The Poisson distribution is a discrete probability distribution that describes the number of times an event is likely to occur in a fixed interval of time or space when those events occur with a known constant average rate and independently of the time since the last event. It is widely used in operations research, queuing theory and finance to model counts such as trades, defaults, jumps, failures, or customer arrivals. (Source: Investopedia)
Key takeaways
– The Poisson distribution models count data (0, 1, 2, …) for events occurring independently at a constant average rate λ (lambda).
– PMF: P(X = x) = (λ^x e^(-λ)) / x!; mean = variance = λ.
– Estimate λ with the sample mean; test fit by comparing mean and variance (overdispersion suggests other models).
– Common finance uses: count of trades, arrival of market “shocks” or jumps, rare-event modeling, and Poisson-based jump-diffusion pricing (e.g., Merton jump model). (Source: Investopedia; Merton 1976)
Understanding the Poisson distribution — formula and properties
– PMF (probability mass function): P(X = x) = (λ^x e^(-λ)) / x!, for x = 0, 1, 2, …
– Cumulative distribution: P(X ≤ k) = Σ_{x=0}^k P(X = x).
– Parameters and moments: λ > 0 is both the expected value E[X] and the variance Var[X].
– Limiting/approximation results:
• Poisson approximates Binomial(n, p) when n is large and p is small with λ = np.
• When λ is large, the Poisson is approximately Normal with mean λ and variance λ.
– Relationship to processes: Numbers of events in disjoint intervals are independent; counts in interval of length t are Poisson(λ t). Interarrival times are exponential with mean 1/λ.
Explain Like I’m 5
If on average 3 apples fall from a tree every hour, Poisson tells you the chance of seeing exactly 0, 1, 2, 3, … apples fall in a given hour, using the average (3) to compute those probabilities.
When to use the Poisson distribution (practical guidance)
Use Poisson when:
1. Your outcome is a count (0, 1, 2, …).
2. Events are approximately independent.
3. Events happen at a roughly constant average rate over the interval of interest.
4. Events are rare relative to the number of opportunities (in some contexts), or mean counts are small.
If the variance is substantially greater than the mean (overdispersion), consider alternatives such as the negative binomial. If events are dependent or rates vary in time, consider non-homogeneous Poisson processes or Cox processes.
Assumptions the Poisson distribution makes
– Independence: occurrences do not influence each other.
– Stationarity: rate λ is constant over the interval.
– No simultaneous occurrences in an infinitesimally small subinterval (practically: probability of two events in an infinitesimal interval is negligible).
– Mean equals variance (Var[X] = E[X] = λ).
Is the Poisson distribution discrete or continuous?
– Discrete. The random variable takes whole nonnegative integers only.
Practical steps: how to apply the Poisson distribution to real data
1. Define the interval and the event you count (per minute, day, trader, branch, decade, etc.).
2. Collect data: record counts across many comparable intervals. Example: daily number of trades for a trader over 100 trading days.
3. Compute sample statistics:
• Sample mean λ̂ = (Σ counts) / N -> MLE for λ.
• Sample variance s^2 = Σ (count – λ̂)^2 / (N – 1).
4. Quick check for fit:
• If s^2 ≈ λ̂, Poisson is plausible.
• If s^2 >> λ̂ (overdispersion) -> consider negative binomial or zero-inflated models.
5. Estimate probabilities:
• Use PMF: P(X = x) = (λ̂^x e^{-λ̂}) / x!.
• For tail probabilities (e.g., P(X > k)) use 1 − CDF.
6. Goodness-of-fit tests:
• Chi-square test comparing observed counts to expected Poisson probabilities (requires sufficient sample size in each cell).
• Likelihood-ratio tests and information criteria (AIC/BIC) when comparing Poisson vs alternatives.
7. Build models:
• Regression-style: Poisson regression (log link) relates expected count λ to covariates: log(λ_i) = X_i β. If overdispersion exists, use quasi-Poisson or negative binomial regression.
8. Estimate confidence intervals for λ:
• Approximate: λ̂ ± z_{α/2} sqrt(λ̂ / T) for counts aggregated over T units (approx normal for large counts).
• Exact/More accurate: use chi-square-based intervals for Poisson rates.
9. Implement computations:
• Excel: POISSON.DIST(x, mean, FALSE) for PMF and TRUE for CDF.
