Predicting Delivery Failures in Technical Courier Services: A Survival Learning Approach for Sustainable Last-Mile Logistics
Year: 2026
Type: preprint
Abstract: Purpose: This study develops a predictive monitoring framework based on statistical and machine learning survival analysis to anticipate delivery failures, operational disruptions, and associated fleet emissions in last-mile logistics. Because technical courier operations involve scheduled appointments, on-site technical tasks, strict SLAs, and right-censored delivery outcomes, time-to-event modeling provides a proactive alternative to reactive exception handling. Design/methodology/approach: Using a structured benchmark of 60,000 technical courier delivery orders engineered to capture operational, spatial, behavioral, and environmental dynamics, the study benchmarks semi-parametric Cox proportional hazards models and regularized extensions (Ridge, Lasso, Elastic Net, and OSCAR) against non-parametric survival ensembles, including Random Survival Forests (RSF), Bagging Survival Trees, and Survival Boosting. Model discrimination, calibration, and prediction error are evaluated out-of-sample using Harrell's concordance index (C-index), time-dependent AUC, and the Integrated Brier Score (IBS), complemented by post-hoc interpretability (variable importance and partial dependence) and an online operational simulation. Findings: Survival learning models—and in particular the Enhanced Random Survival Forest incorporating dynamic operational and environmental covariates (including workload progression, routing/thermal/carbon indicators, and customer/service attributes)—outperform baseline statistical specifications, achieving strong out-of-sample discrimination (C-index≈0.84, AUC=0.8450) and low prediction error (IBS=0.0203). Post-hoc and sensitivity analyses show that cumulative workload, carbon/routing friction, thermal extremes, and urban/regulatory constraints (such as LTZ access) shape failure risk non-linearly, while proactive mid-route risk monitoring reduces preventable disruptions and associated CO2 emissions. Practical implications: The framework supports dispatchers and Transportation Management Systems (TMS) in shifting from post-facto disruption management to proactive, time-aware delivery monitoring—enabling earlier intervention on high-risk orders, better resource and route coordination, fewer failed attempts, and CO2 tracking aligned with ISO 14083:2023. Originality/value: Rather than treating delivery outcomes as static binary classifications or relying solely on reactive reporting, this work demonstrates how survival learning models can monitor time-dependent failure risk in technical courier logistics even when full physical IoT sensorization is not yet available, integrating operational reliability with environmental sustainability Keywords: Technical last-mile logistics, Supply chain resilience, Survival analysis, Delivery failure prediction, Green logistics. 🎯 Objective: This study proposes a dataset and proactive predictive framework based on survival analysis to evaluate and mitigate delivery failure risks in last-mile technical courier services, addressing operational inefficiency, elevated costs, and redundant emissions. • RQ1. Which factors exert the greatest influence on the occurrence of failed deliveries in technical courier services? • RQ2. Are advanced machine learning models capable of accurately predicting the time-dependent risk of delivery failure prior to service execution? • RQ3. What extent can predictive analytics enhance the sustainability of logistics operations by reducing failed deliveries and minimizing unnecessary vehicle movements?
Author: Garofalo, Nunziato
Language: English
Subfield: Transportation
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7236728762
