Explainable machine learning predicts self-reported ESG adoption among bank-connected Malaysian SMEs

Small and medium enterprises account for the vast majority of business establishments in Malaysia and a substantial share of national employment, but their environmental, social and governance adoption remains uneven. A research study using survey data from 457 formal-sector, bank-connected SMEs examined which firm characteristics were associated with reported ESG adoption. The analysis is predictive and does not establish that any characteristic causes adoption.

The survey was distributed through a bank-linked stakeholder network and covered firms in manufacturing, trading, services, agriculture, construction and other sectors. Owners and decision-makers completed the questionnaire, with 285 respondents, or 62.36%, identifying as the sole decision-maker. The survey was administered in English, Bahasa Melayu and Mandarin and recorded firm characteristics, market orientation, ESG awareness, knowledge of support programmes, advisory sources, operational measures, barriers and reported practices.

The outcome was based on the question “Is your company currently practising any ESG-related initiatives?” Respondents selected one of four levels: fully implementing, partially implementing, planning to implement or not implementing. For the main analysis, full or partial implementation was classified as adoption, while planning or no implementation was classified as non-adoption. The target represents self-declared adoption status, not verified ESG performance, quality or sustainability outcomes.

To reduce the risk of circular prediction, the main specification excluded all environmental, social and governance practice items that partly constitute the adoption outcome. It retained 35 non-practice predictors, including firm profile, programme awareness, market exposure, advisory sources and funding or listing orientation. These expanded to 112 model columns after preprocessing and one-hot encoding. Missing-value handling and all distribution-dependent transformations were fitted within each training fold.

The researchers compared Logistic Regression, Random Forest, XGBoost and an Explainable Boosting Machine using a stratified 80/20 train-test split, five outer cross-validation folds and four inner folds for hyperparameter tuning. The training set contained 365 firms and the held-out test set 92. The models were evaluated with accuracy, class-specific F1 scores, ROC-AUC, PR-AUC under both class polarities and Brier score. A stratified bootstrap with 1,000 resamples estimated uncertainty in the test metrics.

Random Forest had the highest held-out discrimination point estimates, with a ROC-AUC of 0.853 and an adopter-positive PR-AUC of 0.904. In paired bootstrap comparisons, its ROC-AUC exceeded XGBoost by 0.052, with a 95% confidence interval of 0.012 to 0.093, and the Explainable Boosting Machine by 0.036, with an interval of 0.004 to 0.078. Random Forest also separated from those two models on both PR-AUC measures. The comparison did not distinguish Random Forest from Logistic Regression on ROC-AUC, and no model was uniformly superior across all reported metrics.

Grouped permutation importance identified six variables in the top ten for all four models on the single held-out ranking: ownership type, use of financial consultants as an ESG information source, awareness of Bank Negara Malaysia’s Low Carbon Transition Facility, intent to list on a stock exchange, awareness of sustainable finance instruments and intent to seek external funding. Ownership type and consultant use held the first two positions in each model’s ranking. Use of financial consultants had a negative predictive association with adoption, but the cross-sectional survey could not determine whether consultants influence adoption or whether firms with lower internal readiness are more likely to seek them.

The rankings were not stable when the models were refitted across cross-validation folds. Financial-consultant use appeared in 19 of 20 model-fold top-ten rankings, ownership type in 18 and company type or market orientation in 18. By comparison, Low Carbon Transition Facility awareness appeared in seven and listing intent in three. Correlated programme-awareness variables can also share or redistribute importance depending on the model and explanation method. The results therefore support recurring but unstable predictive signals rather than a fixed set of determinants.

Several limitations bound the findings. The bank-network sample likely overrepresents formal SMEs with established banking relationships and is not nationally representative. The cross-sectional design cannot support causal inference, and common-method and social-desirability biases may affect associations between self-reported predictors and adoption. The binary outcome removes differences in implementation stage, while the small test set produced wide uncertainty intervals and imprecise calibration estimates. No external validation was performed, so the models are not ready for operational allocation or policy deployment. Future research would require longitudinal data, sampling beyond bank-connected firms and validation against objective records such as emissions data and supply-chain audit outcomes.

Subject of Research: Explainable machine learning classification of self-reported ESG adoption among formal-sector, bank-connected Malaysian SMEs

Article Title: Predicting ESG adoption among Malaysian SMEs using explainable machine learning

Article References: Ad-Din, S., & Seera, M. (2026). Predicting ESG adoption among Malaysian SMEs using explainable machine learning. Discover Artificial Intelligence, 6(1), Article 1446. https://doi.org/10.1007/s44163-026-02136-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02136-9

Keywords: ESG adoption, Malaysian SMEs, explainable machine learning, Random Forest, survey research, predictive modelling, sustainable finance, bank-connected firms, Predicting, adoption, Malaysian, SMEs