Publication

Abstract : Deep Reinforcement Learning (DRL) techniques are gaining strong interest in developing stock trading strategies, given their ability to learn profitable behaviors in complex environments. This paper presents a comparative study evaluating deep RL algorithms, including DQN, Double DQN (DDQN), Dueling Double DQN (Dueling DDQN), and Rainbow DQN, on financial stock trading problems based on historical price data. The performance of each algorithm is analyzed using metrics such as cumulative rewards earned, value function stability measured through Q-values, and loss over the training period, indicating sample efficiency. We benchmark extensions like Rainbow DQN that integrate multiple optimization mechanisms against conventional DQN methods. We discuss the capabilities of these state-of-the-art techniques along with current limitations and challenges that need to be addressed through further research. The results on out-of-sample datasets over validation periods provide guidelines for real-world applicability. The insights from this empirical analysis can inform the design of improved stock trading agents using deep reinforcement learning.