Paper: arXiv 2509.11420

Authors: Yijia Xiao, Edward Sun, Tong Chen, Fang Wu, Di Luo, Wei Wang

Abstract

Developing professional, structured reasoning on par with human financial analysts and traders remains a central challenge in AI for finance, where markets demand interpretability and trust. Traditional time-series models lack explainability, while LLMs face challenges in turning natural-language analysis into disciplined, executable trades. Although reasoning LLMs have advanced in step-by-step planning and verification, their application to risk-sensitive financial decisions is underexplored. We present Trading-R1, a financially-aware model that incorporates strategic thinking and planning for comprehensive thesis composition, facts-grounded analysis, and volatility-adjusted decision making. Trading-R1 aligns reasoning with trading principles through supervised fine-tuning and reinforcement learning with a three-stage easy-to-hard curriculum. Training uses Tauric-TR1-DB, a 100k-sample corpus spanning 18 months, 14 equities, and five heterogeneous financial data sources. Evaluated on six major equities and ETFs, Trading-R1 demonstrates improved risk-adjusted returns and lower drawdowns compared to both open-source and proprietary instruction-following models as well as reasoning models. The system generates structured, evidence-based investment theses that support disciplined and interpretable trading decisions. Trading-R1 Terminal will be released at https://github.com/TauricResearch/Trading-R1.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.0)

Why this score: The paper focuses on LLM fine-tuning and RL techniques, which involve advanced ML methods but minimal heavy mathematical derivations or novel theoretical finance models. It demonstrates empirical rigor through a curated 100k-sample dataset, multi-source heterogeneous data, and evaluation on six equities/ETFs with reported risk-adjusted returns and drawdowns, though it lacks explicit backtest code or detailed statistical metrics in the excerpt.

Research Flowchart

  flowchart TD
  A["Research Goal: Develop Financially-Aware Reasoning LLM for Trading"] --> B["Data: Tauric-TR1-DB<br>100k samples, 14 equities, 18 months"]
  B --> C["Methodology: 3-Stage RL Curriculum<br>Easy to Hard Reasoning Tasks"]
  C --> D["Training: Supervised Fine-tuning &<br>Reinforcement Learning Alignment"]
  D --> E["Computational Process: Thesis Composition &<br>Volatility-Adjusted Decision Making"]
  E --> F["Outcome: Improved Risk-Adjusted Returns<br>vs. Baselines"]
  F --> G["Result: Structured, Evidence-Based<br>Investment Theses"]
  G --> H["Release: Trading-R1 Terminal"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes Trading-R1, a financially-aware LLM that integrates strategic reasoning, thesis composition, facts‑grounded analysis, and volatility‑adjusted decision making, trained via supervised fine‑tuning and reinforcement learning.

Source sentence
We present Trading-R1, a financially-aware model that incorporates strategic thinking and planning for comprehensive thesis composition, facts-grounded analysis, and volatility-adjusted decision making.

Assumes 2

  • Historical price and volume data are taken from Yahoo Finance with a 15‑day rolling window per trading day. data
    Source sentence
    We obtained historical price and volume data from the Yahoo Finance API. For each trading day, a 15-day rolling window was extracted containing open, high, low, close (OHLC) values and trading volume.
  • Trading decisions are evaluated with a neutral band and trading cost term κ in the outcome reward. execution
    Source sentence
    directionally correct calls earn a magnitude- and intensity-calibrated fraction of b and pay trading cost κ

Data

Backtest sample: 2024-06 – 2024-08 · six major equities and ETFs · equities · daily

  • Tauric-TR1-DB text-news public
    Jan 1 2024 – May 31 2025 · 14 major equities
    Source sentence
    Training uses Tauric-TR1-DB, a 100k-sample corpus spanning 18 months, 14 equities, and five heterogeneous financial data sources.
  • Yahoo Finance price data · Yahoo Finance market-prices public
    Source sentence
    We obtained historical price and volume data from the Yahoo Finance API.
  • Technical indicators (stockstats) · stockstats library alternative public
    Source sentence
    In addition to raw prices, we computed a suite of widely used technical indicators using the stockstats library.

Recognised sources: Yahoo Finance FRED

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Not modeled

Paper text extraction: Partial - A trading cost term κ is included in the outcome reward to encourage calibrated selectivity.

Source sentence (the authors' words)
directionally correct calls earn a magnitude- and intensity-calibrated fraction of b and pay trading cost κ

Finds 3

Reported performance (Trading-R1): Sharpe 1.88 · annual return 8.1% · max drawdown 3.7%

  • Trading-R1 achieves a Sharpe ratio of 1.88 with 8.08% returns on NVDA and outperforms GPT‑4.1 on AAPL with a Sharpe of 1.80 versus 1.24, while maintaining lower drawdowns (3.68% vs. 7.88%). performance
    Source sentence
    It achieves a Sharpe ratio of 1.88 with 8.08% returns on NVDA, and outperforms GPT-4.1 on AAPL with a Sharpe of 1.80 versus 1.24, while maintaining lower drawdowns (3.68% vs. RLMs’ 7.88%).
  • Trading-R1 attains leading hit ratios, including 70.0% on NVDA and 64.0% on SPY. performance
    Source sentence
    The model also attains leading hit ratios, including 70.0% on NVDA and 64.0% on SPY.
  • Across all evaluated assets, Trading-R1 delivers improvements over baseline models, achieving the strongest overall performance among SLM, RLM, LLM, Trading‑SFT, and Trading‑RFT. performance
    Source sentence
    Our Trading-R1 approach achieves the strongest overall performance by combining SFT and RFT to capture market dynamics effectively.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 30 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 6 papers in the archive (6 by independent authors)

Code

https://github.com/TauricResearch/Trading-R1 — the authors' own repository

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.