Invisible Footprints, Visible Insights: Machine Learning Reveals Scope 3 Emissions

Nian-Zu YehSzuyung Wang^*

• National Taiwan University, Taipei City, Taiwan

Scope 3 greenhouse gas emissions are critical to firms' carbon footprints yet are often difficult to quantify due to limited direct data, motivating predictive modeling approaches. Methods: We developed and compared four machine learning algorithms (K-nearest neighbors, random forest, AdaBoost, and XGBoost) to estimate corporate Scope 3 emissions using readily available financial and sustainability performance data. We leverage 10,449 listed firm-level data from 2014 to 2023, covering major industries such as semiconductor, steel, textile, and building materials, evaluating performance of each model by a held-out test set with metrics including R², mean absolute percentage error (MAPE), and root mean squared logarithmic error (RMSLE). Results: XGBoost achieved the highest accuracy (R² = 0.85, MAPE = 15%, RMSLE = 0.20), outperforming random forest (R² = 0.80, MAPE = 20%) and AdaBoost (R² = 0.78), while K-NN had the lowest accuracy (R² = 0.60). The results demonstrate that ensemble tree-based models substantially improve Scope 3 emission prediction accuracy over simpler models. Discussion: Notably, random forest's interpretable feature importance provided insight into key emission drivers with only a slight accuracy trade-off, highlighting the balance between predictive accuracy and model interpretability.