In this paper we describe a stock market simulation in which market participants use genetic algorithms to gradually improve their trading strategies over time. A variety of experiments show that, under certain conditions, some market participants can make consistent profits over an extended period of time, a finding that might explain the success of some real-world money managers.
These experiments suggest a four parameter model of market participants. Each participant can be described along four dimensions: information set, constraint set, algorithm set, and model set. The information set captures what data the participant has access to (e.g., the participant has access to all historical price data). The constraint set describes under what restrictions the participant operates (e.g., the participant can borrow money at 1\% above the prime rate). The algorithm set indicates what programs the participant can use (e.g., the participant is restricted to hill-climbing optimization algorithms). The model set specifies the language which the participant employs to describe its findings (e.g., the participant uses stochastic differential equations). This four parameter model better explains the relative strengths and weaknesses of market participants than traditional financial or computational models.
We have applied some of the insights that we have gained from doing this and related research to our own trading accounts. We participated in the 1993 U.S. Investing Championships (options division) and finished with a 43.9\% return over a period of four months. To leverage this success we have formed a money management firm, called Redfire Capital Management Group, that employs evolutionary algorithms to create fully automated trading strategies for bond, currency, and equity markets.