The Poisson process is the workhorse of classical actuarial mathematics: events arrive independently at a constant average rate, and the past tells you nothing about the immediate future. It is elegant, tractable - and often wrong. Earthquakes trigger aftershocks. Cyber attacks spawn copycats. A liability judgment invites a wave of similar claims. Financial defaults propagate through counterparty networks. Real-world events cluster, and the Hawkes process is the natural mathematical language for that clustering.
The idea: events that excite more events
A Hawkes process is a self-exciting point process. Its defining feature is a conditional intensity - the instantaneous arrival rate of events - that jumps upward every time an event occurs and then decays back towards a baseline. Formally, the intensity at time t equals a baseline rate plus a sum of contributions from all past events, each contribution governed by an excitation kernel that fades with elapsed time, often exponentially.
Two parameters carry the intuition. The branching ratio measures how many further events each event triggers on average; below one, the process is stable, and as it approaches one, cascades become long and violent. The decay rate measures how quickly the excitement fades - fast decay produces short, sharp bursts, slow decay produces long-memory clustering.
Why actuaries should care
Ignoring clustering leads to systematic underestimation of tail risk. If claims arrive in bursts but are modelled as independent, the variance of aggregate losses is understated, reserves and capital are set too low, and reinsurance is bought against the wrong loss distribution. Hawkes-based frequency models correct this by making dependence explicit and measurable rather than an afterthought bolted on through correlation assumptions.
Applications across insurance and finance
In catastrophe settings, Hawkes processes describe earthquake aftershock sequences - seismology's ETAS model is a marked Hawkes process - and can represent clustering in severe weather outbreaks. In cyber insurance, self-excitation captures contagion: a successful exploit increases the short-term likelihood of further attacks across a portfolio of insureds sharing common vulnerabilities. In credit and operational risk, mutually exciting variants model how an event of one type raises the intensity of another, giving a natural framework for contagion between risk classes. In market microstructure, Hawkes models are standard for order flow, which is why financial engineers value them beyond insurance.
Calibration in practice
Maximum likelihood estimation is the standard route, and for exponential kernels the likelihood can be evaluated efficiently through a recursive formulation. Practical challenges deserve respect: parameter identifiability requires enough events; the choice of kernel shape matters for tail behaviour; and non-stationarity - a baseline rate that itself drifts - can masquerade as self-excitation. Goodness of fit is usually assessed through the time-rescaling theorem, which transforms event times of a correctly specified model into a unit-rate Poisson process that can be tested.
Conclusion
Hawkes processes bring the clustering that practitioners observe every day inside the model rather than leaving it in the residuals. For pricing, reserving and capital work involving contagious or cascading events - cyber especially - they are becoming an essential part of the modern actuarial toolkit, bridging actuarial science and financial engineering in exactly the territory where the two disciplines meet.