RESEARCH DOSSIER
Stochastic Claim Processes
Abstract. An insurance claim process is a temporally indexed stochastic system that connects risk exposure, occurrence of loss-producing events, the emergence of contractual claims, reporting, assessment, payment, and settlement. Its statistical structure governs the meaning of frequency, claim clustering, waiting time, development delay, and projected liability. This monograph develops the mathematics of claim occurrence as a counting process; derives homogeneous and nonhomogeneous Poisson likelihoods and waiting-time distributions; compares renewal models with memoryless arrivals; formulates marked Poisson processes for severity and reporting development; and examines random-intensity, Markov-modulated, and self-exciting point processes. It then connects latent occurrence processes to observed insurer databases under reporting delays, claim-state transitions, censoring, changing exposures, and incomplete information. Original worked counterexamples quantify how identical expected counts can conceal sharply different zero-claim probabilities and variances, how gamma mixing and slow Markov regimes alter predictive uncertainty, and how reporting delay produces a large stock of incurred but unreported claims even when occurrence intensity is completely known. The monograph integrates exact probabilistic derivations with original methodological synthesis, identifies the limits of parameter identification and causal interpretation from claims histories, and establishes a reproducible sequence of model diagnostics. It treats aggregate-loss and surplus questions only to the extent needed to distinguish their objects from the arrival process; those topics have separate Research Dossiers. Its conclusions are analytical assessments of established stochastic models, not a new named theory or proprietary actuarial system.
Keywords: Stochastic claim processes; insurance point processes; claim frequency; homogeneous Poisson process; nonhomogeneous Poisson process; renewal process; interarrival time; marked point process; reporting delay; incurred but not reported; Cox process; mixed Poisson process; Markov-modulated Poisson process; Hawkes process; claim-state dynamics; counting-process martingale; censoring; exposure adjustment; prediction; actuarial model validation.
Introduction
Insurance losses unfold in time. A fire or traffic collision may occur on one date, become known to the insured later, be reported to the insurer after another interval, prompt a sequence of assessments and partial payments, and end with a settlement that may itself be reopened. A policy portfolio may change in size or composition between all of these events. Thus the date of loss occurrence, the date of claim notification, the date of payment, and the reporting date in an insurer's accounting extract define different stochastic observations. A model of counts by calendar month cannot be interpreted until the event clock and the exposure denominator are identified.
The appropriate mathematical starting point is the counting process N(t), the number of occurrences in an interval beginning at a defined origin and ending at time t. It has nondecreasing, integer-valued, right-continuous paths, begins at zero under the usual convention, and increases by one at a simple event time. Insurance can require a multitype or marked process when several transactions attach to the same occurrence, and simultaneous losses may require event groups or marks rather than an unjustified assumption that each transaction is an independent arrival. The conceptual foundations of random measures and point processes are developed in Daley and Vere-Jones (2003). Their theory supplies definitions; actual insurability and prediction depend on the process observed and the information available.
Let E(t) denote measurable risk exposure at time t, such as insured vehicle-years per year, and let λ(t | H_t) denote conditional claim intensity per portfolio unit of clock time given history H_t. Exposure and conditional intensity play different roles: increasing E(t) can raise expected counts without changing the hazard per policy, whereas a rising conditional hazard per policy can raise the same observed counts even with fixed exposure. The separation is particularly important for seasonal business, policy inception and expiration, portfolio acquisition, natural catastrophe windows, and post-event claims reporting. A claim notification is a delayed observation of an occurrence and is not automatically a second occurrence.
The canonical studies supply complementary evidence. Cox's analysis of event series (1955) treats randomness and components of variance. Arjas's structural reserving framework (1989) connects point processes and martingales to insurance liabilities. Norberg's continuous-time liability analysis (1993) expresses occurrence and development as a marked, nonhomogeneous Poisson process. Later empirical insurance research tests whether overdispersion, reporting delays, and changing exposure require more elaborate intensities. The research object here is the time-indexed process generating and transforming claims, including what can be inferred about its latent history from incomplete records.
The treatment is international in mathematical scope. It makes no country-specific legal assumption about which event creates a contractual claim, which reporting deadlines apply, or which payment constitutes final settlement. Contract wording, claim definitions, coverage exclusions, deductibles, and administrative recording rules affect the observed event process and should be recorded with the model rather than treated as universal constants. The purpose is scholarly explanation, quantitative verification, and comparison of model classes, not a personalized pricing or reserving opinion.
Poisson Counting Processes
Definition, assumptions, and meaning of the clock
For λ > 0, a homogeneous Poisson process has N(0) = 0, independent counts on disjoint time intervals, and N(t) - N(s) distributed as Poisson[λ(t − s)] for 0 ≤ s < t. Its probability mass function is P(N(t) = n) = exp(−λt)(λt)^n/n!, n = 0, 1, ... . Its mean and variance both equal λt. The model thus specifies more than a mean claim count: it imposes independent increments and equality of conditional variance and conditional mean for increments measured at a fixed intensity. The parameter λ carries a time unit. A rate of three per year is not numerically the same as three per month.
The waiting times between events are independent exponential variables with mean 1/λ. The equality P(N(t + h) − N(t) = 0 | H_t) = exp(−λh) expresses memorylessness; an accident's imminent occurrence hazard does not depend on how long the process has waited since its previous event under the model. These are strong physical and informational assumptions. A heterogeneous pool can display an approximately Poisson total without each individual policy being independently exposed, while catastrophe clusters and repeated claims from a single unresolved incident can violate the conditions.
