Reader question: Financial events often arrive in bursts: one trade is followed by more trades, one cancellation by more cancellations, one default by further stress. How can an algorithm model the fact that an event itself temporarily raises the probability of future events?
A Hawkes process is a self-exciting point process. Instead of modelling only counts per minute or returns per day, it models the times of individual events and lets past events change the current event intensity.
This article owns the event-clustering problem: timestamped events + event types + an excitation kernel → conditional intensities, branching ratios, cross-excitation structure, likelihood diagnostics and forecasts of near-term event activity.
It does not own causal inference, personalised trading advice or the whole market-microstructure problem. A fitted Hawkes link says one event type helps predict another under the model; it does not automatically prove economic causation.
This is public mathematical and computational education.
1. Begin with a Poisson process
For a simple Poisson process with constant rate μ, the expected number of events in a short interval dt is:
μdt.
The event rate does not depend on what happened a millisecond ago or an hour ago.
That memoryless structure is often too simple for financial activity, where events cluster strongly.
2. The Hawkes conditional intensity
For a univariate Hawkes process, a standard conditional intensity is:
λ(t) = μ + Σtᵢ<t φ(t−tᵢ).
Here:
- μ is the baseline intensity;
- tᵢ are past event times;
- φ(s) is the excitation kernel;
- λ(t) is the instantaneous event rate conditional on history.
Every past event adds a temporary amount to today’s intensity according to how old that event is.
3. Exponential kernels
A common choice is:
φ(s) = αe−βs, s > 0.
The parameter α controls the initial excitation jump. The parameter β controls decay speed.
Immediately after an event, intensity rises. As time passes without further events, the contribution decays toward zero.
This creates bursts naturally.
4. The branching ratio
The total expected direct excitation caused by one event is the integral of the kernel:
η = ∫₀∞ φ(s)ds.
For the exponential kernel:
η = α/β.
In the branching interpretation, η is the expected number of direct offspring events generated by one event.
For a stationary linear Hawkes process, a key stability condition is:
η < 1.
As η approaches one, endogenous clustering becomes stronger and the process approaches criticality.
5. Baseline versus endogenous activity
The cluster representation developed by Hawkes and Oakes treats baseline events as immigrants and self-excited descendants as offspring.
Under the ideal stationary branching interpretation, the branching ratio gives a useful measure of how much event activity is endogenously generated by prior events rather than baseline arrivals.
But that interpretation depends on model specification. A misspecified baseline can make ordinary seasonality look like self-excitation.
6. Multivariate Hawkes processes
Financial systems have multiple event types: buy trades, sell trades, limit-order submissions, cancellations, price moves, defaults or alerts.
For d event types:
λᵢ(t) = μᵢ + Σⱼ Σtₖʲ<t φᵢⱼ(t−tₖʲ).
The kernel φᵢⱼ measures how an event of type j changes the future intensity of type i.
This creates a directed matrix of predictive excitation.
7. The kernel matrix and stability
Define the integrated kernel matrix:
Gᵢⱼ = ∫₀∞ φᵢⱼ(s)ds.
A standard stationarity condition for a linear multivariate Hawkes process is that the spectral radius of G is below one:
ρ(G) < 1.
This is a matrix generalisation of the univariate branching-ratio condition.
The largest eigenvalue therefore becomes a system-level diagnostic of excitation strength under the model.
8. Likelihood estimation
Given event times on interval [0,T], a point-process log-likelihood has the form:
ℓ = Σₖ log λ(tₖ) − ∫₀ᵀ λ(t)dt.
The first term rewards high intensity where events actually occurred. The second penalises models that predict excessive intensity over the whole interval.
Maximum-likelihood estimation chooses parameters such as μ, α, β to balance these two forces.
9. Why calibration can be computationally expensive
Naively, every event depends on every earlier event, creating quadratic work in the number of events.
For exponential kernels, recursive updates can reduce cost because the accumulated excitation can be updated rather than recomputed from scratch.
High-dimensional multivariate systems can still be expensive because every event type may interact with every other type.
Sparsity, regularisation and structured kernels are therefore computational as well as statistical choices.
10. Time-rescaling diagnostics
If the fitted conditional intensity is correct, transform event times using the cumulative intensity:
zₖ = ∫tₖ₋₁tₖ λ(s)ds.
Under a correctly specified model, these transformed inter-arrival increments should behave like independent Exponential(1) observations.
This gives several diagnostics:
- QQ plot against Exp(1);
- Kolmogorov–Smirnov test;
- serial-correlation test on transformed residuals;
- inspection of residual clustering.
A model that fits the raw event count but fails the time-rescaling test is not capturing the temporal structure correctly.
11. Intraday seasonality is a major weak link
Financial activity often has a U-shaped intraday pattern: high near the open, lower around midday and high again near the close.
If the baseline μ is forced to be constant, the model may interpret predictable clock-time seasonality as self-excitation.
Falsifier: replace the constant baseline with an intraday seasonal baseline. If the branching ratio collapses, the original “endogeneity” estimate was partly a seasonality artefact.
12. Timestamp resolution matters
At high frequency, multiple events can share the same recorded timestamp because the data system rounds or batches timestamps.
