Dynamic Portfolio Management Using Reinforcement Learning and Deep Reinforcement Learning Approaches

The application of machine learning techniques in finance is a rapidly evolving research frontier. Reinforcement learning (RL) presents a particularly promising framework for tackling dynamic and complex decision-making problems. This thesis explores the application of three RL algorithms, specifically SARSA, Q-learning and Deep Q-Learning, to the problem of dynamic portfolio management. A dataset containing 25 years of historical financial data for companies listed in the EUROSTOXX50 index is used as the empirical foundation for this research. The study investigates how RL agents can approximate value functions and optimize asset allocation across multiple assets over time.