Researcher: Peiran Zhao

Introduction

Bitcoin trades often arrive in bursts. This project models that clustering and explores how the estimated rate of trade arrivals can inform volatility forecasting, option pricing, hedging, and execution.

Methodology

The researcher fit a Hawkes process by maximum likelihood to 80,000 BTC–USD trades from Coinbase in May 2026. In the model, each trade temporarily raises the estimated rate of subsequent trades, and that effect decays over time. The researcher used the fitted intensity to build a volatility forecast and compared it with GARCH(1,1) and HAR-RV on five-minute intervals.

Fitted Hawkes parameters alongside a trade intensity curve that jumps after trades and decays toward its baseline
Figure 8. Fitted Hawkes parameters and the estimated trade-arrival intensity over a two-second window.

Results & Analysis

The Hawkes-based forecast achieved 66.3% directional accuracy, ahead of GARCH at 55.6% and behind HAR-RV at 72.2%. HAR-RV also had the lowest mean squared error in the out-of-sample comparison. The fitted Hawkes branching ratio was about 0.185, below the stationarity threshold of 1.

Six panels showing Bitcoin price, realized volatility, trade intensity, return distribution, and forecast comparisons for GARCH, HAR-RV, and Hawkes-Vol
Figure 10. Bitcoin trading data and volatility forecast performance across 19 held-out five-minute intervals. Select the image to read the panel labels at full size.

Conclusion

The fitted Hawkes process gives the study a way to quantify bursts in trading activity. The report also uses that intensity in option pricing and reinforcement-learning experiments for hedging, execution, and market making.