Generating investment policies for nonlinear portfolio optimization with genetic programming

In this paper we present a method for portfolio optimization based on policy optimization with nonlinear optimization methods. The investment policies are generated automatically using genetic programming and special care has been taken to generate policies that can be differentiated automatically. This system is compared with a traditional approach using only genetic programming, and their performance is evaluated with statistical analysis. Both systems take a wide range of technical analysis methods as input and we apply them on five different portfolios composed of stocks from the Stockholm Stock Exchange. We find that the policy optimization system yields a higher return on average than the genetic programming system, but that it's hard to state with statistical significance. We also see that the policy optimization system produces less complex strategies that appear to generalize better than the genetic programming system, even though every strategy requires longer time to evaluate.