Applications of metaheuristics and multi-Agent techniques in financial portfolio optimisation

Abstract

Portfolio Optimisation (PO) is a challenging topic in computational finance for balancing returns and risks through asset allocation. While existing PO methods can be insufficient due to the complex conditions of financial markets, there are recent advances in metaheuristics and multi-agent systems that may show some promising studies and results in relevant fields of studies. Therefore, this thesis is targeted to investigate both techniques for PO under various market conditions and trading strategies.

First, the PO with expected asset returns is a Continuous Optimisation Problem (COP). Yet existing metaheuristic algorithms lack the efficiency for solving complex COPs. Accordingly, an Adaptive and Guided Differential Evolution (AdaGuiDE) framework is proposed where a novel guiding mechanism guides search strategies towards promising regions, while a systematic adaptive scheme enhances the trade-off between exploration and exploitation in solving complex COPs. The experiments on three benchmark datasets of COPs in evolutionary computation demonstrate that the AdaGuiDE outperforms state-of-the-art metaheuristic algorithms in tackling high-dimensional COPs with complex features.

Second, selecting promising assets in a portfolio from hundreds of assets for diversifying risks is important yet often ignored by some existing PO approaches. Besides, many metaheuristic algorithms neglect to integrate domain knowledge in specific areas. Therefore, an enhanced AdaGuiDE framework namely the AdaGuiDE+ is proposed to reduce extreme risks in PO with expected returns. During asset selection, a customised large language model with retrieval-augmented generation selects high-potential assets after analysing multimodal financial data. Subsequently, the allocation of preselected assets is optimised by the AdaGuiDE+ incorporating the specific domain knowledge. The results on three financial datasets show that the AdaGuiDE+ effectively minimises extreme risks while obtaining high returns at different confidence levels. Additionally, the AdaGuiDE+ demonstrates outstanding performance against the state-of-the-art metaheuristic approaches in handling two sets of practical tasks.

Third, many Deep Reinforcement Learning (DRL)-based algorithms are developed to optimise portfolios by analysing financial data when expected asset returns are unknown. Yet these methods are focused on long-term returns and neglecting short-term risks. A Multi-Agent and Self-Adaptive (MASA) framework is proposed where the DRL-based agent predicts asset allocation for maximising long-term profits while the solver-based agent rectifies the suggested portfolios to fulfil short-term risk requirements. Besides, a market observer provides some additional market information to enhance the adaptability of other agents. The experiments reveal that the MASA achieves higher returns and lower short-term risks than many existing PO approaches. Furthermore, an extended MASAC framework is proposed to optimise portfolios with multiple trading requirements. The results obtained on two well-known financial datasets show the advantages of the MASAC against some latest PO approaches both in trading performance and constraint satisfaction.

Lastly, an integrative framework named the MASAC+ incorporates the AdaGuiDE+ as the constraint solver into the MASAC for handling various trading requirements. The MASAC+ attains a better trade-off between return maximisation and constraint satisfaction than the MASAC. Besides, this work compares the ability of the AdaGuiDE+ and MASAC frameworks to optimise portfolios in the PO model with expected returns that highlights the advantages of the AdaGuiDE+ framework.