For a nonhomogeneous Poisson process (NHPP) with deterministic, nonnegative, locally integrable intensity λ(t), the cumulative intensity is Λ(t) = ∫0^t λ(u) du. Independent increments remain, and N(t) − N(s) is Poisson with mean Λ(t) − Λ(s). For a specified observation horizon T and observed distinct event times 0 < t_1 < ... < t_n ≤ T, its point-pattern likelihood, apart from a parameter-independent convention for event ordering, is L(θ) = exp[−Λ_θ(T)] ∏{i=1}^n λ_θ(t_i). Hence log L(θ) = Σ_i log λ_θ(t_i) − Λ_θ(T). This formula assumes complete ascertainment of event times and a correctly defined observation window. Reporting-delayed datasets require an observation model in addition to an occurrence model.
Exposure-standardized intensity and seasonality
An operational construction is λ(t) = E(t) r(t), with E(t) known at risk and r(t) a hazard per exposure unit. It allows one to compare insurers whose policy counts differ and to distinguish a high-claims season from a high-volume season. The integrated exposure-weighted mean is Λ(T) = ∫_0^T E(t)r(t)dt, not simply a calendar-time average multiplied by the end-period portfolio size. A piecewise-constant exposure path can be integrated exactly. If premium records and exposure records use inconsistent date conventions, the likelihood's denominator can be wrong even when the counting process code is numerically correct.
For illustration, a portfolio with 1,000 units exposed for six months and 2,000 for the next six months, each at r = 0.02 claims per unit-year, has Λ(1 year) = 0.02(500 + 1,000) = 30 claims. Using the closing exposure of 2,000 over the entire year would imply 40. The resulting 33.3% overstatement of modeled expected claim count is an exposure-denominator error, not evidence of stochastic overdispersion. All numbers in this example are constructed for mathematical verification, not drawn from insurer data.
Martingales and conditional intensities
In a general simple point-process model, define a predictable nonnegative intensity λ(t | H_{t−}) with respect to a stated filtration. Subject to the regularity conditions of the Doob–Meyer decomposition, A(t) = ∫0^t λ(s | H{s−})ds is the compensator and M(t) = N(t) − A(t) is a local martingale; appropriate integrability promotes the relevant expectation statements. The formalism distinguishes conditional prediction from deterministic intensity and is central to statistical point-process inference. Aalen's counting-process inference (1978) illustrates how cumulative intensity can be estimated under explicit observational conditions.
A common misuse is to infer from the unconditional equality E[N(T)] = E[A(T)] that N(T) is Poisson. A random compensator may yield larger variance or temporal dependence. In fact, under a Cox process that is conditionally Poisson given its entire exogenous intensity path, Var[N(T)] = E[Λ(T)] + Var[Λ(T)], while E[N(T)] = E[Λ(T)]. The second term is the uncertainty in integrated intensity and is absent from a deterministic NHPP. Conversely, under general history-dependent intensity, conditional Poisson arguments require care: the process need not be Cox and conditioning on a path that itself depends on earlier arrivals can invalidate an assumed independent-increments interpretation.
A first distributional counterexample
Compare two processes on a one-year horizon, both with expected count four. Model A is Poisson with mean four. Model B is conditionally Poisson with integrated intensity 4Z, where Z is gamma distributed with shape 2 and rate 2, so E[Z] = 1 and Var(Z) = 1/2. Then E[N] = 4 for both, but Var(N) = 4 for A and 4 + 16/2 = 12 for B. Moreover, P_A(N = 0) = exp(−4) ≈ 0.0183, while P_B(N = 0) = E[exp(−4Z)] = (2/6)^2 = 1/9 ≈ 0.1111. The same annual expected frequency is therefore compatible with a more than sixfold difference in the zero-claim probability. A mean-only comparison cannot distinguish the models. The second model is a gamma-mixed Poisson process for a one-year horizon, not a claim that every real insurance count follows a negative binomial law.
Arrival and Waiting Times
Interarrival distributions, conditional survival, and forward recurrence
Let T_n denote the nth occurrence time and W_n = T_n − T_{n−1} the nth gap, T_0 = 0. For a homogeneous Poisson process, the W_n are independent exponential(λ). The nth arrival T_n is gamma/Erlang with integer shape n and rate λ. The probability that the first event occurs after t is exp(−λt), and the constant hazard λ follows from −d log P(T_1 > t)/dt. This equivalence between exponential interarrival times and a constant-rate Poisson process depends on the renewal construction and absence of additional event types or latent state dynamics.
For an NHPP observed to time s, the probability of no arrival during the following h units is exp{−[Λ(s+h) − Λ(s)]}. Its forward waiting-time survival varies with current calendar time even without feedback from previous events. For example, an annual portfolio with higher winter exposure may have a shorter expected forward wait in winter than in summer although both intervals are governed by independent Poisson increments. An observed seasonal pattern is thus insufficient evidence of contagion or self-excitation.
Conditionally on N(T) = n under an NHPP, unordered occurrence times have density λ(t)/Λ(T) on [0,T]; equivalently, ordered times are the order statistics of n independent draws from that density. Under a homogeneous Poisson process, this becomes uniformity on the window conditional on the count. The conditional property is useful for time-of-year diagnostics when the total count is fixed, but it does not establish that reports observed during the year equal occurrences during the year.
Observation windows, truncation, and delayed entry
The waiting-time distribution of a randomly selected interarrival interval differs from that of a randomly observed forward recurrence interval outside the exponential case. The length bias associated with sampling a point in calendar time becomes relevant when a dataset begins during an ongoing insurance episode or a claim is selected because it remains open. Right truncation arises when claims not yet reported by the extraction date are excluded altogether; right censoring arises when known open claims have incomplete development. These mechanisms have different likelihood contributions. A naive histogram of the observed delays will underestimate long reporting delays when the newest accident cohorts have not matured.
