Poisson Claims And Its Parameters

Why Model Claim Counts?

A claim-count model describes how many claims occur in a specified period or amount of exposure. That count is different from the total dollars paid: a policy with three small claims and a policy with one large claim have different counts even if their total claim amounts match. In ratemaking, the Poisson distribution provides a basic model for counts such as claims per policy-year, provided the count and exposure period are clearly defined.

The essential parameter is not automatically an annual rate. It is the expected number of events over the period represented by the random variable. If the period changes, the parameter must be adjusted consistently with the exposure and the assumed rate of claim occurrence.

Definition: A random variable $N$ has a Poisson distribution with parameter $\lambda$, written $N\sim\operatorname{Poisson}(\lambda)$, if $\lambda\geq 0$, $N$ takes nonnegative integer values, and, for $k=0,1,2,\ldots$, $$ P(N=k)=e^{-\lambda}\frac{\lambda^k}{k!}. $$ For $\lambda=0$, the count is zero with probability 1. For $\lambda>0$, $\lambda$ is the expected count over the stated period.

Meaning of the Rate Parameter

Suppose a hypothetical portfolio segment has an expected $0.20$ claims per policy-year. For one policy observed for one policy-year, the Poisson parameter is $\lambda=0.20$. For the same risk observed for two policy-years, a constant rate assumption gives $\lambda=0.40$. The number $0.20$ is not a claim probability: a count can be greater than one, and the probability of at least one claim is $1-e^{-0.20}$, not $0.20$.

More generally, if the rate is $r$ expected claims per unit of exposure and the observed exposure is $e$ units, then the mean count is $\lambda=re$. This scaling requires the rate to be applicable throughout the exposure being counted. For a homogeneous Poisson process, claim counts in disjoint periods are independent, and the expected count per unit time is constant. In insurance applications, this is a simplifying assumption about the relevant risk group and period, not a guarantee about actual claim behavior.

For a Poisson random variable, the mean and variance are both $\lambda$. This is a useful defining property of the model and a practical diagnostic: if observed counts vary substantially more than their average would suggest, a Poisson model may not describe them well. The count distribution does not, by itself, specify claim severity, the dollars paid per claim, or the distribution of aggregate losses.

Why the Poisson Probability Formula Holds

One way to justify the formula is to divide a period of length $t$ into $n$ very short intervals. Under a constant rate $r$, each interval has an event probability approximately $rt/n$, with the chance of multiple events in one short interval becoming negligible as the intervals get shorter. The total count is then approximated by a binomial count with $n$ trials and event probability $rt/n$. For fixed $k$, its probability is

\binom{n}{k}\left(\frac{rt}{n}\right)^k \left(1-\frac{rt}{n}\right)^{n-k}.

As $n$ increases, the first factors approach $(rt)^k/k!$, while $(1-rt/n)^n$ approaches $e^{-rt}$. The remaining factor $(1-rt/n)^{-k}$ approaches 1. Thus the limiting probability is $e^{-rt}(rt)^k/k!$, which is the Poisson formula with $\lambda=rt$. This justification depends on the constant-rate and independent-increment assumptions; a time-varying rate requires a suitable adjustment to the mean, and dependence between events can invalidate the model.

The mean also follows directly from the probability formula. For $\lambda>0$,

E[N] =\sum_{k=0}^{\infty} k e^{-\lambda}\frac{\lambda^k}{k!} =e^{-\lambda}\lambda\sum_{j=0}^{\infty}\frac{\lambda^j}{j!} =e^{-\lambda}\lambda e^\lambda =\lambda,

where $j=k-1$ in the sum after the $k=0$ term, which contributes zero. The variance is also $\lambda$, consistent with the Poisson mean-variance property covered in the earlier material on mean, variance, and dispersion.

Calculating Common Count Probabilities

For a specified $\lambda$, substitute the desired count $k$ into the probability mass function. The most common cases are a zero count, an exact count, and a count of at least one:

P(N=0)=e^{-\lambda},\qquad P(N=k)=e^{-\lambda}\frac{\lambda^k}{k!},\qquad P(N\geq 1)=1-e^{-\lambda}.

For a threshold greater than one, it is often simpler to use the complement. For example, $P(N\geq 2)=1-P(N=0)-P(N=1)$. Probabilities in the examples below are rounded to four decimal places only at the final step.

Worked Example: Zero, One, or Multiple Claims

A hypothetical insured risk is modeled as having a Poisson claim count over a one-year policy period with mean $\lambda=2$. Find the probability of no claims, exactly one claim, and at least two claims. Also state the expected count and variance.

