Understanding the Poisson Distribution

The Poisson distribution belongs to the family of discrete probability distributions. It answers: given a known average event frequency, what is the probability of observing exactly k events in the next interval?

Consider a call centre receiving an average of 8 calls per hour. What is the probability of exactly 5 calls in the next hour? Or at least 10? The Poisson distribution handles both scenarios elegantly.

This distribution applies when:

  • Events are independent (one occurrence does not influence another)
  • The average rate remains constant across intervals
  • Two events cannot occur at exactly the same instant

Classic examples include website traffic per minute, insurance claims per month, typos per manuscript page, and equipment breakdowns per year.

The Poisson Formula

The probability of exactly x events occurring when the average rate is λ follows this formula:

P(X = x) = (e−λ × λx) ÷ x!

P(X < x) = Σ P(X = k) for k = 0 to x−1

P(X ≤ x) = Σ P(X = k) for k = 0 to x

P(X > x) = 1 − P(X ≤ x)

P(X ≥ x) = 1 − P(X < x)

  • λ (lambda) — The average number of events per interval; the rate parameter
  • x — The exact number of occurrences you want to calculate probability for
  • e — Euler's number, approximately 2.71828
  • x! — Factorial of x (e.g., 5! = 5 × 4 × 3 × 2 × 1 = 120)

Key Properties and Characteristics

The Poisson distribution has elegant mathematical properties that make it powerful for modelling rare events:

  • Mean: Equal to λ. If you average many observations, you get λ.
  • Variance: Also equal to λ. This unique property (mean = variance) distinguishes Poisson from other distributions.
  • Standard deviation: The square root of λ, or √λ.

As λ increases, the distribution becomes more symmetric and bell-shaped, approaching the normal distribution. For λ = 1, the distribution is heavily skewed right; for λ ≥ 10, it approximates normality.

The Poisson distribution is the limiting case of the binomial distribution when the number of trials is very large and the success probability is very small, but their product (n × p) remains fixed.

Common Pitfalls and Practical Guidance

Avoid these mistakes when applying Poisson analysis:

  1. Non-independence assumptions — The Poisson assumes each event is independent. If bus arrivals bunch together (dependent events), or disease cases cluster geographically, the distribution breaks down. Verify independence before analysis.
  2. Changing rates over time — λ must be constant. Traffic intensity differs between rush hour and midnight. Internet traffic varies by season. Always define your time interval carefully and confirm the rate is stable within it.
  3. Integer-only outcomes — Poisson only works for whole-number counts: 0, 1, 2, 3... You cannot have 2.7 defects or 5.3 calls. If your data contains decimals, a different distribution (like gamma or exponential) is appropriate.
  4. Zero-event impossibility — Avoid using Poisson when zero events are impossible by definition. For example, if you are tracking 'number of insurance claims per policy holder' and every policyholder must file at least one claim, the Poisson model is unsuitable.

Practical Worked Example

Suppose a manufacturing plant experiences an average of 2.4 equipment failures per month. What is the probability of exactly 3 failures next month?

Given:

  • λ = 2.4 (average failures per month)
  • x = 3 (desired number of failures)

Calculation:

P(X = 3) = (e−2.4 × 2.43) ÷ 3!

= (0.0907 × 13.824) ÷ 6

= 1.254 ÷ 6

≈ 0.209 or 20.9%

The calculator also instantly shows cumulative probabilities: the probability of at most 3 failures (35.8%), fewer than 3 failures (14.9%), more than 3 failures (64.2%), and at least 3 failures (85.1%). Maintenance teams can use this to plan spare-parts inventory and downtime budgets.

Frequently Asked Questions

When is the Poisson distribution the right choice instead of binomial?

Use Poisson when you know the average number of events (λ) but not the total number of trials, or when n is very large and p is very small. Binomial requires both parameters: number of trials (n) and success probability per trial (p). For instance, if you count meteor impacts over 100 years, Poisson fits better than binomial because defining 'total possible meteors' is impractical. Poisson shines for rare, continuously occurring events.

Why must λ and x be non-negative integers?

The Poisson distribution models discrete counts—whole numbers of events. You cannot have 3.7 website visitors or 1.2 system crashes. Additionally, negative values are meaningless: you cannot experience −5 phone calls. The mathematics itself, involving factorials and exponentials, requires non-negative integers. If your raw data includes decimals, consider rounding or switching to a continuous distribution like gamma.

How does Poisson relate to the normal distribution?

As λ grows larger, the Poisson distribution shape approaches a normal bell curve. This is why some statisticians use a normal approximation to simplify calculations when λ > 10. However, the normal approximation introduces slight error, and the Poisson remains discrete while the normal is continuous. Modern calculators make exact Poisson computation effortless, so the approximation is rarely necessary today unless working without computational tools.

Can I use Poisson if events happen in clusters or are dependent?

No. Poisson fundamentally assumes independence: each event is unrelated to others. If events cluster (like customer arrivals spiking during sales) or correlate (disease spreading through close contact), the Poisson model fails. Consider alternatives like negative binomial distribution, which handles overdispersion when variance exceeds the mean, or Hawkes processes for self-exciting events.

What do the cumulative probabilities tell me that exact probability does not?

Exact probability P(X = x) answers: what is the chance of <em>exactly</em> x events? Cumulative probabilities answer broader questions: what is the chance of <em>at most</em> x events, or <em>at least</em> x events? For decision-making, cumulative probabilities often matter more. A warehouse manager cares about P(X ≥ 50 damaged items) to set acceptable damage thresholds, not P(X = 47).

How do I choose the right time interval for my rate parameter λ?

Your interval definition must match your question and be consistent with historical data. If historical records show 12 equipment failures per year, you can express λ = 12 (per year), λ = 1 (per month), or λ ≈ 0.23 (per week). Shorter intervals yield smaller λ values; longer intervals yield larger λ values. Choose whichever interval makes sense for your application and decision-making timeline.

More statistics calculators (see all)