GreenGuard-AU: A Parsimonious RegTech Framework for Greenwashing Detection in Data-Scarce Regulatory Environments (International Journal of Computer Applications)

Summary

• GreenGuard-AU evaluates 633 company-year observations from the Australian securities market (2021 to 2024) across 38 engineered features.p.1
• The sample comprises 130 positive greenwashing labels, consisting of 19 confirmed ASIC enforcement actions and 111 cases identified through ESG rating divergence.p.1
• The proposed soft-voting tree ensemble achieves an AUC-ROC of 80.35% (95% bootstrap CI: 75.73% to 84.70%, p < 0.01).p.1
• A benchmark deep transformer architecture reaches only 63.7% AUC under identical data-scarce conditions.p.1
• The model operates with 71.54% ± 10.39% recall and 51.04% ± 9.71% precision across 5-fold cross-validation.p.4
• Greenwashing risk prevalence is highest in Financial Services at 30.8%, followed by Energy and Utilities at 20.6%.p.6

Key insights

1. Feature-engineered tree ensembles outperform complex deep neural architectures in tabular regulatory environments with limited positive labels.p.1
2. The classification pipeline relies on a weighted soft-voting ensemble of Gradient Boosting (weight 0.50), XGBoost (weight 0.35), and Random Forest (weight 0.15).p.3
3. High model complexity creates an Over-architecture Penalty, causing a multimodal transformer to reach only 63.7% AUC compared to 80.35% for GreenGuard-AU.p.4
4. The single strongest predictor of greenwashing risk is commitment type change, achieving a mean absolute SHAP value of 0.604.p.6
5. Sector-weighted linguistic signals consistently rank higher in feature importance than their unweighted counterparts by incorporating domain context.p.6
6. Removing standard financial performance indicators has negligible impact on classification performance, shifting AUC from 80.35% to 80.62%.p.7
7. The screening framework trains in under five seconds on commodity hardware, making monthly screening of listed securities computationally practical.p.7