HyFin: An LLM-Driven Platform for Hybrid Text-and-Table Financial Report Analysis
Runze Tian#, Jingyuan Xue#, Zihan Yang, Yifu Tang, Liuyi Chen, Yiming Huang, Jianke Yu, Bocheng Han, Yvonne Fang, Jiahui Jin, Zhengyi Yang*
RAIDS Lab Authors
Details
Research Area
Tags
Resources
Abstract
Annual reports describe a company's businesses, performance, risks, and future plans. However, this information is distributed across textual content and financial tables, making analysis and verification difficult at scale. Manual analysis is labor-intensive, and large language models may generate hallucinations. We present HyFin, a financial report analysis system for investment research that provides a complete pipeline from report processing to analysis. HyFin organizes annual report analysis around industry exposure, risk exposure, and strategic outlook and future prospects. It retrieves and fuses text and table data through separate evidence graphs while preserving provenance to the source reports. On actual annual-report questions, HyFin achieves an average answer quality score of 0.858, compared with 0.674 for Plain Dense RAG and 0.798 for the best reported competition system, corresponding to relative improvements of 27.3% and 7.5%, respectively. Case studies further show that the industry and risk signals produced by HyFin are consistent with independent market and financial evidence. Together, these capabilities support interactive and traceable investment research.