A Comparative Analysis of Deep Reinforcement Learning Techniques in Portfolio Optimization for Global Market Indexes

This paper explores the evolution of portfolio optimization driven by computational advancements like artificial intelligence and machine learning, focusing on the emerging paradigm of Deep Reinforcement Learning (DRL) for portfolio management. DRL, integrating deep neural networks with reinforcement learning, shows promise in adaptive decision-making. The study offers insights into DRL's finance applicability by comparing four key algorithms: Deep Q-Network (DQN), Dueling Double Deep Q-Network (D3QN), Proximal Policy Optimization (PPO), and Asynchronous Advantage Actor-Critic (A3C). These techniques are evaluated on varied financial datasets. Results demonstrate DRL's effectiveness in revealing patterns, capturing complex relationships, and adapting to market dynamics. The DRL algorithms consistently outperform traditional optimization methods lacking adaptive learning. The DRL-Portfolio framework, implemented using these algorithms, showcases dynamic portfolio optimization across stock indices. The paper suggests future research directions, including enhanced risk modeling, transaction cost optimization, and integration of alternative data sources, providing a nuanced view of DRL's potential in portfolio management and addressing areas for further innovation.