Machine Learning for High-Frequency Signals and Execution Costs
Summary
This discussion distinguishes machine learning for execution optimization from machine learning for directional trading. Execution begins with a known trade and seeks to reduce its cost; alpha generation must identify price direction and capture enough of the move to pay spreads and other costs. It highlights the difficulty of turning granular order, trade, and cancellation data into meaningful features, and describes features such as spread, recent price movement, weighted midpoint, trade direction, book imbalance, and signed trade volume.
The cited study reconstructed order books for 19 stocks, trained on 2008 data, and evaluated policies on 2009 data. Under idealized midpoint execution, learned strategies showed positive test results, with broadly similar directional relationships across stocks. The document cautions that this assumption is unrealistic and that predicted moves were smaller than the spread, so the results do not establish executable profits. It proposes longer holding periods, limit orders with fill and adverse-selection models, or better features as possible research directions.
Key ideas
- Execution optimization assumes trade direction and size are known, while alpha generation must predict direction and overcome trading costs.
- Order-book data pose both scale challenges and difficulty in identifying informative features.
- The study used six microstructure features and trained on one year before testing on the next.
- Midpoint-execution tests showed positive results, but that idealized assumption does not demonstrate real-world profitability.
- A relative, mean-adjusted action value can help separate state-specific effects from broad price drift.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.