Robust technical trading strategies using GP for algorithmic portfolio selection

Abstract

This paper presents a Robust Genetic Programming approach for discovering profitable trading rules which are used to manage a portfolio of stocks from the Spanish market. The investigated method is used to determine potential buy and sell conditions for stocks, aiming to yield robust solutions able to withstand extreme market conditions, while producing high returns at a minimal risk. One of the biggest challenges GP evolved solutions face is over-fitting. GP trading rules need to have similar performance when tested with new data in order to be deployed in a real situation. We explore a random sampling method (RSFGP) which instead of calculating the fitness over the whole dataset, calculates it on randomly selected segments. This method shows improved robustness and out-of-sample results compared to standard genetic programming (SGP) and a volatility adjusted fitness (VAFGP). Trading strategies (TS) are evolved using financial metrics like the volatility, CAPM alpha and beta, and the Sharpe ratio alongside other Technical Indicators (TI) to find the best investment strategy. These strategies are evaluated using 21 of the most liquid stocks of the Spanish market. The achieved results clearly outperform Buy&Hold, SGP and VAFGP. Additionally, the solutions obtained with the training data during the experiments clearly show during testing robustness to step market declines as seen during the European sovereign debt crisis experienced recently in Spain. In this paper the solutions learned were able to operate for prolonged periods, which demonstrated the validity and robustness of the rules learned, which are able to operate continuously and with minimal human intervention. To sum up, the developed method is able to evolve TSs suitable for all market conditions with promising results, which suggests great potential in the method generalization capabilities.