Predicting, Using, and Assessing ESG Signals: A Tripartite Systematic Review of Machine Learning in Sustainable Finance

Abstract

Environmental, Social, and Governance (ESG) ratings increasingly shape capital allocation, corporate strategy, and regulatory oversight, yet their credibility is constrained by methodological opacity, rating divergence, and greenwashing risk. Prior reviews treat machine learning (ML) in ESG as a prediction problem. We identify an emerging research trajectory in which ML is increasingly used not only to consume ESG signals but also to verify their construction and credibility. Drawing on signaling theory, we conduct a PRISMA-guided systematic review of 127 peer-reviewed studies from Scopus and Web of Science to examine how machine learning (ML), deep learning (DL), Natural Language Processing (NLP), and Explainable AI (XAI) are transforming ESG rating analysis. We develop a tripartite framework classifying studies by the functional role of the ESG score: predicted (n = 29), used (n = 57), or assessed (n = 41). Our central contribution is the first synthesis of the methodological-assessment stream, organized into four clusters: XAI reverse-engineering of proprietary scoring functions, divergence reconciliation, greenwashing detection, and unsupervised industry-materiality clustering. The evidence assembled in this stream indicates that ESG ratings weight low-cost aspirational disclosure heavily relative to costly performance evidence, suggesting that greater reliance on aspirational disclosure relative to performance evidence may increase greenwashing risk, consistent with signaling-theory concerns. A study-level validation appraisal further shows that the most extreme fit statistics often arise in target-proximal reconstruction or non-temporal validation settings, cautioning against interpreting high R2 as evidence of transferable out-of-time forecasting.