A continuous-time Hawkes likelihood assumes an ordering. Artificial ties can distort short-lag excitation.
Falsifier: repeat estimation at different timestamp resolutions or with a documented tie-handling rule. If short-horizon kernel mass changes sharply, data resolution is controlling the answer.
13. Power-law kernels
Exponential kernels decay quickly and produce a Markov representation. Financial event memory can decay more slowly.
Power-law or sums-of-exponentials kernels can capture longer memory. The cost is more parameters and more difficult identification.
A better in-sample fit is not automatically a better model. The question is whether out-of-sample event timing and residual diagnostics improve.
14. Inputs and outputs
Inputs can include:
- event timestamps;
- event labels;
- observation window;
- baseline-intensity specification;
- kernel family;
- parameter bounds;
- regularisation strength;
- timestamp-resolution rule;
- seasonality adjustment;
- initialisation and optimiser settings.
Outputs can include:
- baseline intensities;
- kernel parameters;
- branching ratios;
- integrated kernel matrix;
- spectral radius;
- log-likelihood;
- time-rescaling residuals;
- event-intensity forecasts;
- cross-excitation graph;
- uncertainty estimates.
15. Evidence polarity
Evidence for confidence includes stable parameter estimates, time-rescaled residuals close to independent Exp(1), good out-of-sample likelihood, branching ratios below the stability boundary, sensible decay scales, and cross-excitation patterns that survive data subsamples.
Evidence against confidence includes residual clustering, unstable kernels, spectral radius at or above one without a deliberate nonstationary interpretation, large changes after seasonality adjustment, implausible excitation concentrated at one timestamp bin, or cross-links that vanish out of sample.
16. Counterexample: Poisson process is enough
If events are independent after accounting for intraday seasonality, a Hawkes model may add complexity without predictive value.
Falsifier: compare out-of-sample log-likelihood and residual diagnostics with an inhomogeneous Poisson baseline. If Hawkes adds no material improvement, self-excitation is not supported.
17. Counterexample: common cause masquerades as excitation
Suppose a macro announcement causes many buy orders and sell orders at once. A multivariate Hawkes model may estimate strong cross-excitation because the events cluster, even though both are responses to an omitted external driver.
Falsifier: include known exogenous-event covariates or split announcement periods. If cross-excitation disappears, the link was predictive but not causal.
18. Counterexample: near-critical estimate from regime mixing
Combining calm and crisis periods into one stationary model can produce a high branching ratio.
Falsifier: fit separate regimes or rolling windows. If the near-critical estimate is a mixture artefact, the static parameter should not be interpreted as one stable property of the market.
19. Counterexample: overfitted multivariate network
With 30 event types, a full kernel matrix contains hundreds of directional interactions before kernel-shape parameters are counted.
Many weak links can fit noise.
Falsifier: use regularisation, holdout likelihood and stability selection. Links that do not survive these tests should not be presented as structural.
20. Counterexample: predictive link mistaken for manipulation
A sell-order event that strongly predicts cancellations does not prove manipulative intent. The relation may be ordinary market-making behaviour.
Authority boundary: Hawkes outputs can flag temporal structure for review; legal or compliance conclusions require broader evidence.
21. Alternatives
Inhomogeneous Poisson processes model time-varying baseline rates without self-excitation.
Autoregressive count models work on binned event counts rather than exact timestamps.
Cox processes use a stochastic latent intensity driven by exogenous randomness rather than direct event excitation.
Hidden Markov models represent discrete regimes instead of event-by-event excitation.
Neural point processes provide flexible intensity functions but can be harder to interpret and validate.
22. Connections to the surrounding Bukit Timah Tutor estate
Execution dynamics connect to Almgren–Chriss optimal execution.
Event-pattern detection connects to bank market-surveillance algorithms.
Network propagation connects conceptually to Eisenberg–Noe clearing algorithms, though Hawkes models event timing while Eisenberg–Noe solves balance-sheet clearing.
The full lane is indexed at Finance & Banking Algorithms | Applied Mathematics in Real Financial Systems.
23. What would falsify confidence?
Confidence should be withdrawn if time-rescaled residuals are not close to independent exponentials; if seasonality explains the estimated excitation; if cross-links disappear out of sample; if timestamp resolution controls the kernel; if stationarity conditions fail unintentionally; or if omitted external drivers explain the apparent event-to-event dependence.
24. Verification and update triggers
Preserve raw timestamps, event labels, baseline specification, kernel family, parameter estimates, optimiser status, residuals and holdout scores. Revalidate after exchange-feed changes, timestamp-format changes, market-structure changes, event-taxonomy changes, major regime shifts or substantial changes in event frequency.
25. Primary and high-quality references
- Alan G. Hawkes, Spectra of Some Self-Exciting and Mutually Exciting Point Processes, Biometrika, 1971.
- Alan G. Hawkes and David Oakes, A Cluster Process Representation of a Self-Exciting Process, Journal of Applied Probability.
- Emmanuel Bacry, Iacopo Mastromatteo and Jean-François Muzy, Hawkes Processes in Finance, survey of financial applications.
Educational boundary: Hawkes processes model conditional event intensity. A fitted excitation link is predictive structure under a model, not automatic evidence of economic causation, manipulation or future profit.