The short answer: if events occur independently at an average rate λ per interval, the probability of exactly k events is P(X = k) = e^(−λ) × λ^k ÷ k!. At 4 calls per hour, exactly 6 calls have probability 10.42%, and at least one call has probability 98.17%. The mean and variance both equal λ.

Key takeaways

  • λ is a rate per interval, not a raw count. The interval must match the rate.
  • Mean equals variance. If the sample variance is much larger than the mean, the model is overdispersed.
  • Rates scale linearly with exposure. Four per hour becomes two per half-hour.
  • Rare-event counts converge to Poisson. It is the limit of a binomial with large n and small p.
  • “At least one” is a complement: 1 − e^(−λ).
  • Clustering and trends break the constant-rate assumption, even when the mean looks right.

What the Poisson model describes

The Poisson distribution counts how many events occur in a fixed interval of time, length, area, or volume when the events happen independently and at a constant average rate. Examples include calls arriving at a help desk in an hour, typos on a page, radioactive decays in a minute, and defects along a length of cable.

The model needs four conditions: the count is of discrete events, events do not occur simultaneously, each event is independent of the others, and the average rate is constant across the interval. It describes counts, not the waiting time between events; that is the exponential distribution. It also does not apply when events arrive in clusters, such as orders placed during a flash sale.

The formula and its terms

For a Poisson random variable with rate λ over one interval:

P(X = k) = e^(−λ) × λ^k ÷ k!
Mean = λ, Variance = λ, Standard deviation = √λ

The term e^(−λ) normalizes the probabilities so they sum to 1. λ^k ÷ k! weighs larger counts: the numerator grows with k, but the factorial in the denominator eventually dominates, which is why the distribution has a right tail. The mode is near λ: for integer λ, both λ − 1 and λ are most likely, so a rate of 4 makes 3 and 4 tied for the peak.

Probabilities are computed for a defined interval. If the rate is 4 per hour, then λ = 4 for a one-hour count, λ = 2 for a half-hour count, and λ = 0.5 for a 7.5-minute count. Mixing a per-hour rate with a per-minute interval is the most common arithmetic error.

Worked example: 4 calls per hour

A support line receives an average of 4 calls per hour. For exactly 6 calls: e^(−4) = 0.0183156, 4^6 = 4,096, and 6! = 720. So P(X = 6) = 0.0183156 × 4,096 ÷ 720 = 0.104196, or 10.42%.

The probability of no calls is e^(−4) = 0.0183156, or 1.83%. Exactly 1 call is 4 × e^(−4) = 7.33%, and exactly 2 is 8 × e^(−4) = 14.65%. The probabilities rise to a plateau at 3 and 4 (19.54% each) and then decline, but the tail never reaches zero: exactly 10 calls still has probability 0.53%.

Cumulative counts and the complement

“At most 2 calls” means P(0) + P(1) + P(2) = e^(−4) × (1 + 4 + 8) = 0.238103, or 23.81%. “At least 1 call” is the complement 1 − P(0) = 1 − 0.0183156 = 0.981684, or 98.17%.

The complement is especially useful for “at least one” questions with small rates. If a server fails on average once every 10,000 hours, λ = 0.0001 per hour and the chance of at least one failure in an hour is 1 − e^(−0.0001) ≈ 0.0001, essentially the rate itself. For rare events, probability and expected count nearly coincide.

For a range, subtract cumulative probabilities. P(2 ≤ X ≤ 5) = P(X ≤ 5) − P(X ≤ 1), which avoids adding four separate terms and reduces rounding drift.

Changing the interval

Poisson rates are proportional to exposure. If calls arrive at 4 per hour, the expected count in 30 minutes is λ = 2, and the probability of no calls is e^(−2) = 13.53%. In 15 minutes λ = 1, so P(0) = 36.79% and P(3) = e^(−1) ÷ 6 = 6.13%.

Scale the rate before computing, then state the interval with the answer. “The probability of 3 calls is 6.13%” is incomplete; “the probability of exactly 3 calls in the next 15 minutes is 6.13%” is testable. The same logic applies to defects per metre, accidents per 1,000 vehicle-hours, or particles per litre.

Mean, variance, and overdispersion

The Poisson model forces the variance to equal the mean. That is a strong claim. For 4 calls per hour, the standard deviation is √4 = 2, so counts of 2 to 6 are ordinary and counts below 0 or above 10 are rare.

