A Deep Reinforcement Learning Approach Combining Technical and Fundamental Analyses with a Large Language Model for Stock Trading

Stock trading strategies are essential for successful investment, yet developing a profitable approach is challenging due to the stock market's complex and dynamic nature. This paper introduces a Deep Reinforcement Learning (DRL) framework for automated stock trading that integrates technical and fundamental analyses with a large language model. We model the trading environment as a Partially Observable Markov Decision Process (POMDP) and propose a hybrid architecture that combines Long Short-Term Memory (LSTM) with Proximal Policy Optimization (PPO) to capture intricate temporal patterns in stock data and make informed trading decisions. Our model in-corporates market indicators alongside financial news headlines, processed through the FinBERT language model, to create a rich state representation. Additionally, we integrate a drawdown penalty into the reward function to further improve portfolio stability. Evaluations on a dataset of 30 U.S. stocks demonstrate that our model outperforms benchmarks in cumulative return, maximum earning rate, and Sharpe ratio, indicating that the hybrid approach yields more resilient and profitable trading strategies than existing methods.