The data-generating process should record occurrence time, contractual incidence time where distinct, reporting time, successive transaction times, closure time, and extract timestamp with their data lineage. Changes in reporting software, migration of policy systems, and backdated administrative entries can introduce artificial bursts into an otherwise regular occurrence process. Event-time precision and time zones matter where multiple claims share a day: a daily count can contain ties even when the underlying mathematical point process is simple in continuous time.
Interarrival inference versus pooled frequencies
Estimating λ by N(T)/T is the homogeneous maximum-likelihood estimator under complete observation and correct exposure normalization. Confidence intervals based on Poisson sampling quantify event randomness conditional on the model; they omit intensity-model misspecification, exposure measurement error, and dependence-induced extra variance. A pattern of similar monthly mean counts but unusually long runs of zeros followed by dense clusters suggests checking a mixed-intensity or state-dependent model. It is diagnostic evidence requiring explicit model comparison, not proof of any unique generating mechanism.
Renewal Models
General renewal structure
A renewal process takes independent identically distributed nonnegative waiting times W_1, W_2, ... with common distribution F and defines N(t) = max{n: W_1 + ... + W_n ≤ t}. Its interarrival law can be Weibull, gamma, lognormal, or another positive distribution; a Poisson process is the exponential special case. If μ_W = E[W] exists, is finite, and the usual renewal theorem conditions hold, N(t)/t approaches 1/μ_W almost surely as t grows. This long-run frequency result does not imply Poisson counts or independent increments at finite horizons.
A renewal model can represent recurring machine failures following repair, replacement, or reset of a claim-relevant asset where the clock since renewal influences the next event. It requires a defensible reset definition. Replacing a cracked windscreen may reset risk of that component; processing an administrative claim does not necessarily reset the hazard of an underlying driver, building, or natural peril. Different renewal clocks for separate insured objects can be superposed, but the superposition of non-Poisson renewals is generally not another renewal process.
The renewal function m(t) = E[N(t)] satisfies the renewal equation m(t) = F(t) + ∫0^t m(t−u)dF(u), with interpretation as a Stieltjes integral. For a continuous distribution F with density f, renewal density u(t), when it exists, is Σ{n≥1} f^{*n}(t), and m(t) = ∫_0^t u(s)ds. The counting distribution follows P(N(t) ≥ n) = F^{*n}(t). These relationships permit exact or numerical comparison with Poisson benchmarks at operational horizons.
Erlang waiting-time counterexample
Take an ordinary renewal process whose independent gaps follow Erlang(shape 2, rate 2 per year). Their mean is one year and variance is one-half year squared, so their long-run average count is one per year, matching a Poisson process with λ = 1. At t = 1 year, however, P(N ≥ 1) = 1 − e^(−2)(1 + 2) ≈ 0.5940, while the Poisson comparison gives 1 − e^(−1) ≈ 0.6321. More strikingly, P(N ≥ 2) = 1 − e^(−2)[1 + 2 + 2²/2! + 2³/3!] ≈ 0.1429, whereas the Poisson process gives 1 − 2e^(−1) ≈ 0.2642. Equality of mean waiting time does not equalize the finite-horizon event-count tail. The two-stage waiting law discourages closely spaced successive arrivals relative to exponential waiting.
Nonstationary and conditioned renewals
An ordinary renewal process begins at a renewal epoch. An equilibrium or stationary renewal construction has a different first-gap distribution, often called the equilibrium excess-life distribution: for finite mean μ_W, P(W_forward > x) = (1/μ_W)∫_x^∞ [1−F(u)]du. Treating a portfolio observed halfway through an unrecorded renewal cycle as if every insured item had just been reset creates bias. A condition-dependent hazard model can attach covariates to gap distributions, but when event times also change covariates a recurrent-event survival model may be preferable to an unqualified iid renewal assumption.
Renewal alternatives should be compared on the operational question. If the central objective is predicting time to the next repeated machinery failure after repair, a duration model offers interpretable parameters. If the objective is measuring aggregate natural catastrophe arrival across many territories, events may have overlapping initiating mechanisms and common shocks that invalidate independent individual renewal gaps. The Markovian arrival-process literature, including Neuts (1979), provides a more general tractable class containing several correlated or phase-type arrival mechanisms.
Marked and Compound Processes
From occurrences to monetary claims
A marked point process associates each occurrence time T_i with a mark Y_i. A mark can contain a vector: contractual loss size, coverage type, geographic cell, report delay, notification channel, sequences of payments, reserve updates, and settlement states. Under a simple compound Poisson model with N(T) independent of iid nonnegative claim severities X_i, aggregate gross losses are S(T) = Σ_{i=1}^{N(T)} X_i. Where E[X] and E[X²] exist, E[S(T)] = λT E[X] and Var[S(T)] = λT E[X²]. The second expression includes both uncertainty in the number of claims and uncertainty in their severities. It is not equal to E[N]Var[X] unless E[X] = 0.
For example, with expected occurrence count three per year, mean severity 5,000, the second severity moment is 15,000 and Var[S] = 3 × 50,000,000 = 150,000,000 monetary units squared; the standard deviation is approximately 16,202. Counting dependence changes the monetary risk even when severity is held fixed. These values are synthetic and are not price or capital recommendations.
