Which Environment, Social and Governance Indicators Influence the Financial Distress Prediction? Evidence From Explainable Neural Networks

Abstract

This study identifies which disaggregated environmental, social and governance (ESG) indicators influence financial distress prediction, combining an explainable artificial neural network, multinomial logistic regression and random forest on 9715 UK private firms (2020–2023). Relative to a benchmark using financial predictors alone, adding disaggregated ESG indicators raises holdout accuracy from 66.5% to 74.4% and the macro F1‐score from 0.657 to 0.748. Material information is concentrated in specific indicators rather than in the aggregate ESG score: executive remuneration and workforce‐ and stakeholder‐related practices provide the most convergent evidence, corporate governance receives moderate support and environmental indicators are heterogeneous. The aggregate score is consistently signed across methods but in the wrong direction, associated with a higher probability of distress, showing how composite ratings can invert the signals of their components. Materiality is stage‐ordered: long‐horizon practices are most informative among sound firms, governance controls among distressed firms and incremental value peaks in the grey zone.