Dynamic portfolio optimization using reinforcement learning

Abstract

This study uses machine learning in the development of a dynamic investment strategy for portfolio optimization. We aim to explore the efficiency of this approach over a passively managed portfolio and assess the whether transaction costs erode the gains in the dynamically managed portfolio. To this end we explore the application of recurrent reinforcement learning for optimal asset allocation of a portfolio consisting of stock prices for six companies in different sectors. We develop an environment based on monthly historic prices of these stocks and a re-balancing agent that acts on the environment. The risk and return factors of the individual stock are taken as the state of the environment. Using a modified version of Sterling Ratio as the performance measure, we select model parameters through direct recurrent reinforcement learning from historical data and test the efficacy of the strategy on unseen data. From the analysis we find that the regularly re-balanced portfolio out performs the market portfolio based on buy and hold strategy based on both the terminal wealth and the risk adjusted return measure.