Poisson distribution

Poisson distribution is a discrete probability model that predicts how often a rare, independent event happens inside a fixed window of time or space. It answers one blunt question: given an average rate, what are the odds of exactly this many events?

Named for French mathematician Siméon Denis Poisson, who published it in 1837, the model now sits under staffing plans, queue forecasts and risk pricing — anywhere arrivals look random but average out.

The maths drives calls you probably never label as statistics. How many agents to roster for the 9pm shift. How many spare servers to hold back. Get the average right, and Poisson hands you the odds of any specific count.

Key takeaways

• Poisson distribution runs on one parameter, lambda (λ), the average number of events per interval.
• It fits events that are independent, rare, and counted in whole numbers inside a fixed window.
• Contact centre planners feed Poisson arrival rates straight into Erlang C to size each shift.
• Insurers, fraud teams, and site reliability engineers price rare event frequency the same way.
• Above roughly λ = 30, the curve flattens and a normal approximation takes over the maths.

How it works

Poisson distribution calculates the probability of exactly k events in an interval when those events arrive at a known average rate λ and stay independent of each other. Mean and variance both equal λ, which makes the model easy to test against real data.

Four conditions have to hold before you apply it. Events are countable in whole numbers, two events never land at the same instant, the rate λ holds steady across the interval, and each event ignores the last.

The formula is P(k) = (λ^k × e^−λ) / k!. Here k is the count you’re asking about, λ is the average per interval, e ≈ 2.71828, and k! is k factorial — so 4! = 24.

Work a small one. A helpdesk takes two tickets per minute on average, so λ = 2. Poisson puts a quiet minute with zero tickets at 13.5%, one ticket at 27.1%, and four tickets at 9.0%.

Source: standard Poisson probability mass function, at λ values typical of small and mid sized call queues.

Before you trust any of it, test the fit. Count events across a sample of equal windows, then compare the mean with the variance. If the two drift far apart, your rate isn’t constant and Poisson will understate the busy stretches.

That count table is exactly what queue maths needs. Erlang C, the staffing formula behind most workforce management tools, assumes Poisson arrivals and exponential handling times, then returns the odds a caller waits — Poisson supplies the arrival half.

Once λ climbs past roughly 30, the curve flattens and starts to mirror a normal distribution, so analysts switch models for high volume queues. At tiny λ values, the shape stays sharply skewed toward zero.

Examples

Poisson turns up wherever counts are small, independent, and worth planning around. The cases below span offshore contact centres, national workforce data, insurance pricing, retail engineering, and hospital triage, roughly the applied statistics a BPO buyer meets.

Contact centre staffing in the Philippines. A Manila site running a US retail account might take 240 inbound calls an hour at evening peak, so λ = 4 per minute. Planners size headcount from that, then run Erlang C for the service level target.

Customer service workforces at national scale. The US Bureau of Labor Statistics, the federal agency that tracks American wages and employment, lists customer service representatives among the country’s largest occupations (BLS Occupational Employment).

Insurance claim pricing. Lloyd’s of London, the specialist insurance market, prices cyber cover with Poisson because attacks on any one firm are rare and roughly independent. Its 2023 systemic risk work layers severity on that base (Lloyd’s systemic risk scenarios).

E-commerce checkout failures. Payment errors are rare, independent, and countable, so platform engineers model them with Poisson to set autoscaling thresholds and on-call rotations long before Black Friday traffic lands.

Hospital emergency arrivals. A regional hospital averaging 12 ambulance arrivals an hour works from λ = 12 and staffs triage against the resulting spread, sizing for the busy tail rather than the comfortable average.

Related terms

These sibling concepts show up constantly in capacity planning and analytics work. Most workforce management dashboards lean on at least three of them at once, and the distinctions matter when you are checking a vendor’s model.

• Binomial Distribution: the fixed trial yes or no model that Poisson approximates when trials are many and successes rare.
• Normal Distribution: the continuous curve Poisson converges toward once lambda climbs past roughly 30.
• Standard Deviation: the spread measure, equal to the square root of lambda for any Poisson variable.
• Probability: the wider discipline Poisson sits inside, covering both discrete counts and continuous measures.
• Forecasting: the planning practice that turns Poisson outputs into rosters, inventory orders, and budgets.
• Predictive Analytics: the modelling layer that extends Poisson counts with machine learning features.
• Queuing Theory: the service time maths, Erlang C included, that sits on top of Poisson arrivals.

FAQ

When should I use Poisson instead of binomial distribution?

Use Poisson when trials are very many, the per trial chance is very small, and all you really know is the average rate. Use binomial when you have a fixed trial count and a clear per trial probability.

What is lambda (λ) in Poisson distribution?

Lambda is the average number of events in the interval you choose. A queue averaging 60 calls an hour gives λ = 60 for an hourly model, or λ = 1 for a per minute model. Both views work while the rate holds steady.

Can Poisson distribution have decimal answers?

The input λ can be any positive decimal, so 3.7 calls a minute is fine. The output k cannot. Probabilities attach to exactly four events, never to exactly 4.3 events.

How accurate is Poisson for real business forecasting?

It holds up while the four assumptions do: independence, a constant rate, rarity, and whole number counts. Real queues break the constant rate rule across a day — so planners split the day into windows where λ stays stable, then model each one.

Is Poisson distribution still used in modern AI workflows?

Yes, because recommendation engines, anomaly detection, and ad bidding systems all model event counts before deeper neural layers take over.

Ready to put these models behind your own staffing plan? Browse the vetted analytics and BPO teams in the Outsource Accelerator directory and shortlist a partner.