Predicting ESG adoption among Malaysian SMEs using explainable machine learning

Abstract Environmental, social, and governance (ESG) considerations are increasingly embedded in regulatory frameworks, financial markets, and corporate strategy worldwide. While large listed firms have been the primary focus of ESG reporting requirements, small and medium enterprises (SMEs) play a central role in employment, supply chains, and regional development, particularly in emerging economies. In Malaysia, SMEs account for the vast majority of business establishments, yet ESG adoption remains uneven across firms. This study examines which non-practice firm-level characteristics are associated with ESG adoption classification among formal-sector Malaysian SMEs using explainable machine learning. The sample of 457 firms is drawn from a bank-linked stakeholder network, so it over-represents formal-sector, bank-connected SMEs, and the outcome is self-declared. The survey records ESG adoption on a four-level scale (fully, partially, planning to, or not implementing), which is recoded into a binary adopter/non-adopter target for the main analysis; it reflects reported adoption status rather than a verified measure of ESG performance. Logistic Regression, Random Forest, XGBoost, and Explainable Boosting Machine (EBM) are compared using nested cross-validation with all preprocessing fitted inside the cross-validation pipeline. On the held-out test set of 92 observations, Random Forest recorded the highest discrimination point estimates (ROC-AUC = 0.853; adopter-positive PR-AUC = 0.904). Paired bootstrap comparisons showed higher Random Forest discrimination than XGBoost and EBM, but did not distinguish Random Forest from Logistic Regression on ROC-AUC. No model was uniformly superior across all reported metrics, and the estimates remain imprecise because of the small held-out sample. On a common measure of variable-level importance computed for all four models, several predictors recur near the top, most consistently the use of financial consultants as an ESG advisory source, ownership type, and company type or market orientation, followed by programme-awareness and funding-intent variables. When the models are refitted across cross-validation folds, the rankings shift and few variables hold a fixed position, so these are reported as recurring but unstable predictive signals rather than a stable set of determinants. The results are predictive and non-causal: the model identifies firm characteristics associated with higher predicted adoption probability, not causal mechanisms. Performance is evaluated internally, with no external validation, so the model is exploratory and not ready for operational allocation or policy deployment. Within that scope, the findings offer an interpretable basis for exploratory segmentation of formal-sector, bank-connected SMEs and may help generate hypotheses for more targeted sustainability support.