Towards Smarter Trading: An AI Trading Framework Combining Reinforcement Learning and Large Language Model

Abstract

The rapid evolution of financial markets demands intelligent trading systems capable of synthesizing heterogeneous information and making adaptive decisions under conditions of un certainty. In this paper, we propose a trading framework that leverages reinforcement learn ing (RL) to fine-tune a Large Language Model (LLM) for autonomous trade decision-making. Unlike prior approaches that depend on supervised pre-training with expert-annotated anal yses or domain-specific corpora for cold-start guidance, our method applies Group Relative Policy Optimization (GRPO) directly to a general-purpose instruction-tuned LLM, enabling the model to develop trading competence purely through reward-driven exploration, without curated professional signals. The model ingests multi-source market observations — encompassing technical indicators, financial news, and corporate financial statements — within a rolling temporal window and outputs structured trading strategies specifying action type, share quantity, take-profit price, and stop-loss price. This formulation enforces strategy completeness through explicit exit conditions while also supporting flexible position sizing; bridging the gap between simplified academic models and practical trade execution. To guide learning, we design a multi-dimensional reward function grounded in profitabil ity and trading discipline. Each strategy is evaluated on path-dependent profit-and-loss, risk exposure relative to stop-loss levels, position sizing appropriateness, and regulatory adher ence; overall, this provides fine-grained feedback that cultivates the model’s awareness of both return potential and downside risk. We conduct comprehensive experiments along three dimensions: (1) model comparisons, contrasting the RL-trained LLM against its base model, alternative LLM architectures, and a DQN-based traditional RL trading agent to quantify improvements from RL fine-tuning and LLM-based reasoning respectively; (2) training budget analyses, investigating how train ing steps influence the model; and (3) strategy ablations, examining the contributions of quantity-based position sizing. Our results demonstrate that the proposed framework pro duces coherent, risk-aware trading strategies without supervised warm-up; the ablation anal yses further yield insights into the respective roles of model capacity, training sufficiency, and strategy design.