Such independent compound-sum assumptions are often violated when a single weather event generates many linked claims, when severity changes with catastrophe state, or when reporting delay increases with complexity or payment size. In these cases conditional marks, event-level clusters, or multivariate point processes preserve the actual dependence. The separate registered dossiers on Aggregate Loss and Compound Risk Models and Loss Severity and Heavy-Tailed Distributions address aggregation and tail behavior in greater depth.
Reporting delay and Poisson thinning
Let original loss occurrences form an NHPP with intensity λ(s). Give each occurrence at s an independent nonnegative reporting delay D with conditional survival S_D(u|s) = P(D > u | occurrence at s), independent across events given the occurrence process and the specified covariates. At valuation time T, an incurred claim remains unreported exactly when s ≤ T < s + D. Poisson marking and thinning imply that the number of incurred-but-not-reported events generated during [0,T] is Poisson with mean B(T) = ∫_0^T λ(s) S_D(T−s|s)ds. The reported portion of the same occurrence cohort has mean Λ(T)−B(T). For the independently marked Poisson model, the two portions are independent; the conclusion need not hold under a shared latent intensity or dependent reporting delays.
A concrete example has constant occurrence intensity 100 claims per month during a twelve-month exposure period and independent exponential reporting delay with mean two months. At the twelve-month valuation date, B(12) = 100∫_0^12 exp[−(12−s)/2]ds = 200(1−e^(−6)) ≈ 199.50 incurred-but-not-reported claims. Expected total occurrences equal 1,200 and expected observed reports from these occurrences equal approximately 1,000.50. A naive frequency estimate based solely on reports would confound the roughly 199.50-claim reporting backlog with the underlying occurrence hazard. This is a stylized fully known model: real IBNR prediction also must estimate exposure, occurrence intensity, delay distribution and their uncertainty.
If severity, propensity to report, or settlement is associated with the delay, the mean of outstanding payment amounts requires the joint distribution of loss size and development mark. Count thinning alone does not identify a reserve in currency. In Jewell's Bayesian delay model (1989), inference about unreported counts critically depends on the information observed about event dates and delay parameters. In Norberg's marked-process approach (1993), the distinction among settled, reported-unsettled, incurred-unreported, and future-covered claims organizes their predictive liabilities. The finer incidence and payment structure is developed by Haastrup and Arjas (1996) with an explicitly Bayesian continuous-time model.
The occurrence–reporting–payment timeline
A defensible observation architecture uses separate stochastic objects. Let N^O(t) count covered loss occurrences up to t, N^R(t) first reports, and N^P(t) payment transactions. A claim may generate zero, one, or several payments, so N^P is not generally a thinned copy of N^O. Nor does a claim-specific reserve update necessarily represent cash paid. The insurer's valuation system may combine current case estimates and payments; any stochastic model must state which quantity is observed, which is an estimate, and which is forecast. A practical record includes identifiers linking all transactions to their original event and contractual coverage, together with the full observation and update times.
Antonio and Plat's micro-level reserving study (2014) applies marked point-process and recurrent-event ideas to individual liability claims, with separate occurrence, report, payment and closure information. Their empirical demonstration supports a potential gain from detailed data in the studied portfolio, but does not establish universal superiority of a micro-model for every insurer. The comparison is with aggregation under particular business, information and validation conditions.
Claim development as state transitions
For an individual known occurrence, define an illustrative state X(t) from a finite set: 0, incurred but unreported; 1, reported and open; 2, closed; and 3, reopened. A continuous-time Markov chain may model transitions with generator Q(t) whose off-diagonal entries q_ij(t) are nonnegative and whose row sums are zero. State 0 can move to 1; open claims may move to closed; closed claims may reopen where contract and data permit. State 3 is unnecessary if reopening is modeled as a direct transition from 2 back to 1, but a distinct state can retain information about reopenings. The probability row vector p(t) solves p'(t) = p(t)Q(t) for a time-inhomogeneous generator. Payment events can be separate marked jumps whose intensity depends on claim state, age, cause, and severity.
This Markov specification assumes the current state and modeled covariates capture the information needed for future transitions. Duration since reopening or time since first report can invalidate the memoryless state assumption. A semi-Markov or multi-state duration model is then more appropriate. Discretionary administrative closures, portfolio transfer, or outsourced claims management can alter observed transition intensities independently of insured peril. Claims development models should therefore distinguish operational process changes from changes in underlying loss hazard.
Calendar-time and accident-time information
When claim data are aggregated into accident-year/development-year triangles, exact event times, within-period exposures, and relationships among repeated transactions can be lost. Wüthrich and Merz's stochastic reserving treatment (2008) presents reserving models and prediction uncertainty at this aggregate level. Mack's chain-ladder uncertainty analysis (1993) derives a distribution-free standard-error calculation under its model assumptions. Those are proper comparators and complementary validation baselines, not interchangeable stochastic representations of individual claim arrival times. A favorable micro-level likelihood can coexist with worse aggregate predictive performance when the detailed data are sparse, noisy, or administratively inconsistent.
Markovian Regime Dynamics
Random intensities and conditional independence
A Cox process, or doubly stochastic Poisson process, is constructed by first drawing a random nonnegative intensity path L(t) and then, conditionally on the path, taking a Poisson process with intensity L(t). This is different from merely estimating a deterministic λ(t) with uncertain coefficients. Conditional on a latent path, disjoint increments are independent; marginally they generally exhibit dependence because they share the random environment. Cox's 1955 event-series study anchors the statistical treatment of such varying event rates. Avanzi, Wong, and Yang's insurance Cox-process analysis (2016) explicitly combines changing exposure, delayed reporting and serially dependent overdispersed claim counts, with a real-data illustration. Such a process can separate latent risk environment from conditional arrival randomness provided assumptions on the latent process and observation mechanism are identifiable.
