Diagnostic and Prescriptive Insights through XGBoost, SHAP and Differential Evolution: Application to Last-Mile Delivery in Portugal

The increasing complexity of last-mile delivery (LMD) operations, intensified by rising e-commerce demands and environmental variables, necessitates advanced decision-support systems that go beyond predictive analytics. This study introduces a comprehensive diagnostic and prescriptive framework for LMD performance enhancement, integrating XGBoost, SHAP, and Differential Evolution (DE) optimization. Using real-world data from a Portuguese logistics provider, the framework identifies root causes of delivery inefficiency and suggests actionable improvements through both multi-feature and single-feature perturbation techniques. The approach enables the classification of delivery services into performance tiers, explains the key factors behind inefficiencies, and recommends feasible interventions to transition services from suboptimal to star-performing. In addition, the framework integrates an LLM to provide a human-readable report based on the previous results. This research contributes to the limited yet growing field of prescriptive analytics in LMD, with potential implications for increased operational efficiency and revenue gains.