Forecasting Bitcoin price movements using multivariate Hawkes processes and limit order book data

Abstract Forecasting short-term returns of Bitcoin is a key challenge in high-frequency trading, due to the cryptocurrency’s extreme volatility, market microstructure complexity, and non-stationary behavior. Limit Order Book (LOB) data offer a rich source of high-resolution information that can improve predictive models beyond what is possible using price series alone. In this study, we investigate the BTC/USD trading pair and develop two return sign forecasting approaches based on multivariate Hawkes processes (MHP), leveraging LOB event streams collected in real time from a centralized exchange. Our first method integrates an MHP with a Continuous-time Output Error (COE) model to jointly model event timing and return dynamics in irregularly sampled data. The second approach directly forecasts return sign and timing via an extended MHP. Empirical results show that the hybrid MHP–COE pipeline consistently outperforms the pure Hawkes-based model in both prediction accuracy and simulated trading profitability. We also evaluate the methods’ computational efficiency to assess their viability in real-world high-frequency environments.