Trading-R1: Financial Trading with LLM Reasoning via Reinforcement Learning

In one sentence

Summary

The paper addresses the challenge of aligning large language models with the disciplined, interpretable reasoning required for professional financial trading. It argues that while general reasoning LLMs excel in math and coding, they struggle with the noise, uncertainty, and path-dependency of markets. The authors propose Trading-R1, a model designed to produce structured investment theses and actionable trade recommendations, bridging the gap between natural language analysis and executable trading strategies. The work positions itself against opaque quantitative models and ungrounded general-purpose LLMs, emphasizing the need for explainability and risk-aware decision-making in high-stakes financial environments. The methodology centers on Tauric-TR1-DB, a 100k-sample corpus spanning 18 months and 14 equities, integrating technical, fundamental, news, sentiment, and macro data. Training employs a three-stage easy-to-hard curriculum: first, supervised fine-tuning (SFT) structures the output format; second, SFT grounds claims in evidence using reverse reasoning distillation from proprietary models; third, reinforcement learning fine-tuning (RFT) aligns decisions with market outcomes. Labels are generated via a volatility-aware, multi-horizon discretization method mapping returns to five trading actions. The backbone is Qwen3-4B, optimized for reasoning tasks. Evaluated on six major equities and ETFs, Trading-R1 demonstrates improved risk-adjusted returns and lower drawdowns compared to open-source and proprietary instruction-following models. The system generates structured, evidence-based investment theses that support disciplined trading. Key innovations include reverse reasoning distillation to reconstruct hidden CoT traces and volatility-adjusted reward labeling. Limitations include reliance on synthetic reasoning traces and evaluation on a limited set of blue-chip assets, potentially restricting generalizability to smaller caps or different market regimes. The code and terminal will be released publicly.

Research problem

The paper addresses the difficulty of applying large language models to financial trading due to the need for interpretability, structured reasoning, and risk-aware decision-making. Traditional time-series models lack explainability, while general LLMs struggle to ground their inferences in financial contexts and produce traceable, executable trades. The core problem is developing a model that can mimic professional analyst reasoning, integrating heterogeneous data sources into coherent investment theses while making disciplined, volatility-adjusted trading decisions.

Core contributions

• Introduction of Trading-R1, a financial reasoning LLM tailored for trading-oriented reasoning and structured thesis composition.
• Creation of Tauric-TR1-DB, a 100k-sample corpus spanning 18 months, 14 equities, and five heterogeneous financial data sources.
• Development of reverse reasoning distillation to reconstruct hidden chain-of-thought traces from proprietary black-box models for SFT.
• Implementation of a three-stage easy-to-hard curriculum interleaving supervised fine-tuning and reinforcement learning.
• Design of a volatility-driven discretization method for generating reliable, multi-horizon trading labels for reinforcement learning.

Method

• Reverse reasoning distillation: Inferring reasoning steps from proprietary model outputs.
• Three-stage curriculum: Structure, Claims, Decision phases.
• Supervised fine-tuning (SFT) for structural and evidential priors.
• Reinforcement learning fine-tuning (RFT) with volatility-adjusted rewards.
• Volatility-driven discretization: Multi-horizon EMA returns normalized by volatility.
• Asymmetric quantile cutoffs for label generation to reflect market drift.
• XML-tagged formatting for consistent thesis structure.
• Self-augmentation with reject sampling to refine reasoning quality.

Results & findings

• Trading-R1 achieves improved risk-adjusted returns and lower drawdowns compared to baseline models.
• The model successfully generates structured, evidence-based investment theses with clear citations.
• Reverse reasoning distillation effectively creates high-quality training data without hosting massive models.
• Volatility-adjusted labeling provides a robust reward signal for reinforcement learning.
• The three-stage curriculum stabilizes training and prevents error compounding in intermediate reasoning steps.
• The system demonstrates generalization across different market conditions (bull and bear) within the training period.

Data & experiments

• Tauric-TR1-DB: 100k samples from Jan 2024 to May 2025, covering 14 major equities.
• Data sources include Finnhub, SimFin, Google News, stockstats, and macroeconomic indicators.
• Evaluation on six major equities and ETFs, comparing against open-source and proprietary instruction-following models.
• Metrics include risk-adjusted returns, drawdowns, and qualitative assessment of thesis structure and evidence grounding.
• Backbone model is Qwen3-4B, fine-tuned using LoRA and full parameter updates in later stages.

Limitations & caveats

• Evaluation limited to six major equities and ETFs, potentially lacking generalizability.
• Reliance on synthetic reasoning traces from proprietary models may introduce biases.
• Training period covers only 18 months, limiting exposure to diverse market regimes.
• Volatility-adjusted labels are proxies and may not capture all market nuances.
• Backbone model size (4B) may limit reasoning depth compared to larger models.
• No live trading evaluation, only backtesting on historical data.

Curator notes

Discovery status: trading_agent_focus_promoted. Source links: AlphaQuanter: An End-to-End Tool-Augmented Agentic Reinforcement Learning Framework for Stock Trading; Design and Empirical Study of a Large Language Model-Based Multi-Agent Investment System for Chinese Public REITs