Poisson Distribution: Formula, Worked Examples, and Real-World Uses

How many customers walk into a store per hour? How many typos appear on a page? How many earthquakes strike a region per year? These questions share a common shape: counting how many times a rare, random event happens over a fixed interval of time or space, when the average rate is known but the exact timing isn't. The Poisson distribution is the standard tool for exactly this kind of question.

The Conditions

The Poisson distribution belongs to the same family of discrete distributions as the binomial distribution, but answers a subtly different kind of question — counting rare events over a continuous interval rather than counting successes across a fixed number of discrete trials.

A situation follows a Poisson distribution when: events occur independently of each other, the average rate (λ, "lambda") is constant over the interval being studied, and two events can't occur at literally the same instant. Unlike the binomial distribution, there's no fixed "number of trials" — instead there's a continuous opportunity for an event to occur across the interval, with a known average count.

The Formula

Here λ is the average number of events in the interval, k is the specific count you're finding the probability for, and e is Euler's number (≈2.71828). The formula depends only on λ and k — no separate probability-per-trial parameter is needed, unlike the binomial distribution.

Worked Example

Problem: A call centre receives an average of 4 calls per minute. What's the probability of receiving exactly 6 calls in a given minute?

λ = 4, k = 6

P(X=6) = (4⁶ × e⁻⁴) / 6! = (4096 × 0.01832) / 720 = 75.03 / 720 ≈ 0.1042

There's roughly a 10.4% chance of receiving exactly 6 calls in a given minute, given the average rate of 4 per minute.

Mean and Variance Are Always Equal

This is one of the Poisson distribution's most distinctive properties — unlike most other distributions, its mean and variance are the exact same number, λ. This makes it easy to spot data that plausibly follows a Poisson distribution: if a count variable's sample variance is close to its sample mean, Poisson is a reasonable candidate model. If the variance is substantially larger than the mean ("overdispersion," common with real-world count data that has more variability than a pure Poisson process would predict), a different or extended model is usually more appropriate.

Cumulative and "At Least" Probabilities

As with the binomial distribution, cumulative probabilities are found by summing the exact-k formula across the relevant range, and the complement rule often shortcuts an "at least" calculation:

Example: what's the probability of at least 1 call in a given minute (same λ=4)? P(X=0) = (4⁰ × e⁻⁴)/0! = e⁻⁴ ≈ 0.0183. P(at least 1) = 1 − 0.0183 ≈ 0.9817 — with an average rate of 4 per minute, it's overwhelmingly likely at least one call arrives in any given minute.

Relationship to the Binomial Distribution

The Poisson distribution is what the binomial distribution converges to as the number of trials n grows very large while the probability of success p shrinks very small, in such a way that np stays constant (equal to λ). This is genuinely useful in practice: when modelling rare events across a huge number of opportunities (like defects across millions of manufactured units, each individually unlikely to be defective), the Poisson distribution is often a more convenient and equally accurate approximation to what is, technically, an underlying binomial process.

The Poisson distribution's discovery predates modern statistics by over a century — Siméon Denis Poisson introduced it in 1837 while studying wrongful convictions in French military courts, and it later found one of its most famous early applications in a study of Prussian cavalry deaths from horse kicks, a genuinely rare, independent, constant-rate event that fit the model almost perfectly. That historical pedigree is part of why the distribution remains a default starting point whenever "rare events over an interval" describes the data at hand.

Real-World Uses

Call centres and queueing: the arrival rate of customer calls or website visitors is a textbook Poisson application, directly informing staffing decisions and queue management models.

Manufacturing defects: the number of defects per unit area of material, or per batch of products, is often modelled as Poisson, informing quality control sampling plans.

Epidemiology: the number of disease cases occurring in a population over a given time period is frequently modelled with Poisson-based methods, particularly for rare diseases.

Sports analytics: the number of goals scored in a football match is commonly modelled as approximately Poisson, underlying many betting-market and match-prediction models — even though real matches have some departures from the pure independence assumption (a team leading late often plays more defensively, changing the effective scoring rate).

Natural events: earthquakes, radioactive decay events, and traffic accidents at a specific intersection are all classic textbook examples of processes reasonably modelled as Poisson, when the rate is roughly constant over the period studied.

A Second Worked Example: At-Most Probability

Problem: A website averages 2.5 server errors per hour. What's the probability of 2 or fewer errors in a given hour?

λ = 2.5. Need P(X≤2) = P(X=0) + P(X=1) + P(X=2).

P(X=0) = (2.5⁰ × e⁻²·⁵)/0! = e⁻²·⁵ ≈ 0.0821

P(X=1) = (2.5¹ × e⁻²·⁵)/1! = 2.5 × 0.0821 ≈ 0.2052

P(X=2) = (2.5² × e⁻²·⁵)/2! = (6.25 × 0.0821)/2 ≈ 0.2565

P(X≤2) = 0.0821 + 0.2052 + 0.2565 ≈ 0.5438

There's roughly a 54% chance of seeing 2 or fewer errors in a given hour — a useful baseline for setting alerting thresholds: an hour with, say, 8 errors would be a genuinely rare event under this model, worth investigating rather than dismissing as normal variation.

Common Mistakes

Using the wrong λ for the interval in question. As covered above, this is the single most common Poisson calculation error — always match the rate parameter to the specific interval (per minute, per hour, per day) the question is actually asking about, scaling linearly as needed.

Applying Poisson to events that aren't actually independent. If one event genuinely makes a subsequent event more or less likely (like a single fault triggering a cascade of related errors), the independence assumption is violated, and the true variance will exceed what the Poisson model predicts (overdispersion) — a different model, like the negative binomial distribution, is often more appropriate in that case.

Assuming a Poisson model fits without checking mean ≈ variance. Before trusting Poisson-based predictions for real data, it's worth checking that the sample mean and sample variance of observed counts are reasonably close — a large mismatch is a signal the constant-rate, independent-events assumptions may not genuinely hold for this specific process.

Calculating Poisson Probabilities in Practice

Like the binomial distribution, Poisson probabilities are almost always calculated with software for anything beyond the simplest cases — statistical packages provide direct functions for exact and cumulative Poisson probabilities, computed using the same underlying formula shown here but without the manual arithmetic. Understanding the formula well enough to interpret and sanity-check that output remains valuable even when a tool does the actual calculation.