A machine learning approach to financial portfolio optimisation

Abstract

Investment bankers, speculators and investors all have one primary goal, that of maximising gains while simultaneously minimising risk. They typically have market positions across a number of financial instruments. These baskets of financial instruments are known as portfolios. Active research on portfolio optimisation can be traced back to the 1950s when Nobel prize winner Harry Markowitz published the highly cited work titled "Portfolio Selection". The dynamic, highly noisy, non-linear characteristics of financial time series makes them highly complex. Consequently, with the recent rapid advances in artificial intelligence and machine learning, their application to the financial domain has seen increased interest. Research into reinforcement learning has yielded a number of recent advancements in the field and has been applied to robotics, optimal control problems, warehouse management systems, the generation of music and visual artefacts, finance and other optimisation problems such as parameter tuning for machine learning algorithms. We propose a portfolio optimisation model based on the Advantage-Actor-Critic (A2C) reinforcement learning algorithm and investigate; the effectiveness of the proposed model in portfolio optimisation; the impact the timeliness of the training period on the performance of the model on an unseen test set; the performance of the resulting model versus an equally weighted buy-and-hold portfolio. We show that although the proposed solution is effective in optimising a portfolio, the benchmark outperforms it. Furthermore, we show that the timeliness of training validation-test schedule is crucial to the performance of the model.