Financial Data Mining with Genetic Programming: a Survey and Look Forward
Automated Trader Magazine
Genetic Programming (GP) is an appealing machine-learning technique for tackling financial engineering problems: it belongs to the family of evolutionary algorithms that have proven to be remarkably successful at handling complex optimization problems, and possesses the unique feature of producing solutions under a symbolic form that can be understood and analyzed by humans. Over the last decade, GP has been applied to generate financial trading strategies, forecast stocks and options prices, or grasp some insight into the dynamics of the markets and the behavior of the agents. In this paper, we first provide a brief survey of the existing studies, then highlight fields of investigations that, we believe, should lead to enhance the applicability and efficiency of GP in the financial domain. By Nicolas NAVET and Shu-Heng CHEN
1 Relevance of GP for creating trading strategies
Genetic programming (GP) applies the idea of biological evolution to a society of computer programs. Specifically, in financial trading, each computer program represents a trading system - a decision rule - which when applied to the market provides trading recommendations. The society of computer programs evolves over the course of the successive generations until a termination criterion is fulfilled, usually a maximum number of generations or some property of the best individuals (e.g., stagnation for a certain number of generations, a minimum performance threshold is reached). Classical genetic operators, namely mutation, crossover and reproduction, are applied at each generation to a subset of individuals and the selection among the programs is biased towards the individuals that constitute the best solutions to the problem at hand.
In the 80s, economists began to be interested in the idea of evolving populations of decision rules1 because of the close similarity with the economic agents who are constantly revising - adapting - their own decision rules as they gain experience and as their environment undergo changes. Since then, evolutionary models have proved to be a powerful toolkit for modeling and understanding the behavior of societies of “imperfectly smart agents exploring their way into an essentially infinite space of possibilities” (in the words of J. Holland, see [Wal92]). In line with what has just been said, it is clear that evolutionary techniques, such as Genetic Algorithms and Genetic Programming, are relevant to serve as devices to generate financial trading rules, and indeed GP in particular has been already quite often used for that purpose2. A simple example of a typical trading rule is given in Figure 1.
...