Action-based feature representation for reverse engineering trading strategies

This paper considers the problem of reverse engineering strategies for trading in the financial markets. We investigate this problem in the context of a trading tournament in which student teams used delta hedging and other mechanisms to attempt to achieve benchmark performance in managing a hedge fund in a simulated market. Our hypothesis is that machine learning models can be trained to solve the apprenticeship learning problem; that is, these models can learn to trade like tournament participants. After reviewing classical return-matching approaches and recent work in inverse reinforcement learning, we propose a supervised learning methodology that makes use of recursive partitioning (RP). Our proposed RP approach is based on a feature representation for actions that, we argue, corresponds to the information structures readily available to tournament participants. RP achieves high accuracy in predicting the type and scale of participant trades and in tracking overall portfolio performance. Our results suggest that further research on our proposed approach is warranted and should include an expansion to testing on data from real markets.