Poisson Distribution in Finance

By Ryan O'Connell, CFA, FRM Published November 5, 2026

The Poisson distribution in finance turns an expected number of events into probabilities for zero, one or many events. A lender expecting two defaults next year still needs to know the chance of no defaults or a much larger count. An operations team may ask similar questions about processing incidents. The calculation is useful when the event definition, exposure and assumptions fit; a correct formula alone does not establish that they do.

This guide explains count probabilities, a checked default example, the exact binomial comparison and the extra assumptions needed for time scaling and waiting times. For the broader choice among count, price and return models, see probability distributions in finance.

What Does the Poisson Distribution Measure?

A Poisson random variable takes nonnegative integer values: 0, 1, 2 and so on. Let X_T be the number of defined events during an interval of length T, and let λ_T be the expected count for that interval. We write X_T ~ Poisson(λ_T). The mean can be fractional even though each realized count is an integer.^[1]

Define the counting unit carefully. Counting distinct borrowers that default differs from counting missed payments: one borrower might miss several payments but contributes only one first default. Likewise, five small processing errors and five major system outages share a count but can have very different financial consequences.

A Poisson distribution describes a count for a specified interval. A homogeneous Poisson process describes how events arrive through time. The process requires additional assumptions about rates and dependence; fitting one interval’s counts does not establish those assumptions.^[3]

Poisson Formula: Exact Counts and Tails

The probability mass function, or PMF, gives the probability of exactly k events:^[1]

Here, e is the base of natural logarithms. Every probability uses the expected count for the same interval as the question. If λ_T = 0, the distribution is concentrated at zero: no events occur under that model.

Zero, One and At Least One Event

Substituting k = 0 or k = 1 gives two useful shortcuts. The probability of at least one event is the complement of zero:

For a larger threshold m, add the probabilities below it and subtract from 1. The last term subtracted is m – 1, because m itself belongs in the requested upper tail:

Mean, Variance and Standard Deviation

A Poisson count’s mean and variance have the same numerical value. Its standard deviation is the square root of that value:^[1]

If the expected count is 2, the variance is numerically 2 and standard deviation is √2 ≈ 1.414 events. This is an exact property of the assumed distribution; a finite sample’s mean and variance need not match exactly. Our standard deviation in finance guide explains the distinction between variance and standard deviation.

Example: Rare Defaults in a Loan Portfolio

Consider a hypothetical fixed portfolio of 400 borrowers. Each has a 0.5% probability of default over the next year. Assume defaults are independent, each borrower can default only once, and no borrowers enter or replace defaulted names. The expected annual default count is 400 × 0.005 = 2.

Under those assumptions, the actual count D is binomial. To explore the rare-event approximation, introduce a separate model count Z ~ Poisson(2). The chart shows that approximation’s probabilities for zero through eight annual defaults; it does not show observed loan outcomes.

One and two defaults are tied as the most likely individual outcomes, at about 27.067% each. Zero defaults still has about a 13.534% probability. An expected count of two therefore leaves substantial uncertainty about the realized count.

That last probability answers a count threshold. Whether five defaults would threaten the lender depends on exposure sizes, recoveries and dependence assumptions, which a simple Poisson count does not resolve. See credit portfolio models for the broader loss-modeling problem.

Poisson vs. Binomial: Check the Approximation

For N independent borrowers with the same horizon default probability p, the exact default count is binomial. Its possible values stop at N:^[2]

The Poisson approximation uses λ = Np. Its rare-event foundation is the limit in which individual probabilities become small while the number of opportunities grows and the expected count stays fixed. A large N alone is insufficient: an approximation also needs sufficiently rare, suitably independent events for the accuracy the decision requires.^[3][4]

For the same hypothetical 400-name portfolio, compute the binomial probabilities with N = 400 and p = 0.005. For example, P(D = 0) = 0.995⁴⁰⁰ ≈ 13.466%; P(D ≥ 5) = 1 – ∑_k=0⁴ C(400,k) × 0.005^k × 0.995^400-k.

The five-or-more difference is about 0.045 percentage points, calculated from the unrounded probabilities. Whether that error is acceptable depends on the application. It does not establish the approximation’s accuracy for a different portfolio, a more extreme threshold or correlated defaults.

