Learning to optimize profits beats predicting returns -: comparing techniques for financial portfolio optimisation

Stock selection for hedge fund portfolios is a challenging problem that has previously been tackled by many machine-learning, genetic and evolutionary systems, including both Genetic Programming (GP) and Support Vector Machines (SVM). But which is the better? We provide a head-to-head evaluation of GP and SVM applied to this real-world problem, including both a standard comparison of returns on investment and a comparison of both techniques when extended with a "voting" mechanism designed to improve both returns and robustness to volatile markets. Robustness is an important additional dimension to this comparison, since the markets (the environment in which the GP or SVM solution must survive) are dynamic and unpredictable. Our investigation highlights a key difference in the two techniques, showing the superiority of the GP approach for this problem.