A gamma mixing example illustrates updateability. With Z ∼ Gamma(shape k, rate k) constant through a year and N|Z ∼ Poisson(μZ), the marginal count is negative binomial with E[N] = μ and Var[N] = μ + μ²/k. If k = 2 and μ = 4, observing n = 8 complete claims in the first window produces posterior Z|N=n ∼ Gamma(k+n, rate k+μ) under this model, with posterior mean (2+8)/(2+4) = 5/3. In a future equal-exposure, nonoverlapping window sharing the same latent Z, posterior expected count is 4×5/3 = 20/3 ≈ 6.67. That forecast requires a persistent shared latent state across the windows. If the frailty is redrawn independently for the future year, the future unconditional mean remains four and the preceding observation does not update its intensity.
Two-state Markov-modulated Poisson process
Let J(t) be a continuous-time Markov chain on states L and H with generator Q = [[−q, q], [q, −q]], q > 0, and stationary initial probabilities 1/2 each. Conditional on its path, claim arrivals are Poisson with rate λ_L = 1 or λ_H = 5 per year according to the current state. The stationary mean arrival rate is three per year. With a one-year horizon, E[N(1)] = 3. The stationary intensity variance is 4, and its autocovariance at lag u is Cov[λ(J(0)), λ(J(u))] = 4exp(−2qu). Consequently,
Var[N(T)] = 3T + 2∫_0^T (T−u)4e^(−2qu)du = 3T + 4T/q − 2[1−e^(−2qT)]/q².
At T = 1 and q = 1 per year, the variance is approximately 5.2707, exceeding the Poisson value 3. At q = 10 it is approximately 3.3800; rapid switching reduces the integrated latent-intensity variance. As q approaches zero, the process tends to a mixture of two fixed annual intensities, with variance approaching 3 + 4 = 7 at T = 1. Thus equal long-run mean frequency can hide materially different one-year variability depending on how persistent high-risk regimes are. Fischer and Meier-Hellstern's MMPP cookbook (1993) details the tractable correlated-arrival representation.
A Markov chain over environmental conditions may be appropriate for alternating weather regimes, operational vulnerability periods, or stable and stressed network states. It must be fitted with caution: a two-state latent chain is an approximation, not a claim that real catastrophe hazards have precisely two physical states. Inference can be weak when states have similar rates, transition times are rare, or reporting delay is mistaken for persistence of the hazard. Both arrival and observation models need validation.
Self-excitation and clusters
A linear Hawkes process specifies conditional intensity λ(t|H_t) = μ(t) + Σ_{t_i<t} g(t−t_i), with nonnegative excitation kernel g and an exogenous baseline μ(t). A past event increases the near-future rate if g is positive, with decay or delayed effects determined by the kernel. For a stationary homogeneous model under standard conditions, the kernel mass η = ∫_0^∞g(u)du must be below one for a finite stationary mean intensity, which equals μ/(1−η) for constant positive baseline μ. Hawkes's original 1971 analysis develops spectral and covariance properties, providing a framework for analyzing second-order statistical structure. Distinct event-generating mechanisms can nevertheless display similar second-order summaries, so those summaries alone cannot establish a causal excitation mechanism.
In insurance, an observed cluster may arise from a series of actual events, an unmeasured common hazard, delayed batch reporting, or a single catastrophe generating many policy-level claims. A Hawkes fit to notification times could mistake a catastrophe-induced claims-reporting campaign for physical contagion among accidents. A causal self-exciting interpretation requires event definition, risk mechanism, temporal ordering, and model checks beyond a fitted excitation coefficient. It is scientifically legitimate to use Hawkes intensity as a predictive device without asserting that one claimant's filing causes the next underlying loss.
Markovian event generators and state dependence
The general Markovian arrival process extends the MMPP by allowing phase changes that coincide with arrivals, and can represent richer correlations and interarrival structure. Neuts's 1979 construction is a foundational reference. A process for claim-arrival regimes and a Markov model for individual claim settlements solve different problems; they may be combined in a hierarchical model, but they cannot be conflated merely because both use finite-state chains. The first governs when new claims occur; the second governs how an existing claim develops. Their data requirements, filtration and forecast targets differ.
Time-Series Diagnostics
Define the observable and observation filtration
The first diagnostic is a data audit rather than a statistical fit. Confirm whether timestamps record accidents, policyholder notifications, claim registration, acceptance, transaction payments or closure. Verify currency, covered territory, duplicates, one-to-many event-to-claim relationships, reopening rules, changing claims codes, and recording precision. Reconstruct exposure at risk by coverage time. A system migration that posts multiple old claims on one date can mimic a self-exciting process in registered claims, while a newly extended reporting hotline can shorten notification delays without changing underlying loss incidence.
Reporting delayed claims induce a triangular observation window: at a valuation T, recent accident cohorts have short observable development, old cohorts have longer development, and very late claims can be missing. Claims still open at the extract date contribute censored time-to-settlement observations. Estimating delay distributions only from closed files selects faster settlements. Conditioning only on reported files selects shorter notification delays. An appropriate delayed-entry, censoring and competing-transition likelihood is essential before interpreting frequency or duration estimates.