Both models have mean 2, but the exact binomial variance is Np(1-p) = 400 × 0.005 × 0.995 = 1.99, compared with 2 for the Poisson approximation. More fundamentally, the Poisson distribution assigns positive probability to counts above 400. Those outcomes are impossible for this fixed portfolio. The approximation does not make repeat defaults by the same name economically possible.^[4]

Use N = 400 and p = 0.005, keeping “exactly” and “at least” distinct. The calculator evaluates the binomial assumptions; it does not test whether the borrowers really default independently.

Scaling a Rate to the Correct Time Interval

Now consider a separate hypothetical example: operational incidents arriving as a homogeneous Poisson process at a constant rate ρ = 0.5 incidents per month, with exposure held fixed. This process assumes independent counts in nonoverlapping intervals and the same count distribution for intervals of equal length, regardless of when they begin.^[3]

For three months, λ₃ = 0.5 × 3 = 1.5 incidents. Use 1.5 in the PMF for a three-month question. The table keeps that rate fixed while changing only the interval length.

The mean scales linearly; event probabilities do not. For example, the three-month probability of at least one incident is 1 – e^-1.5 ≈ 77.69%. Multiplying the one-month probability by three would produce an invalid answer above 100%.

The same rate is 6 incidents per year, so a three-month interval can also be calculated as 6 × 0.25 = 1.5. Do not multiply an annual rate by three months as though the units matched. If activity volume or incident intensity changes, reassess the constant-rate model instead of automatically extending this table.

When Are Waiting Times Exponential?

For the homogeneous process just specified, let W be the waiting time to the next incident. Its survival probability and mean are:^[3][6]

At 0.5 incidents per month, the average wait is 1 / 0.5 = 2 months. The probability of waiting longer than three months is e^-0.5×3 ≈ 22.31%. Waiting longer than T and observing no events during T describe the same outcome here.

Exponential waiting is memoryless: conditional on no incident having occurred, the remaining wait has the same distribution. That property follows from the process assumptions. A long quiet spell does not make an incident “due.”

A Borrower’s Hazard Rate Is Different Again

A single borrower stops being at risk of a first default once that default occurs. If its conditional default hazard is a constant h per year, its default probability over T years is:^[6]

For a separate hypothetical borrower with h = 0.02 per year over three years, p(3) = 1 – e^-0.02×3 ≈ 5.824%. The shortcut hT gives 6.000%. A hazard quoted as 0.02 per year is therefore not exactly a 2% one-year default probability.

For N independent names with that same constant hazard, the horizon count has mean N × (1 – e^-hT) and is binomial under the equal-risk assumptions. As names default, the surviving population shrinks; its aggregate default-arrival intensity is not constant. Replacing that mean by NhT adds a small-hT approximation. The Merton structural credit model takes a different approach by deriving default from the firm’s asset and debt structure.

Default Counts vs. Dollar Losses

Return to the original 400-borrower example. Suppose every borrower has a hypothetical $20,000 exposure at default and a fixed 60% loss given default. Each default then loses $20,000 × 0.60 = $12,000.

The expected loss is $24,000 under these simplified assumptions, including the exact binomial model. Dollars themselves are not Poisson counts: the fixed loss is a multiple of the count. With different loan sizes or uncertain recoveries, the number of defaults alone cannot determine the loss distribution.

For repeated operational events, a model that sums random event severities over a Poisson count is called a compound Poisson model. It needs severity assumptions as well as frequency assumptions. In a finite credit portfolio, name-specific exposures and recovery dependence also matter. A Value at Risk calculation must use the loss distribution relevant to the position, rather than silently treating a default-count quantile as a dollar-loss quantile.

Stock returns are another variable entirely: they can be negative and are not nonnegative integer counts. Do not apply a Poisson PMF directly to percentage returns. A model that separately counts price-jump events still needs a model for their sizes and the remainder of the return. Our normal distribution in finance guide discusses return probabilities and their different limitations.

Limitations: Clustering and Overdispersion

A fixed Poisson mean forces variance to equal the mean. Overdispersion means the count variance exceeds the mean. For comparable exposures, that is a reason to investigate the fixed-mean Poisson model, although sampling uncertainty also matters.

Shared economic shocks can raise many borrowers’ risks together; persistent stressed conditions can also create dependence through time. Neither unconditional default dependence nor serial clustering is represented by the simple independent, constant-rate model. CreditRisk+ addresses shared risk through conditional Poisson approximations and random risk factors; it is more than a single unconditional Poisson distribution.^[4]

Same Average Count, Different Tail

Consider a new hypothetical annual event-count model, separate from the finite borrower portfolio. An annual interval has an expected count of either 0.5 or 3.5, each with 50% probability. Conditional on that interval’s mean, its count is Poisson. Compare this mixture with a fixed Poisson mean of 2; both have the same overall average.

