Machine Learning for Market Microstructure and High Frequency Trading
This paper considers the applications of machine learning when applied to high-frequency trading (HFT) using three studies to solve/optimize a specific trading problem involved with HFT. The cases look to determine price movement and optimize execution within a standard exchange and within dark pools. This study varies from existing literature on this topic by using microstructure data to learn real feature-based control policies.
The first case examines using reinforcement learning to optimize execution of purchasing v shares in t trading steps at the lowest share price. Using the RL methodology to optimize execution, choices for states, actions, impacts and rewards are applied to microstructure data for liquid equities and evaluated the execution performance on an order book. The profitability of the strategies learned by RL are tested and analyzed.
The second case shows the potential of machine learning on execution shortfalls and in effect reducing trading costs. This study considers models that reduce trading costs by actually learning when to trade rather than just reducing costs when the trade is provided to the model. This model is designed to permit reliable prediction of directional provide movements, and its profits are analyzed. This study found that learning produces profitable policies that are consistent and similar among products tested on.
The last section of the paper discusses using RL to apply smart order routing (SOR) in dark pools to divide orders across a specified venue, as dark pools offer many prices via many venues at once. The paper offers a brief discussion of the mechanics of trade execution within a dark pool, and explains how the algorithm works. They found with each additional round of allocation, the results improve rapidly. They conclude the most powerful application of machine learning in trading is for optimization via historical data.