Count dispersion and autocorrelation
For nonoverlapping equal-exposure blocks B_j, compare sample mean and variance of counts with the Poisson benchmark, and inspect count autocorrelation across time. A Fano factor above one can indicate excess dispersion, but is also produced by unmodeled exposure heterogeneity, seasonality, parameter drift, mixture of perils or event batching. Similarly, autocorrelation may be generated by a persistent Cox intensity or an administrative reporting cycle. Tests must correct for estimated trends, overdispersion, uneven exposure, and the number of lags examined. Choosing the most favorable lag after inspecting the data creates its own selection bias.
A Poisson generalized linear model using log exposure as an offset can establish a defensible baseline for covariates and calendar effects. Negative-binomial counts provide a practical marginal alternative where overdispersion is persistent, but their likelihood for independent monthly observations is not a replacement for a full continuous-time Cox model if cross-month dependence is a central target. A marked process model can explicitly separate claim severities and occurrence times; a multi-state model can separate report and settlement evolution. The best class is defined by the forecasting task and the data's ability to identify its additional structure.
Residual point processes and time transformation
For a correctly specified simple point process with continuous conditional compensator and suitably observed event times, residual spacings z_i = ∫{t{i−1}}^{t_i} λ̂(u | H_{u−})du are expected to resemble iid unit-rate exponential variables under the model, with qualifications when parameters are estimated and the last censored interval is included. The transformed cumulative occurrence times should resemble a unit-rate Poisson process. Ogata's residual analysis (1988) gives a classic likelihood-and-time-transformation framework, originally applied to earthquakes. For insurance data, the transformation must use the modeled occurrence or reporting process that matches the observation clock; transforming report dates with an occurrence intensity generally creates misleading residuals.
Practical checks include quantile plots of transformed waiting times, residual autocorrelation, simulated-envelope comparisons, seasonal error by exposure, probability integral transforms for block counts, predictive coverage for IBNR claims, and calibration of observed payment and closure transitions. Holdout windows should extend beyond the longest material reporting and development delays where the use case permits. Prediction of individual claim progression and prediction of ultimate aggregate liabilities should be scored separately.
Empirical evidence on delayed claims
Verrall and Wüthrich's reporting-delay analysis (2016) works with individual property and casualty claim records and accommodates trends and seasonal patterns. It directly illustrates why stationary delay assumptions can distort recent cohort predictions. Avanzi, Wong, and Yang (2016) test a Cox-process approach on Australian claims counts with exposure and reporting delay, while Antonio and Plat (2014) use additional individual payment and closure information in a European liability portfolio. These papers demonstrate model feasibility in particular datasets, and their insurance applications are stronger evidence for that feasibility than a purely generic point-process text. They do not warrant importing calibrated parameters or performance rankings to a different country's insurance book without local evidence.
Governance and external validity
A credible model record gives the definition of event and population, time window, exposure reconstruction, observation mechanism, likelihood, prior distributions where used, parameter uncertainty, sensitivity to tail and delay assumptions, and reproducibility conditions. The International Actuarial Association's ISAP standards repository provides the professional model-standard context. ISAP 1 is a model standard for adoption or adaptation by competent standard setters; it should not be described as automatically binding in every jurisdiction. In international research, methodological diligence includes transparent uncertainty communication and documentation of source and statistical limits, independently of local law.
Model comparison should use at least one simple baseline. A complex state-space, Hawkes or Cox model is justified by improved explanatory adequacy, predictive calibration and reproducibility under an explicitly defined target, after accounting for its increased estimation burden. A purely in-sample likelihood increase or an attractive visualization of clustered claims is insufficient. Unobserved intensity regimes and delayed reporting can mimic one another: sensitivity analysis across alternative data-generating stories is therefore a substantive scientific requirement.
Integrated Synthesis and Model Selection
A common mathematical representation
The processes studied can be organized without creating a new theoretical system. Their shared mathematical language is a time-indexed random event measure and a clearly stated information filtration. Poisson processes fix a deterministic intensity and independent increments. Renewal processes specify waiting-time laws. Cox processes randomize intensity independently of conditional Poisson arrivals. Markov-modulated processes place structured persistence in an exogenous state; Hawkes processes place event-history feedback in conditional intensity. Marks attach loss amounts and delays; claim-state models govern later development. The relationships are established distinctions in probability and actuarial science, assembled here to clarify insurance interpretation.
The same observed annual count mean can be produced by all of these model families. Their predictions diverge when conditioning on recent events, measuring zeros or clustered counts, predicting future waiting times, or estimating the stock of incurred but unreported claims. A one-number annual frequency has insufficient information to identify the process class. Event times and exposure histories improve identification; reporting delays and dataset truncation can remove that advantage. Claims forecasting is consequently a joint problem of event generation and observation.
Comparative model matrix
| Model class | Primary stochastic specification | Key identifiable feature under adequate data | Principal risk of misuse |
|---|---|---|---|
| Homogeneous Poisson | Constant arrival intensity | Exponential interarrivals; Poisson independent increments | Treating variable exposure as constant |
| Nonhomogeneous Poisson | Deterministic calendar/exposure intensity | Independent increments after trend adjustment | Attributing latent clustering to deterministic seasonality |
| Renewal | Iid gap distribution from renewal epochs | Duration dependence and nonexponential interarrivals | Assuming observed entry occurs at a renewal epoch |
| Mixed/Cox Poisson | Exogenous random intensity | Extra-Poisson variance; potentially correlated increments | Ignoring uncertain or delayed ascertainment |
| Markov-modulated Poisson | Latent Markov-state arrival rate | Regime persistence and count correlation | Overinterpreting weakly identified regimes |
| Hawkes | Intensity feedback from past observed events | Short-horizon clustering conditional on history | Mistaking common shocks or batch reporting for causation |
| Marked claim process | Time-stamped events with severity and development marks | Dependence of claim count, size and delay | Treating payments as separate losses |
| Multi-state claim development | Transition intensities for claim statuses | Settlement, reopening and duration dependence | Enforcing a memoryless state where claim age matters |
The matrix is a guide to questions, not a universal ranking. Insurance data often mix multiple generating processes. For example, a local exposure-adjusted NHPP can govern ordinary claims, a separate event-level catastrophe process can generate clusters of policy claims, and delayed reporting can govern visibility into both. Combining components requires estimating or bounding their dependencies; a decorative collection of individually fitted models does not automatically produce a coherent predictive distribution.