The mixture has more empty intervals and a higher probability of large counts. Its mean is 0.5 × 0.5 + 0.5 × 3.5 = 2. Its variance is the average conditional variance, 2, plus the variance of the interval means, 2.25, giving 4.25. Its zero probability is 0.5e^-0.5 + 0.5e^-3.5 ≈ 31.836%; its five-or-more probability is the same weighted average of the two component tails, about 13.736%.

A gamma-distributed random Poisson mean produces a negative binomial distribution, a useful alternative when extra count variation needs to be modeled.^[5] The two-point mixture above is a different mixed-Poisson distribution. Neither choice by itself specifies whether high-count years persist: that requires assumptions about how the underlying rates evolve.

Heterogeneous Risk Needs Careful Interpretation

Different borrowers can have different fixed default probabilities without sharing a shock. If their defaults remain independent, the exact count is Poisson-binomial, with mean ∑p_i and variance ∑p_i(1-p_i). That variance is no greater than the mean. Unequal probabilities alone therefore do not establish overdispersion or contagion; a shared random environment is a different assumption.

Possible alternatives include modeling known exposure changes, using a mixed-rate or negative-binomial count model, or using a credit model that preserves finite-name default indicators and common risk factors. Negative-binomial counts are also unbounded, so they do not eliminate the finite-portfolio approximation issue. The choice should follow the source of the model failure.

How to Evaluate a Poisson Model

1. Define the event and exposure. Specify distinct borrower defaults, claims or operational incidents; state the interval and population or activity volume.
2. Choose the right mean. For equal independent borrower PDs, use Np for the horizon count approximation. For a justified homogeneous process, use ρT. Do not interchange a per-name hazard, aggregate rate and interval mean.
3. Estimate consistently. With comparable equal-length intervals, the sample average count estimates the Poisson mean. Under a constant-rate process with differing observation lengths, total events divided by total observed time estimates the rate.^[1] Changes in exposure require separate adjustment.
4. Check dispersion and dependence. Compare sample mean and variance with appropriate allowance for sampling error; examine common shocks, event batches, changing intensity and serial dependence. Matching two moments does not prove a Poisson fit.
5. Calculate the actual question. Separate exactly k, at least k and more than k. Check the approximation against the exact binomial where the portfolio assumptions permit it.
6. Stress the decision. Vary the mean and dependence assumptions, examine larger counts and uncertain severity, and retain limits arising from sparse event data. A precise tail calculation can still have substantial model uncertainty.

Common Mistakes

• Using a rate as an interval mean. Convert events per month to expected events over the chosen number of months before applying the PMF.
• Multiplying probabilities across time. A mean can scale linearly under a constant-rate process; an at-least-one probability generally does not.
• Subtracting one term too many. For P(X ≥ 5), subtract probabilities for zero through four, leaving five in the tail.
• Calling finite defaults exactly Poisson. One name contributes at most one default, and a fixed N-name portfolio cannot have more than N defaults.
• Inferring exponential waits from a histogram. A marginal count fit does not establish independent or stationary increments.
• Replacing 1 – e^-hT with hT without qualification. The shortcut is for small hT; a hazard is not an exact horizon probability.
• Confusing counts with losses. Add exposure and severity assumptions before translating default counts into dollars or applying them to capital decisions.

Sources and Calculation Notes

Sources were checked November 5, 2026. All portfolio, incident-rate, hazard and loss inputs are hypothetical. Probabilities are rounded from unrounded calculations; no observed event data or forecast performance is claimed.

1. NIST/SEMATECH: Poisson Distribution, PMF, cumulative probabilities, support, moments and mean estimation.
2. NIST/SEMATECH: Binomial Distribution, finite support, PMF and moments.
3. Robert G. Gallager, MIT: Poisson Processes, Spring 2011, sections 2.2.1–2.2.5; process assumptions, waiting times, count probabilities and Bernoulli limits.
4. Michael B. Gordy, Federal Reserve Board: A Comparative Anatomy of Credit Risk Models, December 8, 1998, section 1.1 and section 5; conditional default-count approximations, risk-factor mixing and finite-name limits.
5. R stats documentation: The Negative Binomial Distribution, Details; gamma-Poisson mixing and overdispersed counts.
6. NIST/SEMATECH: Exponential Distribution, survival function, constant hazard and mean waiting time.