Enhancing ESG Data Analysis in Green Banking Using Large Language Models

Abstract

Environmental, Social and Governance (ESG) data is an integral part of the sustainable finance landscape, enabling financial institutions to understand long-term risks, align portfolios with climate goals, and meet growing disclosure requirements from regulators. However, such data is often not available in structured formats. Instead, it is embedded in longer textual disclosures such as sustainability reports, corporate filings and other narrative disclosures. Traditional rule-based or keyword driven approaches are insufficient to deal with the complexity of the language, context, and idiosyncrasy of real-world ESG disclosures. This paper presents a three-layer LLM-based framework to automate the extraction of ESG signals from financial documents and their multi-class classification into environmental, social and governance categories, as well as the identification of greenwashing practices. The framework is developed based on fine-tuned transformer architectures, and a separate decision support layer integrates the extracted ESG signals in credit risk and green loan classification processes. Experiments on publicly available ESG datasets show important improvements in classification accuracy, precision, and F1-scores over rule-based baselines and domain-adapted BERT models, and the proposed system can be used to reliably identify discrepancies between narrative sustainability claims and quantitative performance indicators. It discusses model bias, explainability, and compliance with the EU Sustainable Finance Disclosure Regulation. It concludes that LLMs could become a transformative green banking analytics tool, but appropriate governance structures will require development to fully realize their potential.