A Comparative Empirical Evaluation of Deep Recurrent Neural Networks versus Ensemble Learning Models for Dynamic Scope 3 Emission Risk Forecasting in Intermodal U.S. Freight Logistics

• Billy Elly

  Lautech

  Author

The imperative to decarbonize supply chains has intensified global focus on Scope 3 emissions, which constitute the majority of logistics-related greenhouse gas output yet remain notoriously difficult to predict due to their indirect nature and dependency on dynamic multi-modal operations. Existing forecasting approaches predominantly rely on retrospective emission factor calculations and static regression models, limiting proactive risk management capabilities. This study presents a comparative empirical evaluation of deep recurrent neural networks (specifically Long Short-Term Memory networks) against ensemble learning models (Random Forest and XGBoost) for dynamic Scope 3 emission risk forecasting in U.S. intermodal freight logistics. An enriched dataset integrating shipment records, multi-modal transport parameters, and energy intensity metrics was constructed and used to train models for classification of shipment-level emission risk. Empirical results demonstrate that ensemble learning models, particularly XGBoost, achieved superior predictive performance with an accuracy of 89.4% (AUC = 0.91), outperforming the LSTM-based deep learning approach which achieved 84.7% accuracy. Feature importance analysis identified shipment distance, transport mode mix, and cargo weight as the most influential predictors. These findings challenge the prevailing assumption that deep learning architectures inherently outperform traditional machine learning for structured logistics data and provide a practical, interpretable framework for proactive sustainability planning in freight operations. The study contributes to sustainable supply chain management literature by establishing benchmark performance metrics for predictive emission risk modeling and offering actionable insights for logistics practitioners.

• PDF

Copyright (c) 2026 Billy Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.