Automated Trading Framework Using LLM-Driven Features and Deep Reinforcement Learning

Abstract

Stock trading faces significant challenges due to market volatility and the complexity of integrating diverse data sources, such as financial texts and numerical market data. This paper proposes an innovative automated trading system that integrates advanced natural language processing (NLP) and deep reinforcement learning (DRL) to address these challenges. The system combines two novel components: PrimoGPT, a Transformer-based NLP model fine-tuned on financial texts using instruction-based datasets to generate actionable features like sentiment and trend direction, and PrimoRL, a DRL model that expands its state space with these NLP-derived features for enhanced decision-making precision compared to traditional DRL models like FinRL. An experimental evaluation over seven months of leading technology stocks reveals cumulative returns of up to 58.47% for individual stocks and 27.14% for a diversified portfolio, with a Sharpe ratio of 1.70, outperforming traditional and advanced benchmarks. This work advances AI-driven quantitative finance by offering a scalable framework that bridges qualitative analysis and strategic action, thereby fostering smarter and more equitable participation in financial markets.