The paper introduces RETuning, a two-stage framework to enhance LLMs for stock movement prediction (SMP). In the first stage, supervised fine-tuning (SFT) teaches the model to generate structured chain-of-thought reasoning: dynamically constructing an analytical framework from diverse information sources, scoring evidence for price up/down/hold, and reflecting before making a prediction. This cold-start approach addresses the observed issues that LLMs tend to parrot analysts' opinions and fail to weigh adversarial evidence. The second stage applies rule-based reinforcement learning (GRPO) with reward shaping (format, accuracy, consistency) and curriculum learning that focuses training on medium-difficulty samples, which are identified by the cold-start model's prediction consistency.

The authors also construct Fin-2024, a large-scale dataset covering 5,123 A-share stocks across all of 2024, with over 200K samples and 32K-token contexts integrating news, fundamentals, analyst opinions, quantitative reports, macroeconomic indicators, and similar stocks. Experiments show that RETuning significantly outperforms existing LLMs and specialized SMP models, achieving F1 scores of 0.4196 (14B) and 0.4071 (32B) on the December 2024 test set. The method enables inference-time scaling via majority voting, with monotonic improvements up to 8–16 repeated samples. RETuning also generalizes to out-of-distribution stocks, future dates (6 months later), and other financial tasks on BizFinBench, demonstrating broad transferability. The work establishes that structured reasoning fine-tuning and reinforcement learning can unlock LLMs' predictive capabilities in finance, paving the way for trustworthy, reasoning-driven financial AI.