• Python: scipy.stats.poisson.pmf, .cdf, .rvs. Example:
from scipy.stats import poisson
poisson.pmf(k, mu=lam)
• R: dpois(x, lambda), ppois, rpois.
Worked examples
1) Small-mean example (errors): Suppose an operational process has error rate 3% with 100 independent trials → λ = 100 * 0.03 = 3.
• P(0 errors) = e^{-3} ≈ 0.0498.
• P(1 error) = 3 e^{-3} ≈ 0.1494.
• P(2 errors) = (3^2/2!) e^{-3} ≈ 0.2240.
• These probabilities are easy to compute by hand, Excel or code.
2) Large-mean example (customers): If average customers per night λ = 200, probability of more than 300 customers is extremely small. Approximating with Normal: mean 200, sd ≈ sqrt(200)=14.14, z ≈ (300.5−200)/14.14 ≈ 7.1 → near-zero probability.
Poisson in finance — practical applications and caveats
– Trade counts: Model number of trades per time unit for high-level activity metrics; useful in market microstructure and order arrival analysis.
– Market “shocks” / jumps: Poisson processes are used to model the random arrival times of jumps. Combined with jump-size distributions, they produce compound Poisson processes. Merton’s jump-diffusion model uses a Poisson jump arrival process for option pricing. (See Merton, 1976.)
– Credit defaults / rare events: For portfolios with many exposures and rare defaults, Poisson approximations may be used to estimate tail counts under simplifying assumptions.
– Operational risk: counting incidents such as system failures or fraud events over a period.
Caveats:
– Real financial event rates often vary (intraday seasonality, volatility clustering) and may violate stationarity and independence. Use non-homogeneous Poisson or Cox processes when λ(t) varies stochastically.
– Overdispersion and excess zeros are common—consider negative binomial or zero-inflated Poisson models.
– When modeling jump sizes and their impact on returns, use compound Poisson or jump-diffusion frameworks rather than raw Poisson counts alone.
Diagnosing and handling poor fit
– If variance ≫ mean: consider Negative Binomial (adds dispersion parameter).
– If too many zeros: use zero-inflated Poisson or hurdle models.
– If rate varies over time: fit a nonhomogeneous Poisson process (time-varying λ(t)), or model λ with covariates via Poisson regression.
– If events cluster: consider Cox processes (doubly stochastic Poisson) or Hawkes processes (self-exciting); Hawkes is often used in finance to model trade clustering.
Advanced extensions
– Poisson process: continuous-time counting process N(t) where N(t) ~ Poisson(λ t). Interarrival times ~ Exponential(λ).
– Compound Poisson: sum of random jump sizes occurring at Poisson times → used to model aggregated loss amounts or jump components of asset returns.
– Mixed Poisson: λ itself is random (leading to marginal distributions such as negative binomial) and is useful when there is unobserved heterogeneity in event rates.
How to report results and communicate them
– Report λ̂ and its confidence interval, show observed vs expected frequency table, and include goodness-of-fit statistics.
– If you use Poisson regression, present incidence rate ratios (exp(β)) and goodness-of-fit or dispersion diagnostics.
– If using Poisson in pricing or risk models, spell out assumptions about stationarity and independence and test or stress those assumptions.
Quick reference formulas
– PMF: P(X=x) = λ^x e^{-λ} / x!
– Mean: E[X] = λ
– Variance: Var[X] = λ
Tools and implementation references
– Excel: POISSON.DIST (Microsoft support).
– Python: scipy.stats.poisson (SciPy docs).
– R: dpois, ppois, rpois (base R).
– For jump-diffusion and applied finance: Merton (1976), and standard texts on stochastic processes (e.g., Ross, “Introduction to Probability Models”).
Bottom line
The Poisson distribution is a simple, tractable model for counts of independent events occurring at a constant average rate. It is especially useful in finance for modeling counts of trades, rare events, and the arrival of jumps when assumptions are reasonable. Always check the data (mean vs variance, time variation, clustering) and consider more flexible models (negative binomial, non-homogeneous Poisson, Hawkes processes, compound Poisson) when real-world features violate Poisson assumptions.
Sources
– Investopedia: “Poisson Distribution” (Joules Garcia).
– Merton, R. C. (1976). “Option Pricing when Underlying Stock Returns are Discontinuous.” Journal of Financial Economics.
– SciPy documentation: scipy.stats.poisson.