Denoising ESG: quantifying data uncertainty from missing data with Machine Learning and prediction intervals

Environmental, Social, and Governance (ESG) datasets are frequently plagued by significant data gaps, leading to inconsistencies in ESG ratings due to varying imputation methods. This study addresses the missing data issues in ESG datasets using machine learning techniques, comparing K-Nearest Neighbors, Gradient Boosting, Multiple Imputation by Chained Equations (MICE) and Neural Networks. We focus on quantifying the risk induced by data anomalies and provide tools to assess the impacts of this risk on the variability of the scores. By introducing prediction uncertainty using methods such as Predictive Mean Matching and Local Residual Draw, in order to assign confidence measures to individual predictions, we provide a nuanced understanding of prediction uncertainty. Empirical analyses show that these methods improve imputation accuracy and quantify uncertainty, which is required for reliable ESG scoring in banking and finance.