Reproduce “AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework” on OpenResearch

About this paper

Xiang Li, Zhenyu Li, Chen Shi, Yong Xu, Qing Du, Mingkui Tan, Jun Huang, Wei Lin

The paper introduces the AlphaFin dataset and the Stock-Chain framework, a two-stage Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) approach to enhance Large Language Models for financial analysis. The framework achieved 55.7% accuracy and a 30.8% Annualized Rate of Return in stock trend prediction, also demonstrating superior performance in financial question answering compared to other LLMs by integrating real-time data.

Key insights

• Combining Retrieval-Augmented Generation (RAG) for real-time data and Chain-of-Thought (CoT) reasoning improves LLM accuracy and interpretability in financial analysis.
• Fine-tuning Large Language Models on specialized, high-quality financial datasets, including CoT examples, significantly enhances their domain-specific capabilities.
• A continuously updated vector database allows LLMs to overcome the issue of outdated information prevalent in pre-trained models, providing current financial insights.

Summary by alphaXiv.