Entity Overlap and Generalisation in Last-Mile Route-Deviation Prediction: Evidence from Driver Holdout, Address Suppression, and Cross-Country Transfer

Planned last-mile routes often change during execution, making pre-dispatch identification of likely deviations useful. Performance measured when test routes involve drivers and addresses represented in training may not extend to unseen entities. This study audits training–test entity overlap in Norwegian–Danish route data, using ridge-regularised linear models as instruments and gradient-boosted trees as a nonlinear check. Rolling-origin, leave-driver-out, paired known-versus-unseen driver, graph-blocked leave-community-out, and bidirectional cross-country evaluations share a strict cohort of 18,723 routes. In the paired design, identical test routes were scored under matched training size, deviation prevalence, and country mix. The driver-identity gain in the area under the receiver operating characteristic curve (AUC) was 0.0837 for drivers observed in training and −0.0007 for unseen drivers, as indicator coding implies, and a within-dataset experience descriptor added little. Address identity added 0.0426 for unseen drivers with the linear baseline; this increment was not reproduced in the tested gradient-boosting specification with target encoding or under graph-blocked evaluation with reduced address overlap. Under cross-country transfer, AUC exceeded chance modestly in both directions, every unadjusted Brier skill score was negative, and identity-fitted specifications ranked routes better than structural features alone only into Norway. Findings are limited to one operator and two countries.