Estimating Market Clustering with a Hawkes Process
Summary
The article builds a Hawkes-process workflow to measure whether large market moves tend to cluster. It converts price data into event times when absolute log returns exceed a multiple of a trailing return-volatility estimate, then fits an exponential-kernel process by maximum likelihood. The branching ratio summarizes the model’s estimated self-excitation, while conditional intensity provides a time-varying view of event clustering. The exponential kernel supports efficient recursive calculations, and the article describes validation through simulated event trains and an independent implementation cross-check.
A key design choice is measuring time in bars and using a relatively long trailing volatility window. The article reports that short windows can raise the threshold during a burst and suppress subsequent events, biasing the estimated clustering downward; accordingly, it advises reading the estimate mainly as a relative measure. The method is descriptive rather than predictive: it characterizes recent event timing and does not forecast the next move. Event definitions, window choices, and model assumptions limit how literally the fitted branching ratio should be interpreted.
Key ideas
- A Hawkes process models event timing in which past events temporarily raise the rate of future events.
- The branching ratio summarizes expected self-excitation, while conditional intensity tracks its changing level.
- Large moves are extracted when absolute log returns exceed a threshold based on prior volatility.
- A short volatility window can suppress clustered events by adapting too quickly to the burst.
- The fitted process describes recent clustering and should not be treated as a price forecast.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.