Quantitative Finance and Economics | 2026
Authors: Lee K.
DOI: 10.3934/QFE.2026005
Journal: Quantitative Finance and Economics
Year: 2026
Publisher: American Institute of Mathematical Sciences
Document Type: Article
Open Access: All Open Access; Gold Open Access
Cited by: 0
This paper examines the extent to which standardized financial statement disclosures can anticipate both overall environmental, social, and governance (ESG) ratings and their evolution over time for listed firms. Utilizing a longitudinal dataset of 851 companies spanning 11 years, we systematically extract 279 International Financial Reporting Standards (IFRS)-compliant accounts and generate an extensive set of financial ratios to capture the static and dynamic features of corporate performance. Through an empirical comparison of traditional machine learning, gradient boosting (LightGBM, XGBoost), and neural network methods (multilayer perceptrons, convolutional neural networks, and TabNet), the analysis finds that boosting algorithms deliver consistently superior accuracy in ESG prediction tasks involving high-dimensional tabular data. Notably, the overall ESG composite score exhibits the highest level of predictability from financial information, while environmental ratings remain more elusive. Further investigation reveals that balance sheet variables most strongly explain absolute ESG levels, whereas cash flow metrics are pivotal in predicting annual changes in the scores. These findings indicate that both corporate financial structures and resource flows encode substantial amounts of information that is relevant to sustainability evaluations. By linking financial accounting theory with ESG analytics, this study provides a rigorous, data-driven framework that offers practical insights for researchers, policy-makers, and market participants seeking to enhance the reliability and timeliness of ESG evaluations using objective financial data. © 2026 the Author(s)
deep learning; ESG; ESG prediction; financial disclosure; machine learning