ESG and Financial Data Extraction and Correlation Analysis Using Multi‐Agent LLMs on Bank Reports

Abstract

Assessing the relationship between environmental, social and governance (ESG) practices and financial performance has become a central concern in sustainable finance. However, existing studies largely depend on commercial ESG ratings, which lack transparency and obscure the contribution of individual indicators. In addition, ESG disclosures vary substantially across institutions, years and reporting formats, making cross‐bank comparison difficult. To address these challenges, this study introduces a unified multi‐agent framework powered by large language models (LLMs) to extract ESG and financial indicators directly from banks' annual reports. The framework employs category‐specific prompt designs that enable the models to recognise synonymous terminology, differing measurement scales and inconsistent reporting units, thereby transforming heterogeneous textual disclosures into structured and comparable datasets. Applying this LLM‐based approach to hundreds of reports from approximately 50 global banks over 7 years, we construct a harmonised dataset and conduct indicator‐level correlation analyses between ESG and financial performance measures. The results reveal heterogeneous yet systematic associations between specific ESG indicators and financial outcomes, with environmental variables showing relatively stronger and more consistent correlations with return on assets (ROA) and return on equity (ROE). This study presents the first scalable, transparent and reproducible LLM‐driven framework for ESG–finance analysis based on unstructured public disclosures, offering a methodological foundation that can be extended to other industries and domains.