Identification hierarchy and inference limits
The first inference target is the well-defined observable: claims registered on a date, for example. The second is the occurrence process that is only partially observed. The third is a counterfactual causal mechanism, such as whether the occurrence of one event changes the hazard of another. These targets require progressively stronger assumptions and evidence. The spectral results of Hawkes (1971) describe properties of particular self-exciting point processes. Neither variance-to-mean ratios nor autocorrelation alone uniquely identify a Cox, renewal, or self-exciting mechanism.
The primary analytic result of this monograph is a conditional interpretation of prediction: the insured occurrence process, the reporting observation model, and the development/payment model must be separately specified before they can be joined into a forecast. That synthesis follows directly from the mathematical decomposition and the cited insurance studies. Its value is methodological coherence and transparent conditions of use; it does not claim priority for a new theorem, classification or stochastic doctrine.
Research Methods and Reproducible Calculations
Research procedure and evidence grading
The investigation drew on primary publisher and DOI records for mathematical point-process sources, original actuarial studies, and an international professional standard repository. Peer-reviewed abstracts and publisher bibliographic records establish precise publication identity and the high-level scope of studies; mathematical formulas in this manuscript are derived from declared assumptions and checked independently rather than inferred from abstracts. The theoretical examples use synthetic inputs and explicitly display each transformation. No confidential claims database, individualized actuarial liability calculation, novel experiment on insurer data, or external peer review is represented as performed.
Source hierarchy is mathematical derivation and primary paper for the core model; insurer-specific empirical articles for claim-data applicability; original professional standards for governance. The publication date displayed by a modern electronic journal platform may be a digitization date; References use the original issue year where verified. Author-year references in the text link to the same exact source as the bibliography, and every external research URL is recorded in the corresponding EVIDENCE.json. The scope of verification for each source, including full text versus abstract/metadata availability, is documented there. An original analytic synthesis is distinguished from a newly discovered scientific result.
Deterministic verification ledger
The following calculations use standard exponential, gamma, Poisson, renewal and covariance identities. Every parameter is fully stated and values are rounded only after calculation.
| Construction | Exact expression | Computed value |
|---|---|---|
| Poisson zero probability, mean 4 | e^(−4) | 0.0183156 |
| Gamma-mixed Poisson zero probability, shape 2 and mean 4 | (2/6)² | 0.1111111 |
| Gamma-mixed Poisson variance, mean 4 and shape 2 | 4 + 4²/2 | 12 |
| Erlang(2, rate 2) first arrival by year 1 | 1 − 3e^(−2) | 0.5939942 |
| Erlang renewal at least two arrivals by year 1 | 1 − e^(−2)(1 + 2 + 2 + 8/6) | 0.1428765 |
| Poisson(rate 1) at least two arrivals by year 1 | 1 − 2e^(−1) | 0.2642411 |
| One-year exposure-adjusted Poisson expected counts | 0.02(1000×0.5 + 2000×0.5) | 30 |
| Reporting backlog, rate 100/month, exponential mean delay 2 months | 200(1 − e^(−6)) | 199.50425 |
| MMPP one-year variance with q = 1 | 3 + 4 − 2(1 − e^(−2)) | 5.2706706 |
| MMPP one-year variance with q = 10 | 3 + 0.4 − 0.02(1 − e^(−20)) | 3.3800000 |
| Posterior future expected count under shared gamma frailty | 4(2+8)/(2+4) | 6.6666667 |
To reproduce these examples, use IEEE-754 double precision or arbitrary precision with e = exp(1), ordinary gamma and Poisson moment identities, and the published formulas. Simulations are unnecessary for these particular exact calculations; stochastic Monte Carlo would introduce an additional convergence issue. A production actuarial application would require independent code review, data rights, access controls, reserve-purpose validation and local professional requirements. None of the illustrative values is a calibrated insurance rate.
Substantive QA findings and boundaries
The mathematical distinctions were explicitly checked for: (i) the difference between unconditional and conditional intensity; (ii) the difference between Cox and history-dependent intensities; (iii) the renewal initial-condition problem; (iv) the necessity of joint marks for severity and reporting-delay dependence; (v) the independence claim in Poisson thinning being conditional on its stated assumptions; (vi) the separation of occurrence, report, payment and closure clocks; (vii) stable stationarity assumptions in two-state Markov modulation; and (viii) the exact difference between martingale compensators and Poisson increment laws. The exposure, overdispersion, renewal, MMPP, severity and IBNR examples are all independently calculable from printed inputs.
Material empirical limits remain visible: none of the cited real-data papers licenses transferring its fitted parameters to a new portfolio; source access for some archival articles is limited to publisher bibliographic record and abstract; and claim recognition is governed by jurisdiction-specific contracts and reporting conventions that are outside this mathematical monograph. These are limits of generalization, not defects deferred to the publication stage. The present release packet requires an intact bibliography, contextual source-link QA, two matching SHA-256 fingerprints for the reviewed manuscript and one fingerprint for the checked evidence ledger.
