Empirical analysis of Q learning, DQN and DDQN for intraday trading

The objective of this study is to enhance intraday trading strategies in the Indian stock market, specifically for the NIFTY 50 index, through the application of Reinforcement Learning (RL). This study explores how RL techniques can be leveraged for trading strategy optimization by estimating empirical Q-values for individual stocks. The experimental framework proposed in this study comprises of different stages namely data collection and preprocessing, environment design, model selection, training and testing, followed by performance evaluation. Five years historical data was obtained via the API of a leading Indian brokerage platform, which was later cleaned and enriched with selected technical indicators. We have compared the effectiveness and efficiency of three RL algorithms: Q-Learning (QL), Deep Q-Network (DQN), and Double Deep Q-Network (DDQN) in the context of stock trading. Performance of the algorithms is evaluated using key metrics such as cumulative returns, maximum drawdown, win-loss ratio, Sharpe ratio, total trades, and overall profitability. Experimental results show that the DDQN consistently outperforms QL and DQN, highlighting its robustness under realistic trading conditions. The insights derived from the experiments conducted in this study contribute to financial machine learning research and offers practical recommendations for the design of algorithmic trading systems.