The Poisson Process: From Theory to Insurance Claims
How the Poisson process models claim arrivals and its extensions for real-world insurance applications.
The Homogeneous Poisson Process
The Poisson process is the most fundamental model for counting events that occur randomly over time. A homogeneous Poisson process with rate lambda has three key properties: the number of events in any interval of length t follows a Poisson distribution with mean lambda times t, counts in non-overlapping intervals are independent, and events do not occur simultaneously. In insurance, this process models claim arrivals, where lambda represents the expected number of claims per unit time.
The exponential distribution of inter-arrival times follows directly from these assumptions. If N(t) is a Poisson process with rate lambda, the time between successive events is exponentially distributed with mean 1/lambda. The memoryless property of the exponential distribution means that the probability of a claim arriving in the next small interval does not depend on how long it has been since the last claim. While this assumption is often a simplification of reality, it provides a tractable and useful baseline model.
Properties and Calculations
Several properties of the Poisson process are particularly useful in actuarial applications. The superposition property states that the sum of independent Poisson processes is itself a Poisson process. If claims from different coverage types arrive as independent Poisson processes with rates lambda_1, lambda_2, and so on, the combined claim arrival process is Poisson with rate lambda_1 + lambda_2 + ... This property justifies modeling total claims across coverages as a single Poisson process.
The thinning property provides the converse: if each event in a Poisson process is independently classified into one of several categories with fixed probabilities, the resulting sub-processes are independent Poisson processes. For example, if claims arrive at rate lambda and each claim has probability p of exceeding a large loss threshold, then large claims arrive as a Poisson process with rate lambda p.
The conditional distribution of event times, given the total count, is particularly elegant. Given that exactly n events occur in [0, T], the event times have the same joint distribution as n independent uniform random variables on [0, T] sorted in ascending order. This result is useful for simulation and for understanding the structure of Poisson processes.
The Non-Homogeneous Poisson Process
Real insurance claim patterns often violate the constant-rate assumption. The non-homogeneous Poisson process (NHPP) allows the rate to vary over time through a rate function lambda(t). The number of events in any interval [s, t] follows a Poisson distribution with mean equal to the integral of lambda(u) from s to t. This generalization captures seasonal patterns in claims, such as higher auto accident rates in winter months, increased property claims during hurricane season, or elevated health insurance utilization at the beginning of the year when deductibles reset.
The NHPP maintains the independence of counts in non-overlapping intervals, but inter-arrival times are no longer exponentially distributed (they follow a distribution that depends on the rate function). Simulating an NHPP typically uses either the thinning method (generating events from a homogeneous process at the maximum rate and randomly deleting excess events) or the time transformation method (transforming the non-homogeneous process into a homogeneous one through a change of time scale).
The Compound Poisson Process
The compound Poisson process pairs random claim arrival times with random claim amounts, forming the basis of aggregate loss models. If N(t) is a Poisson process with rate lambda and X_1, X_2, ... are independent, identically distributed claim amounts (also independent of N(t)), then the aggregate loss S(t) = X_1 + X_2 + ... + X_{N(t)} is a compound Poisson process.
The moments of the compound Poisson distribution are computed using conditioning formulas:
- E[S] = lambda t E[X], where E[X] is the expected claim size
- Var[S] = lambda t E[X^2], where E[X^2] is the second moment of claim size
- The moment generating function of S is M_S(r) = exp(lambda t (M_X(r) - 1)), where M_X is the claim size MGF
These formulas are central to actuarial pricing and reserving. The compound Poisson model appears in the Cramer-Lundberg ruin model, in aggregate loss distributions for reinsurance pricing, and in the collective risk model taught in early actuarial exams.
The Mixed Poisson Process
The mixed Poisson process introduces a random rate parameter, naturally producing overdispersion (variance exceeding the mean) commonly observed in insurance data. Instead of a fixed rate lambda, each risk or group has its own rate drawn from a mixing distribution. Common mixing distributions include the gamma distribution (producing a negative binomial marginal for the claim count) and the inverse Gaussian distribution (producing a Poisson-inverse Gaussian marginal).
The negative binomial distribution, arising from a Poisson-gamma mixture, is one of the most widely used claim count distributions in insurance. Its probability function and moments are: E[N] = alpha beta and Var[N] = alpha beta (1 + beta), where alpha and beta are the gamma mixing distribution parameters. The variance exceeds the mean by a factor of (1 + beta), providing a natural mechanism for modeling the extra variability that pure Poisson models cannot capture.
Mixed Poisson models connect naturally to credibility theory. The mixing distribution represents the unobservable heterogeneity across risks, and Bayesian updating of the mixing parameter based on observed claims corresponds to the credibility-weighted estimate of the individual risk's true claim rate.
Applications in Actuarial Practice
These Poisson process variants appear throughout actuarial practice. In ratemaking, the compound Poisson model provides the foundation for aggregate loss projections. In reserving, non-homogeneous processes help model reporting and settlement patterns. In catastrophe modeling, spatial Poisson processes model the locations of insured properties. In reinsurance pricing, compound Poisson models are used to estimate losses in excess layers.
When studying these models for actuarial exams, practice deriving moments using the conditioning approach (condition on the Poisson count, then apply the law of total expectation and total variance). Become comfortable with the moment generating function technique for computing the distribution of aggregate losses. Work through examples of thinning, superposition, and mixing to develop intuition for how different model choices affect the shape and tail behavior of the resulting distributions.