The application of reinforcement learning and signal processing in dynamic investment management

Abstract

An innovative approach is adopted to develop a rigorous active portfolio management system that explicitly makes investment decisions and processes financial market information. This approach addresses two unique challenges in portfolio management: the ability to effectuate market-sensitive asset allocations and alleviate the effects of financial market uncertainty. These challenges are resolved by utilising Recurrent Reinforcement Learning (RRL) as a sequential decision-making tool. Additionally, signal processing is employed to enhance performance stability. The study proposes the Augmented Recurrent Reinforcement Learning (ARRL), a hybrid portfolio management system that integrates the RRL and signal processing modules. Using shares from nine of South Africa’s primary economic sectors, the ARRL system is used to perform dynamic asset allocation, thereby taking advantage of the changes in the market opportunity set. The performance of the system is compared to standard passive portfolio management strategies. ARRL-based strategies outperform standard passive strategies by a wide margin, according to the findings