Real count data often show variance far above the mean. If the mean count is 4 but the sample variance is 9, the data are overdispersed: some hours are busy for reasons shared by all calls, such as a promotion or an outage. A Poisson model will then produce intervals that are too narrow. Quasi-Poisson and negative binomial models add a dispersion parameter and are the standard remedies.

Poisson as a limit of the binomial

When n is large and p is small but λ = n × p stays moderate, the binomial distribution converges to the Poisson. A process with 1,000 opportunities and a 0.3% chance each has λ = 3; the binomial probability of zero events is 0.997^1,000 ≈ 0.0496, while the Poisson value is e^(−3) ≈ 0.0498.

The approximation is useful because it replaces two parameters with one and avoids huge factorials. It is weakest when p is not small or when n × p is large; in those cases use the binomial directly. The Poisson limit also explains why the model appears whenever there are many opportunities for a rare event, from insurance claims to data-entry errors.

If counts follow a Poisson process with rate λ, the time between events follows an exponential distribution with mean 1 ÷ λ, and the time until the r-th event follows a gamma distribution. For 4 calls per hour, the average gap between calls is 15 minutes, and the chance that the next call arrives within 5 minutes is 1 − e^(−4 × 5/60) = 28.3%.

When the rate itself varies randomly, a gamma mixture of Poisson counts produces the negative binomial distribution, which is why that model handles overdispersion. When events are contagious — one arrival makes another more likely — the Poisson assumption fails and a self-exciting or Hawkes model may be needed.

Common mistakes

  • Using a count instead of a rate for λ. λ must be the expected count in the interval you are asking about.
  • Mismatching rate and interval. A per-hour rate with a per-minute question gives an answer that is off by a factor of 60.
  • Adding rates for overlapping intervals. Rates add only for independent, non-overlapping segments; averaging them is not the same as pooling counts.
  • Assuming variance equals mean in real data. Check the sample variance before trusting Poisson intervals.
  • Forgetting the k! denominator. λ^k alone is not a probability.
  • Applying Poisson to clustered events. Arrivals during a sale or an outage are not independent.

Practical caveats

Constant rate is an idealization. Call volumes peak at certain hours, accidents rise in bad weather, and website requests spike during events. If the rate changes within the interval, either shorten the interval until the rate is roughly constant or model the rate as a function of time.

Zero-inflation is another failure mode: more zeros than a Poisson model predicts, often because some units cannot experience the event at all. Conversely, a hard cap on counts — a queue that can hold only 20 people — truncates the distribution. Check the observed frequency of zeros and the maximum count against the model before quoting probabilities.

Frequently asked questions

What is the difference between Poisson and binomial?

Binomial counts successes in a fixed number of trials with a fixed success probability. Poisson counts events in an interval with a fixed rate, which is the limit of binomial when trials are numerous and the per-trial probability is tiny.

Can λ be greater than 1?

Yes, and it usually is. λ is an expected count, so a rate of 4 per hour gives λ = 4. The restriction λ ≤ 1 applies only to probabilities of a single event in a very short interval.

Why does the Poisson distribution have only one parameter?

Because the mean and variance are tied together by the model. One rate λ sets both, which makes the model simple but also easy to reject when real data are overdispersed.

What does it mean if the variance is greater than the mean?

The data are overdispersed relative to Poisson. Clustering, unobserved heterogeneity, or a varying rate is present; use a negative binomial or quasi-Poisson model instead of trusting Poisson standard errors.

How do I find the probability of at least one event?

Use P(X ≥ 1) = 1 − P(X = 0) = 1 − e^(−λ). For λ = 4 this is 98.17%.

Is the Poisson distribution symmetric?

No. It is skewed right, especially for small λ. As λ grows, the shape becomes more symmetric and a normal approximation with mean and variance λ becomes reasonable.

Further reading

Keep exploring

Read next Logarithm: Base, Product Rule, and Scale A logarithm is an exponent in disguise. Learn the definition, the product, quotient and power rules, change of… Read next Order of Operations: PEMDAS and Ambiguous Expressions PEMDAS is a shared convention, not a law of mathematics. Learn the real precedence rules, the exponent and… Read next Significant Figures: Counting, Rounding, and Precision Significant figures record how precisely a value was measured. Learn the counting rules, the addition and multiplication rules,…