Poisson Probability Calculator

Understanding the Poisson Probability Calculator

A Poisson calculator gives the probability of a specific number of events when events occur independently at a constant average rate. Its defining feature is that the mean and variance are equal, and real count data frequently violates that in a way worth detecting.

How it actually works

Enter the average rate and the number of events. The calculator applies the Poisson formula and reports the mean, variance, and standard deviation. A rate of 3 with 2 events gives a probability of 0.224.

The deeper context most people miss

The mean equalling the variance is a strong constraint and a useful diagnostic. If your count data has variance substantially exceeding its mean, the Poisson model is wrong, and the excess spread usually signals clustering or a rate that varies between units.

Why overdispersion is so common

Real count data frequently shows more variability than the Poisson permits, and the reasons are structural rather than exotic. If the underlying rate varies between individuals or periods, the resulting mixture has greater variance than any single Poisson, which is the situation whenever the population is heterogeneous: accident rates differ between drivers, infection rates between individuals, and defect rates between production batches. Mixing Poisson distributions with a gamma-distributed rate gives the negative binomial distribution, which is the standard remedy and has an extra parameter controlling the excess spread. Clustering violates independence directly, since one event making another more likely produces bursts, which describes contagious processes, correlated failures, and anything with feedback. Zero inflation is a related pattern where more zeros appear than the model predicts, arising when a subpopulation cannot experience the event at all, which is why zero-inflated models exist for counts of things like doctor visits or purchases where some individuals are structurally at zero. The practical consequence of ignoring overdispersion is that standard errors are too small and significance is overstated, sometimes dramatically, so results appear far more certain than the data supports. Checking whether the variance exceeds the mean is a cheap diagnostic that should be routine before fitting any Poisson model.

A worked example: rare events and the law of small numbers

A rate of 3 with 2 events gives 0.224, and the distribution's historical name, the law of small numbers, reflects its origin in modelling rare events. Bortkiewicz's 1898 analysis of Prussian cavalry deaths from horse kicks is the classic illustration, showing that a rare event with a low constant rate across many units follows the distribution closely. The same shape describes calls arriving at a switchboard, which is why queueing theory rests on it, and Erlang's work on telephone traffic founded the field. Radioactive decay counts follow it. Typographical errors per page, mutations per genome region, and goals per football match are commonly modelled with it. The distribution arises as the limit of the binomial when the number of trials grows large and the success probability shrinks with their product held constant, which explains why it describes rare events among many opportunities. That derivation also explains the equal mean and variance, since the binomial variance np(1−p) approaches np as p shrinks. A practically important consequence is that Poisson counts are considerably more variable than intuition suggests: with a mean of 3, observing 0 or 7 events is entirely unremarkable, which matters when small count differences between periods or locations are interpreted as meaningful signals.

Deciding whether count differences are meaningful

This is where the distribution does its most valuable practical work, because people routinely over-interpret variation in small counts. A hospital recording 12 infections one year and 18 the next has seen a 50% increase in the headline, and under a constant Poisson rate that difference is well within ordinary variation. The standard deviation of a Poisson count is the square root of its mean, so counts around 15 vary by roughly plus or minus 4 as a matter of course. This underlies the phenomenon of apparent disease clusters, where a locality shows more cases than average purely by chance, and given enough localities examined some will always show excesses, which is the multiple comparisons problem in geographic form. Investigating every apparent cluster consumes resources and generates false alarms, and epidemiological practice has developed methods accounting for the number of areas examined. The same reasoning applies to comparing crime figures between periods, defect counts between batches, and incident rates between teams. The practical test is whether the difference exceeds what Poisson variation alone would produce, and control charts based on the distribution formalise this for process monitoring. Reporting counts without an indication of expected variation invites over-interpretation, which is why rate estimates should carry confidence intervals.

Queueing theory and where the distribution leads

Poisson arrivals combined with exponentially distributed service times produce the simplest queueing model, and its results are counterintuitive and practically important. The exponential distribution describes the gaps between Poisson events and is memoryless, meaning the time until the next event does not depend on how long you have already waited, which is a strong and frequently unrealistic assumption that nonetheless makes the mathematics tractable. The central result is that queue length grows non-linearly with utilisation and approaches infinity as utilisation approaches capacity, so a system running at 90% utilisation has average queues several times longer than one at 70%, and the last few percent of capacity costs disproportionately in waiting time. This is why systems designed to run near full utilisation perform badly under any variability, and why slack capacity is not waste but the price of acceptable response times. It explains hospital emergency departments, road congestion, call centres, and server capacity planning alike. Variability makes it worse: the same average load with more variable arrivals or service times produces longer queues, which is why reducing variability improves performance without adding capacity. Little's law relates average queue length, arrival rate, and waiting time in a form that holds regardless of distribution, and is one of the more useful results in operations because it requires almost no assumptions.

Variations: related distributions and extensions

The binomial converges to the Poisson as trials grow and probability shrinks with the product fixed. The negative binomial handles overdispersion by allowing variance to exceed the mean. Zero-inflated models handle excess zeros from a structurally unexposed subpopulation. The exponential distribution describes waiting times between Poisson events and the gamma distribution the time until several events. Poisson regression models counts against predictors using a log link, and quasi-Poisson or negative binomial regression handles overdispersion, with using plain Poisson regression on overdispersed data being a common and consequential error. Poisson processes generalise to non-constant rates as inhomogeneous processes, and to space as spatial point processes used in ecology and epidemiology. Compound Poisson distributions sum a random number of random amounts and appear in insurance claim modelling. For hypothesis testing, exact Poisson tests compare observed counts against an expected rate, and confidence intervals for rates from small counts should use exact methods rather than normal approximations, which behave poorly when counts are low.

Using the Poisson distribution correctly

Check whether the variance approximately equals the mean in your data, since substantially greater variance means the model is wrong and standard errors will be understated. Use the negative binomial where overdispersion is present, which adds a parameter for the excess spread rather than ignoring it. Consider zero-inflated models where more zeros appear than expected, which happens when part of the population cannot experience the event. Verify that the rate is genuinely constant across the period or units being pooled, since a varying rate produces overdispersion. Verify independence, since clustering from contagion or correlated failures violates it directly. Expect substantial variation in small counts, since the standard deviation is the square root of the mean and counts around 15 vary by roughly plus or minus 4 routinely. Resist interpreting small count differences as signals, particularly across many locations or periods where some excess is guaranteed by chance alone. Use exact methods for confidence intervals on low counts rather than normal approximations. And remember that queue length rises non-linearly with utilisation, so slack capacity buys response time.

What people get wrong

• Fitting a Poisson model without checking that variance approximately equals the mean, when overdispersion understates standard errors and overstates significance.
• Interpreting a change from 12 to 18 events as a meaningful increase, when Poisson variation alone produces differences of that size routinely at those counts.
• Investigating apparent clusters across many locations without accounting for the number examined, since some excess is guaranteed by chance in any large set of areas.
• Pooling counts across heterogeneous units or periods, which mixes different rates and produces overdispersion even when each component is genuinely Poisson.

Where the math comes from

P(X = k) = (λᵏ · e^(−λ)) / k!, where λ is the average rate over the interval. The mean and variance both equal λ, so the standard deviation is √λ. The distribution arises as the limit of the binomial as the trial count grows and the success probability shrinks with their product held constant.

Questions and answers

Related calculators

Decimal to Fraction · Circle · Golden Ratio · Standard Deviation · Greatest Common Factor