Module 3: Distributions · Lesson 3.3 · about 20 min
The Poisson: orders per hour and rare events
The Poisson distribution counts events in a fixed window of time or space: orders per hour, calls per minute, defects per metre. Its one parameter λ, why mean equals variance, and how to staff for busy hours.
Module 3 · about 20 minutes · school maths. You’ll be able to recognise a Poisson situation, compute Poisson probabilities, scale λ to a window, and spot when counts are more spread out than Poisson.
In everyday terms: raindrops on a paving stone: you know about how many land per minute on average, but each minute the count is a bit different, and you cannot know which minute gets more.
1. Counts with no fixed number of tries
How many orders arrive in the next hour? Calls in the next 10 minutes? Website errors today? There’s no “number of tries”, just events arriving at some average rate. When events
- happen one at a time,
- independently of each other,
- at a steady average rate over the window,
the count in a fixed window follows a Poisson distribution with parameter (lambda), the average count per window:
2. One number does everything
The Poisson’s mean and variance are both :
So a store averaging 100 orders an hour usually sees 100 ± 10 (), while one averaging 4 sees 4 ± 2: relatively much bumpier. Small counts are noisy, which is why a small branch’s daily numbers jump around so much.
scales with the window: 6 orders an hour is 1 per 10 minutes, or 48 in an 8-hour day.
Set λ to 8 and the team’s limit to 12: about 1 hour in 16 is over the limit. Over a 10-hour day about half of all days have at least one busy hour, even though the average is far below 12. Staffing to the average leaves you short in exactly the hours customers notice.
3. Checking the fit: overdispersion
Real counts often vary more than Poisson, because the rate isn’t steady (lunch rush, weekends, campaigns) or events come in clusters (one customer files three tickets). Compare the sample variance with the mean:
- variance ≈ mean: Poisson is a fair model;
- variance ≫ mean: overdispersion. Model hours separately (λ for 12–1 pm, λ for 3–4 pm), or use a wider distribution (the negative binomial).
4. Link to the binomial
A binomial with large and small looks like a Poisson with : 10,000 visitors each with a 0.03 % chance of reporting a bug is close to Poisson with λ = 3. This “law of rare events” is why the Poisson fits so many rare counts.
- λ
- The average count in the window (2 urgent tickets per hour).
- P(X = k) = e^(−λ) λ^k ÷ k!
- The chance of exactly k events in one window.
- Mean = variance = λ
- The SD is √λ, so busier windows wobble more in counts but less as a share.
- Physical meaning
- Poisson is what independent, steady arrivals look like: no memory, no planning. If your counts spread out much more than λ (overdispersion), arrivals are clumping, such as lunch rushes or one customer placing many orders.
“An average of 2 an hour means 2 every hour.” With λ = 2, about 14 % of hours have none at all and about 14 % have 4 or more. Staffing for the average leaves you short one hour in seven.
Watch it happen, one step at a time:
Maths and principle
Principle: chop the window into a huge number of tiny slices, each with a tiny chance of one event. The count is then binomial with enormous and tiny , which tends to the Poisson formula.
Equations: , . P(no events) = .
In code: Excel POISSON.DIST(k, λ, FALSE) (exactly) or TRUE (k or fewer); Python scipy.stats.poisson.pmf(k, lam).
Try it: λ = 3. P(0) = ≈ 0.050; P(1) = 3 ≈ 0.149; P(2) = 9/2 ≈ 0.224.
5. Worked example
Made-up numbers. A help desk gets on average 2 urgent tickets per hour. One person can handle up to 3 an hour.
- P(0) = ≈ 0.135; P(1) = 2 ≈ 0.271; P(2) = 4/2 ≈ 0.271; P(3) = 8/6 ≈ 0.180.
- P(X ≤ 3) ≈ 0.135 + 0.271 + 0.271 + 0.180 = 0.857.
- P(more than 3) ≈ 14.3 % of hours, about 1 hour in 7.
- Over a 40-hour week, expect about 40 × 0.143 ≈ 6 overloaded hours.
If the target is “fewer than 1 overloaded hour a week”, one person isn’t enough at peak times, though the average (2 an hour) looks comfortable.
6. Practice
- Errors average 0.5 a day. What is the chance of an error-free day?
- Calls average 90 an hour. About what range covers most hours (± 2 SD)?
- Daily complaints: mean 3.1, variance 3.0. Poisson or overdispersed?
Answers
- ≈ 0.607, about 61 %.
- SD = √90 ≈ 9.5, so about 71 to 109.
- Variance ≈ mean, so a Poisson model is reasonable.
Try it on your own data: the probability distribution explorer draws the Poisson for any λ and gives exact tail probabilities.
New words: Poisson distribution, rate λ, overdispersion, law of rare events.
- Poisson counts events in a fixed window at a steady average rate λ.
- P(X = k) = e^(−λ) λ^k ÷ k!; mean and variance are both λ.
- Scale λ with the window: 2 per hour is 16 per 8-hour shift.
- Variance far above the mean (overdispersion) means clumped events, so Poisson underestimates busy periods.
Check yourself
Sources
- OpenIntro Statistics, 4th edition (Diez, Cetinkaya-Rundel, Barr), Section 4.5 Poisson distribution https://www.openintro.org/book/os/
- NIST/SEMATECH e-Handbook, 1.3.6.6.19 Poisson Distribution https://www.itl.nist.gov/div898/handbook/eda/section3/eda366j.htm
Your ticks stay in this browser. Save them as a file on the course home.
Updated 2026-11-22. Numbers in the examples are made up to show the idea unless a source is given. Simulations run in your browser; nothing you type is sent anywhere.