The List / Reports, Filings, Accounting, and Risk / paper
Fin-R1: A Large Language Model for Financial Reasoning through Reinforcement Learning
Abstract
In recent years, general-purpose large language models (LLMs) such as GPT, Gemini, Claude, and DeepSeek have advanced at an unprecedented pace. Despite these achievements, their application to finance remains challenging, due to fragmented data sources, intransparent reasoning processes, and weak transferability to business applications. In response, we introduce Fin-R1, a reasoning LLM designed for financial scenarios. With a compact size of 7 billion parameters, Fin-R1 reduces deployment costs while addressing the aforementioned challenges. Its development follows a two-stage pipeline. First, we construct Fin-R1-Data, a high-quality financial dataset consisting of 60,091 chain-of-thought (CoT) samples, distilled and filtered from multiple authoritative benchmarks to ensure consistency and reliability. Second, we train Fin-R1 using Fin-R1-Data through supervised fine-tuning (SFT), followed by reinforcement learning (RL). This stage substantially improves the model's ability to solve complex financial reasoning tasks, yielding outputs that are both accurate and interpretable. Despite its relatively small parameter scale, Fin-R1 achieves competitive empirical performance across established financial benchmarks and demonstrates practical utility in compliance checking and robo-advisory. Our code is publicly available at https://github.com/SUFE-AIFLM-Lab/Fin-R1, and has already attracted over 700 stars.
In one sentence
Summary
The paper addresses the challenge of applying general-purpose large language models to finance, where fragmented data, opaque reasoning, and weak transferability hinder deployment. The authors propose Fin-R1, a compact 7-billion parameter model designed to enhance financial reasoning through a structured two-stage post-training pipeline. This approach aims to produce outputs that are not only accurate but also interpretable and compliant with regulatory standards, addressing the 'black box' nature of existing models. By focusing on reasoning capabilities rather than just text generation, the model seeks to integrate economic, legal, and quantitative logic effectively. The methodology begins with the construction of Fin-R1-Data, a bilingual dataset of 60,091 high-quality chain-of-thought samples. This data is distilled from authoritative benchmarks using DeepSeek-R1-671B and filtered using Qwen2.5-72B-Instruct as an LLM-as-a-Judge to ensure logical consistency and domain alignment. The model training involves supervised fine-tuning on this dataset to instill reasoning patterns, followed by Group Relative Policy Optimization (GRPO). GRPO utilizes format and accuracy reward functions to optimize the policy without requiring a separate value network, thereby improving computational efficiency and reasoning robustness across diverse financial tasks. Empirical results demonstrate that Fin-R1 achieves an average score of 75.2 on financial reasoning benchmarks, ranking second overall and outperforming other 7B-scale models by over 17 points. The model shows strong performance in numerical reasoning, sentiment analysis, and code generation for quantitative strategies. Practical applications include compliance checking and robo-advisory, where the explicit reasoning traces provide necessary transparency. However, the study notes limitations in handling highly evolving market environments and potential biases in the distilled data, suggesting that while effective for static reasoning tasks, further work is needed for dynamic real-time financial decision-making.
Research problem
General-purpose LLMs struggle in finance due to fragmented data sources, lack of transparent reasoning processes, and weak transferability to specific business applications like compliance and risk assessment. Existing models often fail to integrate complex economic, legal, and quantitative logic coherently, leading to unreliable outputs in high-stakes scenarios.
Core contributions
- Introduction of Fin-R1, a 7B-parameter reasoning LLM optimized for financial scenarios with reduced deployment costs.
- Construction of Fin-R1-Data, a curated bilingual dataset of 60,091 chain-of-thought samples distilled from multiple benchmarks.
- Development of a two-stage post-training pipeline combining Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO).
- Implementation of an LLM-as-a-Judge filtering mechanism using Qwen2.5-72B-Instruct to ensure data quality and logical consistency.
- Demonstration of competitive performance on financial benchmarks, outperforming other 7B-scale models significantly.
Method
- Data Distillation: Using DeepSeek-R1-671B to generate reasoning traces and answers from raw datasets.
- Data Filtering: Employing Qwen2.5-72B-Instruct as an LLM-as-a-Judge to check answer accuracy and reasoning quality across seven dimensions.
- Supervised Fine-Tuning (SFT): Training the base Qwen2.5-7B-Instruct model on the filtered Fin-R1-Data to learn reasoning patterns.
- Group Relative Policy Optimization (GRPO): Post-training with GRPO using format and accuracy rewards to optimize the policy without a value network.
- Reward Functions: Designing specific rewards for output format compliance and numerical/logical accuracy.
- Evaluation: Comparing model outputs against ground truth and human annotations for validity.
Results & findings
- Fin-R1 achieves an average score of 75.2 on financial reasoning benchmarks, ranking second overall.
- Outperforms existing state-of-the-art 7B-scale models by more than 17 points on average.
- Demonstrates strong capability in numerical reasoning, sentiment classification, and causal relationship extraction.
- Shows practical utility in compliance checking and robo-advisory applications with interpretable outputs.
- The GRPO training stage substantially improves reasoning accuracy and structural consistency compared to SFT alone.
- Qwen2.5-72B-Instruct proves superior to GPT-4o as a judge for data filtering in this context.
Data & experiments
- Fin-R1-Data: 60,091 bilingual samples covering advanced business knowledge, basic business knowledge, professional knowledge, and financial code.
- Data sources include FinQA, ConvFinQA, TFNS, FinCUGE, FinanceIQ, and proprietary exam questions from Shanghai University of Finance and Economics.
- Experiments evaluate performance on established financial reasoning benchmarks and specific business tasks like compliance checking.
- Comparative analysis against general-purpose LLMs and other domain-specific models of similar parameter scales.
- Ablation studies on data filtering methods and reward function designs in the GRPO stage.
Limitations & caveats
- The model's performance may degrade in highly dynamic or evolving market environments not covered in the static benchmarks.
- Reliance on distilled data from other LLMs may propagate biases or errors present in the source models.
- The GRPO training is limited to objective questions with verifiable solutions, potentially limiting generalization to open-ended subjective tasks.
- Computational efficiency is improved, but training still requires significant resources for the distillation and filtering stages.
- The study focuses on Chinese and English, limiting applicability to other languages without further adaptation.
Curator notes
Discovery status: round2_promoted. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework; DISC-FinLLM: A Chinese Financial Large Language Model based on Multiple Experts Fine-tuning; FinChain: A Symbolic Benchmark for Verifiable Chain-of-Thought Financial Reasoning; Open-FinLLMs: Open Multimodal Large Language Models for Financial Applications; Unlocking Data Value in Finance: A Study on Distillatio