RETuning: Upgrading Inference-Time Scaling for Stock Movement Prediction with Large Language Models

TL;DR

The paper tackles stock movement prediction with LLMs by diagnosing reliance on contextual cues and limited independent reasoning. It introduces Reflective Evidence Tuning (RETuning), a two-stage framework (SFT cold-start plus rule-based RL) that trains models to build and reflect on an evidence-driven analytical framework before predicting direction. A large, long-context Fin-2024 dataset with diverse information sources underpins the approach, enabling significant improvements in SMP and transfer to other financial tasks. Inference-time scaling via repeated sampling further demonstrates robustness, including under out-of-distribution conditions, suggesting practical potential for reasoning-driven LLMs in finance while acknowledging computational costs and domain specificity.

Abstract

Recently, large language models (LLMs) have demonstrated outstanding reasoning capabilities on mathematical and coding tasks. However, their application to financial tasks-especially the most fundamental task of stock movement prediction-remains underexplored. We study a three-class classification problem (up, hold, down) and, by analyzing existing reasoning responses, observe that: (1) LLMs follow analysts' opinions rather than exhibit a systematic, independent analytical logic (CoTs). (2) LLMs list summaries from different sources without weighing adversarial evidence, yet such counterevidence is crucial for reliable prediction. It shows that the model does not make good use of its reasoning ability to complete the task. To address this, we propose Reflective Evidence Tuning (RETuning), a cold-start method prior to reinforcement learning, to enhance prediction ability. While generating CoT, RETuning encourages dynamically constructing an analytical framework from diverse information sources, organizing and scoring evidence for price up or down based on that framework-rather than on contextual viewpoints-and finally reflecting to derive the prediction. This approach maximally aligns the model with its learned analytical framework, ensuring independent logical reasoning and reducing undue influence from context. We also build a large-scale dataset spanning all of 2024 for 5,123 A-share stocks, with long contexts (32K tokens) and over 200K samples. In addition to price and news, it incorporates analysts' opinions, quantitative reports, fundamental data, macroeconomic indicators, and similar stocks. Experiments show that RETuning successfully unlocks the model's reasoning ability in the financial domain. Inference-time scaling still works even after 6 months or on out-of-distribution stocks, since the models gain valuable insights about stock movement prediction.

Figures (38)

• Figure 1: Overview of data sample for one stock at trading day t.
• Figure 2: Existing LLMs without RETuning CoT perform no better than random guessing in stock movement prediction, and most (except Qwen3 32B) fail to scale at inference time.
• Figure 3: Abaltion on CoT prompting. The proposed CoT in Section \ref{['sec:method:1']} can help improve the prediction performance on Qwen3 32B, but not smaller models. Others lose to random guessing.
• Figure 4: Two-stage stock movement prediction model training framework: Stage 1 (Cold Start) uses multi-source data with human labeling and synthesis pipeline for Reflective Evidence Tuning; Stage 2 applies Reinforcement Fine-Tuning with curriculum learning, reward shaping, and inference-time scaling for final label determination.
• Figure 5: Difficulty distribution given by the cold-started model (DeepSeek_R1_14B_SFT).