Model and assumptions. Let $N$ be the number of claims in the one-year period, with $N\sim\operatorname{Poisson}(2)$. The period is fixed, and the Poisson model is assumed appropriate for this risk. No interest or mortality assumptions are relevant to this claim-count calculation.

No claims. Set $k=0$:

P(N=0)=e^{-2}\frac{2^0}{0!}=e^{-2}=0.135335\ldots

Thus $P(N=0)=0.1353$.

Exactly one claim. Set $k=1$:

P(N=1)=e^{-2}\frac{2^1}{1!}=2e^{-2} =2(0.135335\ldots)=0.270671\ldots

Thus $P(N=1)=0.2707$.

At least two claims. Use the complement of zero or one:

P(N\geq 2)=1-P(N=0)-P(N=1) =1-0.135335\ldots-0.270671\ldots =0.593994\ldots

Thus $P(N\geq 2)=0.5940$. The mean and variance are both $\lambda=2$, so $E[N]=2$ claims and $\operatorname{Var}(N)=2$ claims squared. As a check, the three disjoint probabilities for zero claims, one claim, and at least two claims sum to $0.135335\ldots+0.270671\ldots+0.593994\ldots=1$. The method avoids treating the mean of two claims as a maximum count or as the probability of a claim.

Adjusting for Exposure

When the rate is given per unit of exposure, first determine the exposure represented by the count, then calculate its mean $\lambda$. Do not use a rate for one time unit as the parameter for a different time unit without adjustment.

Worked Example: A Rate Over a Multiyear Policy Period

A hypothetical policy has an expected claim count of $0.30$ per policy-year. Assume a constant rate over a four-year observation period. Find the mean count over the four years, the probability of exactly two claims, and the probability of at least one claim.

Model and assumptions. The count $N$ covers four policy-years of exposure and is modeled as Poisson. The annual expected count is constant, and counts are assumed to follow the Poisson model over the period. No interest or mortality assumptions are relevant.

Step 1: Calculate the period mean. The exposure is $e=4$ policy-years and the rate is $r=0.30$ claims per policy-year:

\lambda=re=(0.30)(4)=1.20.

Therefore $E[N]=1.20$ claims over the four-year period.

Step 2: Calculate exactly two claims. Substitute $k=2$ and $\lambda=1.20$:

P(N=2)=e^{-1.20}\frac{1.20^2}{2!} =(0.301194\ldots)\frac{1.44}{2} =(0.301194\ldots)(0.72) =0.216859\ldots.

Thus $P(N=2)=0.2169$.

Step 3: Calculate at least one claim.

P(N\geq 1)=1-P(N=0) =1-e^{-1.20} =1-0.301194\ldots =0.698805\ldots.

Thus $P(N\geq 1)=0.6988$. The expected count is for four years, not one; using $\lambda=0.30$ in the four-year probability calculation would omit three years of exposure.

Combining Independent Counts

If $N_1$ and $N_2$ are independent Poisson counts with parameters $\lambda_1$ and $\lambda_2$, then their sum is Poisson with parameter $\lambda_1+\lambda_2$. The parameters add because expected counts add. Independence is important: adding the means is valid for expected counts even without independence, but the Poisson distribution for the sum is not generally guaranteed without the appropriate independence assumption.

Theorem: If $N_1\sim\operatorname{Poisson}(\lambda_1)$ and $N_2\sim\operatorname{Poisson}(\lambda_2)$ are independent, then $N_1+N_2\sim\operatorname{Poisson}(\lambda_1+\lambda_2)$. The result extends to any finite collection of independent Poisson counts.

Worked Example: Combining Two Independent Exposure Groups

A hypothetical book has two independent groups of policies. The count for Group A is Poisson with mean $0.60$ claims over the observation period; the count for Group B is Poisson with mean $1.10$. Find the expected total count and the probability of exactly one claim across both groups.

Model and assumptions. Let $N_A$ and $N_B$ be the group counts. They are assumed independent, with $N_A\sim\operatorname{Poisson}(0.60)$ and $N_B\sim\operatorname{Poisson}(1.10)$. There is no interest or mortality component.

Step 1: Combine the parameters. By the independent-sum result, the total $N=N_A+N_B$ is Poisson with

\lambda=0.60+1.10=1.70.

The expected total count is therefore $E[N]=1.70$ claims.

Step 2: Calculate exactly one claim.

