A Machine Learning Study of Country-Level ESG Modeling Using Open Indicators

Environmental, Social, and Governance (ESG) indicators are increasingly used to assess sustainability and institutional quality at the country level. Despite their growing importance, ESG data pose significant challenges for machine learning due to heterogeneity, high dimensionality, temporal dependence, and limited transparency in many existing scoring frameworks. This paper presents a systematic machine learning study of country-level ESG modeling using publicly available World Bank indicators. Rather than proposing a new ESG score, we formulate ESG analysis as four computational tasks: representation learning, level forecasting, directional change classification, and explainability. Using a time-aware evaluation protocol, we compare linear and nonlinear models across these tasks. Our results show that ESG indicators form a largely linear, low-rank space and that simple linear models trained on raw indicators outperform more complex approaches for forecasting ESG levels. While ESG levels are highly predictable, directional changes are substantially harder to anticipate. Explainability analysis further reveals that governance-related indicators dominate predictive performance across tasks. Together, these findings highlight both the opportunities and inherent limits of machine learning for ESG modeling and provide a transparent, reproducible benchmark using open data.