Department of Industrial Engineering

Permanent URI for this community

https://scholar.sun.ac.za/handle/10019.1/104

Browse

Browsing Department of Industrial Engineering by browse.metadata.advisor "Andries, Engelbrecht"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Incremental reinforcement learning for portfolio optimisation.
    (Stellenbosch : Stellenbosch University, 2023-03) Refiloe, Shabe; Andries, Engelbrecht; Stellenbosch University. Faculty of Engineering. Dept. of Industrial Engineering.
    ENGLISH ABSTRACT: Portfolio optimisation is a decision-making problem that involves allocation of a certain fund across different financial assets, with the objective of maximising profit and minimising risk, simultaneously. Portfolio optimisation is a difficult problem to analyse. There is a wide range of research on various portfolio optimisation approaches in finance and computational intelligence. The two fields overlap. Thus, the use of meta-heuristics to make intelligent investment decisions is a result of the intersection of finance and computational intelligence. Meta-heuristics formulated the portfolio optimisation problem as a static optimisation problem and successfully obtained optimal portfolios. However, in the real world, investment decision-making is a dynamic problem that involves daily trading. Therefore, it is more representative of real-world investments to formulate the portfolio optimisation problem as a dynamic optimisation problem. This thesis explores a reinforcement learning approach to formulate a dynamic investment strategy. The concept of reinforcement learning has improved the development of multistage stochastic optimisation; a primary component in sequential portfolio optimisation. A recurrent form of a reinforcement learning algorithm called proximal policy optimisation (PPO), that allocates portfolios based on historic asset prices is presented. The results provide a conclusive support for the ability of PPO to identify good-quality portfolios. The results also show that the strategy becomes outdated overtime as it fails to perform as well during the COVID-19 pandemic. Based on this finding, the recurrent PPO approach was improved in order to take into account the presence of concept drift caused by pandemics and potential financial contagions. The approach was adapted to incrementally learn the financial market as the portfolio optimisation process takes place. The incremental recurrent PPO algorithm is shown to be able to adapt to drastic changes in the market and obtain optimal portfolios.