Deep Hawkes Process for High-Frequency Market Making
Deep Hawkes Process for High-Frequency Market Making
This paper introduces a novel high-frequency market making strategy based on the Deep Hawkes Process (DHP), which models limit order book events using a deep learning architecture that combines stacked denoising autoencoders (SDAE) with deep long short-term memory (DLSTM). The DHP captures self- and cross-excitation effects and feedback loops between order arrivals and the state of the limit order book, overcoming limitations of traditional Hawkes processes that assume linear, additive excitation. The authors develop a realistic multi-agent simulation framework with a matching engine and FIX protocol to train and test the market making agent. Experimental results show that the DHP-based agent outperforms a baseline strategy that uses a probability density estimate of the fundamental price, both in predicting order events and in trading profitability. The study also examines the impact of order cancellations on market quality and agent performance, confirming empirical findings that cancellations affect queue position, spread, and order value. The simulation framework successfully reproduces stylized facts from real order book data, demonstrating its validity for developing and evaluating high-frequency trading strategies.
Highlights
- 1First to incorporate Deep Hawkes Process (DHP) into high-frequency market making, enabling feedback loops between order arrival and LOB state with self- and cross-excitation.
- 2Develops a realistic multi-asset simulation framework with scalable architecture, matching engine, and FIX protocol for training and testing market making strategies.
- 3Proposes DLSTM-SDAE architecture that combines stacked denoising autoencoders with deep LSTM to handle noisy order book data and improve predictive performance over Neural Hawkes Process.
- 4Investigates the impact of order cancellations on queue position, profitability, spread, and order value, confirming empirical findings from prior literature.
- 5Outperforms baseline strategy using probability density estimate of fundamental price in both predictive accuracy and trading profitability.
Methods
- MDeep Hawkes Process (DHP): A neural point process model combining DLSTM with SDAE to capture self- and cross-excitation in order book events.
- MMulti-agent simulation framework: Realistic limit order book simulator with price-time priority, matching engine, and FIX protocol for high-frequency market making.
- MStacked Denoising Autoencoders (SDAE): Unsupervised pre-training to initialize DLSTM weights and filter noise from high-frequency order book data.
- MBenchmark comparison: Probability density estimate of fundamental price as baseline for market making strategy evaluation.
Results
- RDHP-based market making agent outperforms the baseline strategy in terms of profitability and predictive accuracy for order types and timing.
- ROrder cancellations significantly affect queue position, profitability, bid-ask spread, and order value, consistent with empirical studies.
- RThe simulation framework reproduces stylized facts from real limit order book data, validating its realism.
- RDLSTM-SDAE achieves better predictive performance than Neural Hawkes Process in forecasting next order event type and time.
- RThe feedback loop between order arrival and LOB state improves market making performance compared to models without such coupling.
Analyze Paper
Generate insights from "Deep Hawkes Process for High-Frequency Market Making".