Explainable machine learning predicts self-reported ESG adoption among
Small and medium-sized businesses make up most businesses in Malaysia and provide a large share of the country's jobs. However, they do not all follow environmental, social and governance (ESG) practices to the same extent. A research study used survey answers from 457 formal businesses connected to banks to examine which company characteristics were linked to reported ESG adoption. The study used predictive models, so it could identify patterns but could not prove that any characteristic caused a company to adopt ESG practices.
The researchers shared the survey through a network of people and organisations connected to banks. It included businesses in manufacturing, trading, services, agriculture, construction and other industries. Business owners and decision-makers answered the questions. Of the respondents, 285, or 62.36%, said they were the only person making decisions for their company. The survey was available in English, Bahasa Melayu and Mandarin. It collected information about company characteristics, target markets, ESG awareness, knowledge of support programmes, sources of advice, business activities, barriers and reported practices.
The researchers measured adoption by asking, “Is your company currently practising any ESG-related initiatives?” Respondents chose one of four answers: fully implementing, partially implementing, planning to implement or not implementing. For the main analysis, companies that were fully or partly implementing ESG initiatives counted as adopters. Companies that were planning to start or had not started counted as non-adopters. This measure reflected what companies said about their own activities. It did not confirm their actual ESG performance, the quality of their efforts or their sustainability results.
To reduce the risk of a model making predictions using information that was already part of the outcome, the researchers left out all questions about specific environmental, social and governance practices from the main model. Instead, they used 35 other predictors, including company details, awareness of support programmes, exposure to markets, sources of advice and plans for funding or stock exchange listing. After data preparation and one-hot encoding, these predictors became 112 model columns. The researchers handled missing data and applied all transformations that depended on the data distribution separately within each training fold.
The researchers compared four machine learning models: Logistic Regression, Random Forest, XGBoost and an Explainable Boosting Machine. They divided the data into training and test sets, using 80% for training and 20% for testing while keeping the proportions of the outcome groups similar. They also used five outer cross-validation folds and four inner folds to tune the models' settings. The training set contained 365 companies, while the separate test set contained 92. The researchers measured accuracy, F1 scores for each class, ROC-AUC, PR-AUC for both possible positive classes and Brier score. They used 1,000 stratified bootstrap samples to estimate how uncertain the test results were.
Random Forest had the highest estimated scores for distinguishing between adopters and non-adopters in the test set. Its ROC-AUC was 0.853, and its PR-AUC when adopters were treated as the positive class was 0.904. In paired bootstrap comparisons, its ROC-AUC was 0.052 higher than XGBoost's, with a 95% confidence interval from 0.012 to 0.093. It was 0.036 higher than the Explainable Boosting Machine's, with a confidence interval from 0.004 to 0.078. Random Forest also performed better than these two models on both PR-AUC measures. However, the researchers did not find a clear difference between Random Forest and Logistic Regression in ROC-AUC. No model performed best on every measure reported.
When the researchers ranked the importance of groups of predictors by measuring how much model performance changed when information was shuffled, six variables appeared in the top ten for all four models in the ranking based on the single held-out test set. These were ownership type, use of financial consultants as a source of ESG information, awareness of Bank Negara Malaysia's Low Carbon Transition Facility, plans to list on a stock exchange, awareness of sustainable finance instruments and plans to seek outside funding. Ownership type and use of financial consultants ranked first and second in every model. Using financial consultants was negatively associated with predicted ESG adoption. However, because the survey captured information at one point in time, the researchers could not tell whether consultants affected adoption or whether companies that were less prepared internally were more likely to seek their advice.
The rankings changed when the researchers retrained the models using different cross-validation folds. Use of financial consultants appeared in the top ten in 19 out of 20 model-and-fold rankings. Ownership type appeared in 18, as did company type or market orientation. By comparison, awareness of the Low Carbon Transition Facility appeared in seven rankings, while plans to list on a stock exchange appeared in three. Variables related to awareness of different support programmes may also share or shift their importance depending on the model and the method used to explain its predictions. These findings point to recurring but unstable patterns that help predict adoption, rather than a fixed set of factors that determine it.
Several limitations affect how the findings should be understood. Because the sample came from a bank-connected network, it likely included more formal businesses with established banking relationships and may not represent Malaysian SMEs as a whole. Since the study collected information at one point in time, it cannot show that one factor caused another. Biases in how people answered the survey, including a tendency to give socially acceptable answers, may also have affected the links between reported company characteristics and ESG adoption. Turning the four adoption stages into just two groups removed differences between companies that were at different stages of implementation. The small test set also led to wide uncertainty intervals and imprecise estimates of how well the predicted probabilities matched actual outcomes. The models were not tested on an independent external dataset, so they are not yet ready to guide real-world decisions about resource allocation or policy. Future studies would need to follow companies over time, include businesses beyond those connected to banks and check survey responses against objective evidence, such as emissions records and supply-chain audit results.
Subject of Research: Using explainable machine learning to classify self-reported ESG adoption among formal-sector Malaysian SMEs connected to banks
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