New AI Framework Audits ESG Data for Climate Risk Validation

In a new preprint on arXiv, researchers Karan Sehgal and Khawar Naveed Bhatti present a deterministic framework for auditable climate risk intelligence from fragmented ESG data. The system addresses the heterogeneity of Scope 1, 2, and 3 emissions reporting by combining single-source-of-truth orchestration with temporal drift analysis, SMOTE-based rare event optimization, and ensemble learning. It uses TreeSHAP for interpretability and governance inspection, enabling full provenance chains from flagged anomalies to escalation decisions. The framework is evaluated against statistical classifiers, anomaly detection methods, and temporal forecasting baselines using classification metrics (recall, F1, ROC AUC), calibration metrics (ECE, Brier score), and a governance audit trace completeness metric.

To support open reproducibility, the authors constructed and released a synthetic ESG validation benchmark calibrated to the GHG Protocol, PCAF, and ISSB standards. The methodology achieved strong results across stratified five-fold cross-validation with paired significance testing. By reframing ESG reporting as deterministic climate risk governance infrastructure, the framework promises to make corporate sustainability data more reliable, explainable, and auditable—critical for regulatory compliance (e.g., EU CSRD, SEC climate rules) and investor confidence. The paper (22 pages, 7 figures) is available on arXiv under a Machine Learning subject area.