Machine Learning for High-Frequency Order Book Trading

Summary

This article outlines a high-frequency strategy research framework for China’s order-driven markets, using Level 1 order book data and support vector machines (SVMs). It proposes extracting features such as best bid and ask prices and sizes, derived depth and spread measures, volume, open interest, and technical indicators. Short-horizon midpoint price changes are grouped into directional classes, with larger changes treated as candidate trading opportunities. The system design links incoming quote events, feature construction, model training and validation, and trade decisions.

The empirical example uses an index futures contract’s October data. The article compares predictions across several tick horizons and reports classification accuracy reaching about 70%, suggesting that accuracy near 60% could be usable for a strategy. A one-day simulation reports trades, win rate, and net profit after assumed fees, alongside substantial modeled slippage. These results are limited: the sample period is short, the simulation assumes unrestricted trade counts and a fixed slippage estimate, and execution quality is explicitly critical. The reported model accuracy alone does not establish durable, live profitability.

Key ideas

• The proposed system predicts short-term direction from order book features using an SVM classifier.
• Features include best quotes and sizes, derived spread and depth measures, and market activity variables.
• The method labels midpoint movements over short tick horizons to define directional opportunities.
• The example reports model accuracy and a one-day simulated result, but relies on a short historical sample.
• Slippage and execution quality can materially change the simulated outcome.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.