Choosing Between Binomial And Poisson Counts

Why the Source of a Count Matters

A count of claims can arise in different ways. A portfolio might contain a fixed number of policies, each of which can have at most one claim during a specified period. Alternatively, claims might occur over a period of exposure at a stated average rate, with no fixed maximum count. These situations can produce similar expected counts, but the assumptions that justify their probability models are different.

The key selection question is not simply, “What is the expected number of claims?” Ask instead what creates each possible claim: a fixed set of opportunities, or events occurring over exposure? Then check whether the assumptions of the chosen model are reasonable for the scenario.

Definition: A count $N$ has a binomial distribution with parameters $n$ and $p$, written $N\sim\operatorname{Binomial}(n,p)$, when it counts successes in $n$ independent trials, each with the same success probability $p$. Its possible values are $0,1,\ldots,n$, and $$ P(N=k)=\binom{n}{k}p^k(1-p)^{n-k},\qquad k=0,1,\ldots,n. $$ A count $N$ has a Poisson distribution with parameter $\lambda$, written $N\sim\operatorname{Poisson}(\lambda)$, when it models events over a specified exposure under a constant-rate Poisson process. Its possible values are all nonnegative integers, and $$ P(N=k)=e^{-\lambda}\frac{\lambda^k}{k!},\qquad k=0,1,2,\ldots. $$

How to Select the Model

Use a binomial model when the count is the number of successes among a known, fixed number $n$ of opportunities. For example, the opportunities could be a fixed set of policies, each classified as having at least one claim or having none during the observation period. The model requires that each opportunity has the same probability $p$ of success and that the outcomes are independent. In this setup the count cannot exceed $n$.

Use a Poisson model when events are counted over a stated amount of exposure and the scenario supports a constant event rate. A standard Poisson-process model assumes events occur at a constant rate over the exposure and counts over disjoint exposure intervals are independent. If the rate is $\lambda$ events per exposure unit and the observation covers $t$ exposure units, then the count has parameter $\lambda t$. There is no fixed upper limit on the count.

Neither model follows from the expected count alone. A binomial count has mean $np$; a Poisson count has mean $\lambda$. Equal means do not make the models interchangeable: their underlying assumptions and possible values still matter. In particular, if a question describes a fixed number of one-claim-or-no-claim opportunities, a count greater than $n$ is impossible, even if a Poisson model with the same mean assigns it a small probability.

Question to ask	Binomial indication	Poisson indication
What creates the count?	A fixed number of trials or opportunities	Events over a stated exposure
What assumptions are central?	Independent trials with a common success probability	Constant rate and independent counts over disjoint exposure intervals
What is the maximum count?	The fixed number of trials, $n$	No fixed maximum
How does time or exposure enter?	It defines the period for each trial’s outcome	The rate is multiplied by the exposure length

Worked Example: Claims Among a Fixed Set of Policies

A hypothetical administrator reviews 12 independent short-term policies. For each policy, the probability of at least one claim during the review period is 0.08. Treat each policy as one claim/no-claim trial. Find the probability that exactly two policies have a claim.

Method and assumptions. There are exactly $n=12$ opportunities, the outcomes are independent, and every policy has the same claim probability $p=0.08$. The count $N$ of policies with a claim is therefore binomial. Its support is $0$ through $12$.

Substitute into the point-probability formula. For $k=2$, there are $\binom{12}{2}=12\cdot11/(2\cdot1)=66$ ways to select the two policies. The probability of a claim on each selected policy is $0.08^2=0.0064$, and the probability of no claim on each of the other ten is $0.92^{10}=0.4343884544$.

P(N=2)=\binom{12}{2}(0.08)^2(0.92)^{10} =66(0.0064)(0.4343884544) =0.1834857\ldots

The probability is 0.1835 to four decimal places. The expected count is $12(0.08)=0.96$, but the model choice comes from the 12 independent claim/no-claim opportunities, not from that mean. A common mistake is to treat the count as Poisson just because it counts claims; that discards the stated fixed-trial structure.

Worked Example: Converting a Claim Rate to a Four-Month Count

A hypothetical book of business is modeled as generating claims at a constant rate of 1.5 claims per month. Under a Poisson-process assumption, find the probability of at least two claims during four months of exposure.

Method and assumptions. The scenario specifies a rate per month, and the model assumes a constant rate over the period with independent counts over disjoint intervals. The exposure is four months, so the Poisson parameter is $\lambda=1.5(4)=6$. The count has no fixed upper limit.

The event “at least two” is the complement of zero or one claim. The Poisson probabilities for those two counts are $e^{-6}$ and $6e^{-6}$, respectively. Thus:

P(N\geq 2)=1-P(N=0)-P(N=1) =1-e^{-6}-6e^{-6} =1-7e^{-6}.

Using $e^{-6}=0.002478752$, we obtain $7e^{-6}=7(0.002478752)=0.017351264$, and hence $P(N\geq2)=1-0.017351264=0.982648736$. The answer is 0.9826 to four decimal places.

