The Limits of Emission-Based Learning in 3PL Operations: Evidence from Medical and Pharmaceutical Last-Mile Deliveries

Medical and pharmaceutical last-mile deliveries are simultaneously expected to be fast, reliable and temperature-safe for patients and to become measurably greener, yet these objectives often pull transport operations in opposite directions. Third-party logistics (3PL) providers are therefore increasingly required not only to report transport CO₂ emissions, but also to learn from them; however, it remains unclear whether the routine operational data they collect are sufficiently informative to enable such emission-based learning in this regulated and service-critical setting. This study examines the predictive limits of machine learning models in estimating CO₂ emissions in medical and pharmaceutical last-mile deliveries performed by a 3PL operator. Using operational data from six customers, we compare global and customer-specific models for the following two dependent variables: total CO₂ emissions per transport operation and CO₂ emissions per pallet. Linear and non-linear models, including linear regression, ElasticNet, Random Forest, HistGradientBoosting and XGBoost, are evaluated using chronological train-test splitting and cross-validation. The results show that global models fail to outperform a naïve benchmark, with negative R² values for both emission measures. Customer-level models reveal substantial heterogeneity as follows: for selected customers, especially those with more regular operational patterns, moderate predictive performance is achieved, while for others, emissions remain largely unpredictable using the available variables. The findings suggest that routine shipment-level data are insufficient for robust emission prediction in 3PL last-mile operations. Emission-based learning requires richer contextual, vehicle, route, traffic and telematics data, as well as customer-sensitive modelling approaches. The study contributes by identifying the data and modelling limits of sustainability intelligence in medical and pharmaceutical last-mile logistics.