Performance-weighted ensembles of random forests for predicting price impact

For any large player in financial markets, the impact of their trading activity represents a substantial proportion of transaction costs. This paper proposes a novel machine learning algorithm for predicting the price impact of order book events. Specifically, we introduce a prediction system based on ensembles of random forests (RFs). The system is trained and tested on depth-of-book data from the BATS and Chi-X exchanges and performance is benchmarked using ensembles of other popular regression algorithms including: linear regression, neural networks and support vector regression. The results show that recency-weighted ensembles of RFs produce over 15% greater prediction accuracy on out-of-sample data, for 5 out of 6 timeframes studied, compared with all benchmarks. Feature importance ranking is used to explore the significance of various market features on the price impact, finding them to be highly variable through time. Finally, a novel procedure for extracting the directional effects of features is proposed and used to explore the features most dominant in the price formation process.