The common mistake is to use $\lambda=1.5$ for the full four-month period. That parameter describes one month of exposure; the four-month parameter is $1.5\cdot4=6$. The result is checked by noting that the probability of at least two claims must be close to one when the expected count is six.

When a Poisson Count Approximates a Binomial Count

A Poisson model can sometimes approximate a binomial count when the number of opportunities is large, the success probability for each opportunity is small, and the product $np$ is a suitable moderate value. This is an approximation, not a change to the original trial structure. The binomial model remains the exact model when its assumptions are given.

The reason for the approximation can be seen directly from the binomial probability for a fixed count $k$. Set $p=\lambda/n$, so the binomial mean $np$ equals $\lambda$. Then:

P(N=k) =\binom{n}{k}\left(\frac{\lambda}{n}\right)^k \left(1-\frac{\lambda}{n}\right)^{n-k} = \frac{\lambda^k}{k!} \left(\prod_{j=0}^{k-1}\left(1-\frac{j}{n}\right)\right) \left(1-\frac{\lambda}{n}\right)^n \left(1-\frac{\lambda}{n}\right)^{-k}.

For fixed $k$ and $\lambda$, as $n$ grows, each factor in the product tends to 1, $(1-\lambda/n)^n$ tends to $e^{-\lambda}$, and $(1-\lambda/n)^{-k}$ tends to 1. The expression therefore tends to $e^{-\lambda}\lambda^k/k!$, the Poisson probability. This explains the approximation; it does not justify using it when $n$ is small or $p$ is not small.

Worked Example: Exact Fixed-Opportunity Model and Poisson Approximation

A hypothetical device program monitors 80 independent devices. For each device, the probability of at least one incident during the period is 0.025. The quantity of interest is the number of devices with an incident, not the total number of incidents. Find the exact probability that no device has an incident, then compare it with the Poisson approximation having the same mean.

Exact model. Each of the 80 devices contributes one binary outcome: incident or no incident. With a common probability $p=0.025$ and independence, the number of affected devices $N$ is binomial with $n=80$. Its mean is $np=80(0.025)=2$. The exact probability of no affected devices is:

P(N=0)=(1-0.025)^{80}=0.975^{80} =0.1319\ldots

Thus the exact probability is 0.1319 to four decimal places. For a Poisson approximation, use $\lambda=np=2$, giving $P(N=0)\approx e^{-2}=0.135335\ldots$, or 0.1353 to four decimal places. The approximation is higher by about $0.1353-0.1319=0.0034$.

The method is to select the binomial model as exact, then use a Poisson model only as an approximation because there are many opportunities with a relatively small probability. A common mistake is to report $e^{-2}$ as the exact answer merely because the mean is 2. The exact calculation follows from the 80 independent device outcomes; the Poisson value is an approximation to it.

Selection Checks and Common Traps

Before calculating a probability, make the model choice explicit. State what one trial or unit of exposure means, identify the relevant parameter, and verify that the count support matches the situation. This short check often prevents a longer calculation under the wrong model.

Identify the count being measured.
Distinguish, for example, the number of policies with claims from the total number of claim events. They need not be the same count.

Look for a fixed number of opportunities.
If there are $n$ independent opportunities, each with the same probability $p$ of success, use a binomial model for the success count.

Look for a rate and exposure.
If a constant rate applies over exposure $t$, set the Poisson parameter to rate multiplied by exposure. Keep the time units consistent.

Check the support and the exactness.
A binomial count cannot exceed $n$. A Poisson approximation may be useful under suitable conditions, but it does not make the Poisson model exact.

A further trap is assuming that “independent” alone determines the model. Independence is important for both frameworks, but a binomial model also needs a fixed number of trials with a common success probability. A Poisson-process model instead uses a rate over exposure and its process assumptions. If the scenario does not support the needed assumptions, the correct response is not to force one of these models simply because a formula is familiar.

Key takeaway: Choose the binomial model for a fixed number of independent, equally likely success opportunities. Choose the Poisson model for events over exposure when a constant-rate Poisson process is appropriate. Treat a Poisson calculation based on $np$ as an approximation unless the scenario itself justifies a Poisson model.

Check Your Understanding

For each scenario, select the model and calculate the requested probability. Give probabilities to four decimal places.

A hypothetical insurer studies 25 independent policies, each with claim probability 0.04 during a stated year. Find the probability of exactly one policy having a claim. (Answer: Binomial; $0.3754$)
A hypothetical portfolio has a constant modeled claim rate of 0.7 claims per week. Find the probability of exactly two claims over six weeks. (Answer: Poisson with $\lambda=4.2$; $0.1323$)
There are 150 independent accounts, each with probability 0.01 of at least one claim in a period. Find the exact probability that no account has a claim. (Answer: Binomial; $0.2215$)
A hypothetical system is exposed for 30,000 machine-hours, with a constant event rate of 0.00008 per machine-hour. Find the probability of at least one event. (Answer: Poisson with $\lambda=2.4$; $0.9093$)
For each of two independent years, one policy has probability 0.15 of at least one claim. If the count is the number of years with a claim, find the probability of claims in both years. (Answer: Binomial with $n=2$, $p=0.15$; $0.0225$)

What You'll Learn