P(N=1)=e^{-1.70}\frac{1.70^1}{1!} =(0.182683\ldots)(1.70) =0.310561\ldots.

Thus the probability is $0.3106$. A check is that the total mean $1.70$ equals the sum of the group means, $0.60+1.10$. The method avoids calculating group probabilities separately and then overlooking ways the total can equal one.

Common Mistake: Confusing a Rate With a Period Mean

A rate stated per year is not automatically the Poisson parameter for a shorter or longer policy period. For a constant rate $r$ per year and a period of $t$ years, use $\lambda=rt$. For example, at $0.08$ expected claims per policy-year, a half-year exposure has parameter $0.08(0.5)=0.04$, not $0.08$. The incorrect calculation $P(N=0)=e^{-0.08}=0.9231$ uses a full-year mean; the correct calculation is $P(N=0)=e^{-0.04}=0.9608$.

Worked Example: Half-Year Exposure and a Policy Cohort

A hypothetical policy has an expected $0.08$ claims per policy-year. Assume a constant rate and Poisson claim counts. For one policy observed for six months, find the probabilities of zero claims and at least two claims. Then find the expected number of claims for 500 independent policies, each observed for six months.

Model and assumptions. The rate is $r=0.08$ claims per policy-year. Each policy is exposed for $e=0.5$ years, and policy counts are independent Poisson counts. No interest or mortality assumptions are relevant.

Step 1: Find the individual policy mean.

\lambda=re=(0.08)(0.5)=0.04.

Step 2: Find the probability of zero claims.

P(N=0)=e^{-0.04}=0.960789\ldots,

so the probability is $0.9608$.

Step 3: Find the probability of at least two claims. First calculate the probabilities of zero and one:

P(N=1)=e^{-0.04}\frac{0.04^1}{1!} =(0.960789\ldots)(0.04) =0.0384316\ldots.

Then use the complement:

P(N\geq 2)=1-P(N=0)-P(N=1) =1-0.960789\ldots-0.0384316\ldots =0.0007792\ldots.

Thus $P(N\geq 2)=0.0008$. The small probability is plausible because the mean for one half-year is only $0.04$.

Step 4: Find the cohort expected count. Each of 500 policies has mean $0.04$, so the expected total is

$$ 500(0.04)=20 $$

claims. The independent aggregate count is also Poisson with parameter $20$. The answer avoids using a full-year parameter for a half-year policy; doing so would incorrectly double the individual mean and the cohort expected count.

When the Model Is and Is Not Suitable

The Poisson model is a useful starting point for claim counts when the observation period and exposure are defined, the rate is reasonably stable for the risks being modeled, and event counts are consistent with the model's independence and mean-variance assumptions. It also gives a compact way to calculate probabilities for counts that can be zero, one, two, or more.

It should not be applied mechanically. For example, strong differences in risk among policies, clustering of events, or changing claim conditions may make counts more variable than a Poisson model predicts. Likewise, a count of policies with at least one claim is not the same random variable as the number of claims: a policy with multiple claims contributes one to the first count and multiple units to the second. Confirm what is being counted, the exposure period, and whether the Poisson assumptions are defensible before interpreting a calculated probability.

A final practical check is to ask whether the chosen $\lambda$ matches the time or exposure attached to the random variable. Then verify that the requested event is expressed in count terms and that complementary probabilities cover the intended cases. These checks prevent many avoidable errors in count calculations aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

For each problem, identify the relevant Poisson mean before calculating. Round probabilities to four decimal places.

A hypothetical account has a Poisson count with mean $0.70$ claims in one year. Find the probability of zero claims and state the expected count. (Answer: $P(N=0)=0.4966$; expected count $=0.70$.)
A hypothetical risk has $N\sim\operatorname{Poisson}(2.50)$ over its stated observation period. Find the probability of exactly one claim. (Answer: $0.2052$.)
A hypothetical policy has an expected $0.12$ claims per policy-year and is observed for three and one-half years at a constant rate. Find its mean count and probability of zero claims. (Answer: mean $=0.42$; $P(N=0)=0.6570$.)
Two hypothetical independent groups have Poisson claim counts with means $0.60$ and $1.10$ over the same period. Find the probability that their combined count is exactly one. (Answer: $0.3106$.)
A hypothetical policy has a constant rate of $0.40$ expected claims per policy-year. Find the probability of at least two claims in a half-year, and the expected total claims for 80 independent policies each observed for that half-year. (Answer: individual probability $=0.0175$; cohort expected count $=16$.)

What You'll Learn