ETA Prediction in Last-Mile Logistics Under Domain Shift: From Zero-Shot Failure to Few-Shot Recovery

Estimated time of arrival (ETA) models for last-mile delivery are commonly evaluated within the same data source, leaving their reliability under cross-dataset distribution shift insufficiently characterized. This study evaluates internal generalization, independent external transfer, limited target-domain supervision, support-set sensitivity, and uncertainty calibration using five LaDe-D cities as the source domain and an independent planned-versus-actual route dataset from two countries as the target domain. CatBoost, LightGBM, XGBoost, and an uncertainty-aware neural reference model (DG-UCETA) were evaluated under a harmonized 21-feature Common-Core representation. External zero-shot MAE reached 150.76–185.47 min, despite substantially lower internal errors. A predefined 1% target-support condition reduced MAE to 57.60–64.87 min, although five independent support-set selections revealed substantial instability in this ultra-low-data regime; performance became markedly more stable at 5–10% support. A matched target-only control further showed that source pretraining improved MAE in only 4 of 16 model-support combinations, indicating that much of the observed recovery resulted from exposure to labeled target-domain data rather than from source initialization itself. Grouped feature ablation also localized the Common-Extended negative-transfer effect primarily to speed-derived variables, with speed-history features increasing external MAE by 28.02 min relative to Common-Core. Conformal recalibration improved coverage but did not correct systematic point-prediction bias. These findings indicate that representative target-domain supervision, feature compatibility, and adaptation stability are more critical to reliable cross-domain ETA deployment than architectural complexity alone.