Conclusion
Stochastic claim processes provide the temporal foundation of insurance risk measurement. Their scientific object is an event history whose stochastic structure changes when occurrence, notification, payment and settlement are separately observed. The Poisson model offers a tractable baseline with independent increments and explicit exposure integration. Renewal processes reveal how the waiting-time law alters short-horizon prediction even under identical long-run frequency. Cox and Markov-modulated models represent persistent heterogeneity and overdispersion; Hawkes models describe event-history excitation while demanding particular care before any causal interpretation. Marked processes and multi-state development models connect occurrence events to the liabilities that insurers eventually observe and pay.
The original worked calculations establish several concrete consequences. Equal mean claim counts can entail different variances and zero-event probabilities; regimes with identical stationary mean intensity can generate different finite-horizon count variability; and known occurrence frequency can coexist with a substantial unreported-claim inventory solely because notifications arrive late. These differences matter for forecasting, reserving, exposure control, and evaluation of model uncertainty. They are consequences of stated model assumptions, verified algebraically and linked to established actuarial research, rather than claims of a newly invented mathematical framework.
The principal methodological conclusion is that every predictive claim about insurance losses must identify its event clock, its at-risk population, its conditional information, and its observation process. The scientific choice among Poisson, renewal, Cox, Markovian and self-exciting structures follows the forecast target, the data-generating mechanism, diagnostic evidence, and stability under perturbation. Source-aware interpretation and independent numerical verification complete the model's scientific argument. This Research Dossier supplies a unified, bounded analytical treatment for Risk Theory and its related Science areas; the exact published identity and release timing remain matters for the separate authorized publication process.
References
Aalen, Odd. 1978. “Nonparametric Inference for a Family of Counting Processes.” The Annals of Statistics 6 (4): 701–726. https://doi.org/10.1214/aos/1176344247
Antonio, Katrien, and Richard Plat. 2014. “Micro-Level Stochastic Loss Reserving for General Insurance.” Scandinavian Actuarial Journal 2014 (7): 649–669. https://doi.org/10.1080/03461238.2012.755938
Arjas, Elja. 1989. “The Claims Reserving Problem in Non-Life Insurance: Some Structural Ideas.” ASTIN Bulletin 19 (2): 139–152. https://doi.org/10.2143/AST.19.2.2014905
Avanzi, Benjamin, Bernard Wong, and Xinda Yang. 2016. “A Micro-Level Claim Count Model with Overdispersion and Reporting Delays.” Insurance: Mathematics and Economics 71: 1–14. https://doi.org/10.1016/j.insmatheco.2016.07.002
Cox, D. R. 1955. “Some Statistical Methods Connected with Series of Events.” Journal of the Royal Statistical Society, Series B (Methodological) 17 (2): 129–157. https://doi.org/10.1111/j.2517-6161.1955.tb00188.x
Daley, D. J., and David Vere-Jones. 2003. An Introduction to the Theory of Point Processes. Volume I: Elementary Theory and Methods. 2nd ed. New York: Springer. https://doi.org/10.1007/b97277
Fischer, Wolfgang, and Kathleen S. Meier-Hellstern. 1993. “The Markov-Modulated Poisson Process (MMPP) Cookbook.” Performance Evaluation 18 (2): 149–171. https://www.sciencedirect.com/science/article/pii/016653169390035S
Haastrup, Svend, and Elja Arjas. 1996. “Claims Reserving in Continuous Time: A Nonparametric Bayesian Approach.” ASTIN Bulletin 26 (2): 139–164. https://doi.org/10.2143/AST.26.2.563216
Hawkes, Alan G. 1971. “Spectra of Some Self-Exciting and Mutually Exciting Point Processes.” Biometrika 58 (1): 83–90. https://doi.org/10.1093/biomet/58.1.83
International Actuarial Association. 2018. ISAP 1: General Actuarial Practice (model standard; current standards repository checked November 14, 2026). https://actuaries.org/publications/international-standards-of-actuarial-practices/
Jewell, William S. 1989. “Predicting IBNYR Events and Delays: I. Continuous Time.” ASTIN Bulletin 19 (1): 25–55. https://doi.org/10.2143/AST.19.1.2014914
Mack, Thomas. 1993. “Distribution-Free Calculation of the Standard Error of Chain Ladder Reserve Estimates.” ASTIN Bulletin 23 (2): 213–225. https://doi.org/10.2143/AST.23.2.2005092
Neuts, Marcel F. 1979. “A Versatile Markovian Point Process.” Journal of Applied Probability 16 (4): 764–779. https://doi.org/10.2307/3213143
Norberg, Ragnar. 1993. “Prediction of Outstanding Liabilities in Non-Life Insurance.” ASTIN Bulletin 23 (1): 95–115. https://doi.org/10.2143/AST.23.1.2005103
Ogata, Yosihiko. 1988. “Statistical Models for Earthquake Occurrences and Residual Analysis for Point Processes.” Journal of the American Statistical Association 83 (401): 9–27. https://doi.org/10.1080/01621459.1988.10478560
Verrall, Richard J., and Mario V. Wüthrich. 2016. “Understanding Reporting Delay in General Insurance.” Risks 4 (3): 25. https://doi.org/10.3390/risks4030025
Wüthrich, Mario V., and Michael Merz. 2008. Stochastic Claims Reserving Methods in Insurance. Chichester: Wiley. https://www.wiley.com/en-us/stochastic-claims-reserving-methods-in-insurance-p-9780470723463