Action-based Feature Representation for Reverse Engineering Trading Strategies

Stock trades now take place at unprecedented speeds. Aided by trading algorithms, market professionals can now buy and sell securities within milliseconds. These algorithms are capable of absorbing and then reacting to vast quantities of information, such as price movement, literally in the time it takes to blink an eye. Given this high speed, some trading algorithms do not take outside information, and in particular market news, into account. That is, trading algorithms are designed only to absorb endogenous data and thus tend to be deterministic. Furthermore, when provided with a constant or predictable type of information, these algorithms will always perform the same action.
To develop a model for incorporating variant trading algorithms, this study proposed to recover a subset of automated trading strategies derived from trade-level data. To determine the feasibility of this approach, we examined trading strategies used by participants during the Fall 2009 University of Virginia McIntire School of Commerce Hedge Fund Tournament. We determined that by using a variation of recursive partitioning it was in fact possible to recover trading strategies employed during the course of the tournament. This conclusion suggests that further research is warranted and provides justification for expanding the study to include trading strategies derived from real data.