Hawkes Process

A Hawkes process is a self-exciting point process: the conditional intensity of events at time t depends on the times of past events, λ(t) = μ + Σ φ(t − tᵢ) over events tᵢ < t, with a decaying excitation kernel φ. The arrival of one event raises the probability of further events for a short window, so events cluster in time; a multivariate Hawkes process adds cross-excitation (and empirically inhibition) between event types. It appears in this vault because limit-order-book events — orders, cancellations, trades — empirically cluster and excite one another, so Hawkes processes model real Limit Order Book flow far better than the memoryless Poisson arrivals assumed by Avellaneda-Stoikov 2008.

Crucially, a Hawkes-driven order book is non-Markovian: because intensity is a functional of the entire past event stream, the decision-relevant future cannot be recovered from a finite memoryless state. This is the precise mechanism by which the Markov property — the load-bearing assumption of every Markov chain, hidden Markov model and Markov Decision Process Trading Model — is empirically violated at the microstructure level. Lalor Swishchuk 2025 drive their market-making environment with semi-Markov and Hawkes jump-diffusion dynamics for exactly this reason.

Connections