

Independence matters. Even when individual counts have Poisson distributions, dependence between them can change the variance of their sum. In that case, the variance includes covariance terms and need not equal the total mean.

Exposure and the Count Parameter

In an exposure-based Poisson model, a constant claim rate \(r\) per exposure unit over total exposure \(e\) gives a count parameter \(\lambda=re\). This form assumes that the rate is stable over the relevant exposure and that the count process satisfies the Poisson assumptions. The mean and variance of the total count are then both \(re\). The square-root result applies to the total expected count, not directly to the rate \(r\).

If exposure is divided into disjoint periods or groups with independent Poisson counts, the counts can be summed by adding their parameters. This is useful for calculating the fluctuation in a combined count. It does not mean that all groups have the same underlying risk rate; it means their independent counts have parameters that add.

How the Equality Supports Limited-Fluctuation Standards

A count-based limited-fluctuation assessment compares a count’s random variation with a required precision. For absolute precision, the relevant Poisson standard deviation is \(\sqrt{\lambda}\) claims. For relative precision, the corresponding model quantity is \(1/\sqrt{\lambda}\). The equality of mean and variance supplies these expressions without requiring a separate variance parameter.

The equality is an input, not a complete credibility criterion. A probability requirement also specifies how much probability must lie within the chosen precision band, and the appropriate calculation depends on the criterion being used. The standard is found by applying that criterion to the Poisson count’s fluctuation. In particular, one standard deviation is not automatically a 95% interval, and the mean-variance equality alone does not determine a probability level.

When the Poisson Equality May Fail

The Poisson model represents counts with a fixed intensity and independent occurrence increments. In a simple claim-count setting, this means that, conditional on the modeled rate and exposure, claim occurrences do not become more or less likely merely because other claims have occurred. When these assumptions are reasonable, the model provides a useful baseline for fluctuation calculations.

Several features can produce a variance different from the mean. If risks have different underlying claim rates and that heterogeneity is not accounted for, the pooled counts may vary more than a single fixed-rate Poisson model predicts. Dependence, clustering, or a rate that changes over time can also alter the variance. Conversely, constraints or dependence that suppress variation can produce a variance below the Poisson value. In these circumstances, using \(\sqrt{\lambda}\) without checking the model may understate or overstate fluctuation.

For exam work, keep the distinction clear: a stated Poisson model licenses the equality; a count that merely describes claims does not. If the problem supplies a different count model or indicates overdispersion, do not replace it with the Poisson variance by habit. The assumption is especially important when a credibility conclusion depends on how much random variation is expected.

Common Exam Trap