Hawkes Processes for Order Flow
Order flow clusters. A buy market order makes another buy more likely soon after; cancelations often arrive in bursts; a price move can trigger both aggressive flow and quote replenishment. Independent Poisson arrivals cannot represent this behavior. Hawkes processes model it with intensities that increase after events and decay over time, making them a useful bridge between descriptive event data and executable microstructure models.
For event types \(i=1,\ldots,d\), a linear multivariate Hawkes intensity is
\[ \lambdai(t)=\mui+\sum{j=1}^d\sum{t^jk<t} \alpha{ij}\exp[-\beta{ij}(t-t^jk)]. \]
The baseline \(\mu_i\) is exogenous activity. Each past event of type \(j\) raises—or, in generalized models can inhibit—the near-term intensity of type \(i\). This lets trades, adds, cancelations, and price moves interact in event time.
Designing the event taxonomy
An overbroad taxonomy loses economic structure, while a high-dimensional one creates parameters that cannot be estimated reliably. Begin with trade sign and best-queue events, then add categories only if they improve an out-of-sample decision.
| Event type | Example | Useful excitation question |
|---|---|---|
| Buy market order | Trade lifts ask | Does it trigger more buys? |
| Sell market order | Trade hits bid | Does it trigger bid cancelations? |
| Ask cancelation | Best ask size removed | Does it precede upward price moves? |
| Bid add | New best-bid liquidity | Does it mean-revert sell pressure? |
| Midprice move | One-tick revision | Does it cause follow-through flow? |
Label events in exchange sequence order, with a clearly defined venue scope. Mixing consolidated trades with a single venue's quote updates creates artificial lead-lag structure. Book state and event normalization need the discipline described in market microstructure.
Branching interpretation and stability
For an exponential kernel, the integral \(\alpha{ij}/\beta{ij}\) is the expected number of direct type-\(i\) offspring from a type-\(j\) parent. The matrix of these integrals is the branching matrix \(K\). A stationary linear Hawkes process requires its spectral radius to be below one:
\[ \rho(K)<1. \]
The endogenous fraction is often summarized in a univariate model as \(\alpha/\beta\). A high value means much activity is internally propagated rather than explained by the baseline. It does not prove that traders are irrational or that the system is about to destabilize; unobserved common news can also appear as mutual excitation.
import numpy as np
def exponential_intensity(t, history, mu, alpha, beta):
age = t - np.asarray(history)
return mu + alpha * np.exp(-beta * age[age > 0]).sum()
def branching_ratio(alpha, beta):
return alpha / beta
For multivariate estimation, use a tested library or a constrained optimizer with nonnegative kernels and explicit stability checks. A short code snippet is not a production calibrator: likelihood gradients, observation boundaries, and numerical overflow matter materially.
Fitting without fooling yourself
The log likelihood includes each observed event's log intensity minus the integrated intensity over the sample. Include market open/close, halts, auctions, and feed gaps in the observation window. Treating a lunch break as zero event intensity makes the decay parameter absorb session boundaries.
Intraday seasonality is particularly important. A constant baseline interprets the predictable opening surge as self-excitation. Use a piecewise or smooth time-of-day baseline, and fit per liquidity regime where feasible. Test on later days, not randomly held-out events from the same burst.
| Check | Good sign | Failure interpretation |
|---|---|---|
| Time-rescaling residuals | Approximately unit exponential | Misspecified intensity |
| Simulated event counts | Match mean and dispersion | Missing clustering |
| Sign autocorrelation | Similar decay profile | Incorrect cross-kernels |
| Branching radius | Safely below one | Unstable fitted process |
| Out-of-sample log score | Beats seasonal Poisson baseline | No useful dependence |
Hawkes features for execution
The intensity itself can be a causal state feature. A burst of sell-market-order intensity accompanied by ask cancelations may increase the expected cost of a buy order. A buy program should not simply chase a rising buy intensity, however: doing so may join the same feedback loop and amplify footprint.
Combine intensities with order-flow imbalance signals, spread, and depth to predict short-horizon markouts or fill risk. For passive quoting, the valuable target is signed post-fill return. For aggressive execution, it is the distribution of cost conditional on waiting versus trading now.
Hawkes models are not full order-book simulators. They know an event was a cancellation, but not necessarily where it sat in the queue or how it affects your rank. Integrate them with a queue-aware model when simulating fills.
Key takeaways
- Hawkes processes represent clustered order flow through intensities excited by past events.
- Event definitions, venue consistency, and intraday baselines are more consequential than kernel sophistication.
- The branching matrix must be stable; a large endogenous fraction is not evidence of trader intent.
- Validate through time-rescaling, simulation, and forward-day likelihood, not in-sample fit alone.
- Use conditional intensities as execution-risk features alongside book state and post-fill markouts.