Trading-R1: Open-Source LLM Framework for Explainable Financial Trading

Background

Financial trading demands high explainability and reliability. Traditional time‑series models rely on handcrafted features and lack logical explanations, while generic large language models (LLMs) struggle to convert natural‑language analysis into executable trading decisions.

Challenges

Data integration difficulty: multimodal financial data (news, fundamentals, technical indicators) are noisy and hard for LLMs to fuse.

Decision reliability: hallucination leads to reasoning that deviates from real data.

Sparse training data: public financial data are fragmented and lack structured supervision signals.

Task mismatch: trading is a path‑dependent, uncertain process unlike the mathematical or programming tasks LLMs are usually optimized for.

Problem Definition

The paper asks how to enable LLMs to generate professional, structured investment arguments that incorporate technical analysis, fundamentals, news, etc.; how to translate LLM reasoning into executable trading decisions that adapt to market risk; and how to train LLMs on limited high‑quality financial data for multi‑stage reasoning.

Method

Data Construction (Tauric‑TR1‑DB)

The authors build a dataset of 100 k samples covering 14 large‑cap stocks (e.g., NVDA, AAPL) and 2 ETFs (SPY, QQQ) over 18 months (2024‑01 to 2025‑05). Each sample contains five modalities: technical indicators, fundamentals, news, sentiment, macro‑economic data. Samples are created by time‑bucket splitting (3 days, 4‑10 days, 11‑30 days), noise filtering via LLM relevance scoring, and random multimodal fusion, yielding inputs of 20‑30 k tokens and supervision signals consisting of an investment argument and a trade label.

Supervised Fine‑Tuning (SFT) – Reverse‑Reasoning Distillation

Proprietary LLMs (e.g., OpenAI o3‑mini) generate final trade suggestions from structured inputs. A lightweight LLM (e.g., GPT‑4.1‑nano) parses these suggestions to infer intermediate reasoning steps (technical contribution, fundamental contribution, etc.). The inferred steps are concatenated with multimodal evidence to form a structured investment argument, which becomes the supervision target for SFT.

Reinforcement Learning (RL) – Three‑Stage Curriculum

The model is trained sequentially through three stages:

Stage I (Structure): Reward the model for producing investment arguments with a standardized XML‑like layout (technical, fundamental, news sections).

Stage II (Argument): Reward the inclusion of direct citations from the input data (e.g., “revenue growth 15 %” must cite the earnings report) to reduce hallucination.

Stage III (Decision): Use volatility‑adjusted labels (strong buy, buy, hold, sell, strong sell) derived from multi‑window EMA returns and standardized volatility to reward decisions that align with market performance.

Key Techniques

Volatility‑adjusted label generation computes EMA returns over 3, 7, 15‑day windows, normalizes them with rolling volatility, and maps them to discrete labels using quantile thresholds (85 %, 53 %, 15 %, 3 %). Strategy optimization employs Group Relative Policy Optimization (GRPO), which scores new versus old policies by group‑wise advantage rather than a value function. The objective includes a KL‑penalty term β.

Experiments

Setup

Training uses the 100 k‑sample Tauric‑TR1‑DB. Testing covers six assets (AAPL, NVDA, AMZN, etc.) and two ETFs (SPY, QQQ) from 2024‑06 to 2024‑08. Baselines include small models (Qwen‑4B, GPT‑4.1‑nano), large models (GPT‑4.1, LLaMA‑3.3), and RL models (DeepSeek, O3‑mini). Evaluation metrics are cumulative return (CR), Sharpe ratio (SR), hit rate (HR), and maximum drawdown (MDD).

Overall Performance

Trading‑R1 achieves the highest Sharpe ratio and the lowest maximum drawdown across all assets, outperforming both open‑source and proprietary instruction‑following or inference‑only models.

Ablation Study

Removing SFT yields HR = 62.5 % (NVDA) and MDD = 2.73 %; removing RL yields HR = 45.7 % (NVDA) and MDD = 1.66 %; the full Trading‑R1 obtains HR = 70.0 % with MDD = 3.80 %, demonstrating the complementary benefit of structure and decision learning.

Interpretability

The generated arguments cite concrete evidence such as MACD crossovers, a gross margin of 68.7 %, or Azure cloud growth, providing traceable reasoning for each trade decision.