From hundreds to dozens: stochastic hill climbing for parsimonious environmental, social, and governance scoring

Environmental, social, and governance (ESG) metrics are widely adopted tools for assessing corporate sustainability. Despite their importance for investors, regulators, and other stakeholders, ESG scores suffer from major drawbacks, including score divergence across providers, opaque and biased methodologies, and data quality issues. This paper focuses on Eikon Refinitiv, a prominent ESG data provider, to explore two interrelated challenges: the distortion of association measures due to missing data imputation and the potential redundancy among ESG key performance indicators (KPIs). First, we show that Refinitiv’s percentile ranking scheme preserves rank correlations (Kendall’s tau, Spearman’s rho) but not Pearson’s correlation when data are complete, but Refinitiv’s imputation for missing values with zeros inflates all association measures. Second, we develop an optimization-based tool to identify informational redundancy among KPIs. We formulate it as a cardinality-constrained rank correlation maximization problem and solve it via stochastic hill climbing with random restarts. Across three sectors and 10 years, we find that only 20–25% of KPIs are needed to approximate pillar scores, suggesting that ESG scoring models can be substantially streamlined without sacrificing accuracy. Our theoretical and empirical findings have important implications for regulatory efforts to standardize ESG ratings and offer practical guidance for constructing interpretable, reliable, and efficient ESG scores.