Integrating Lean Logistics and Machine Learning to Improve Last-Mile Delivery Performance: Evidence from a Logistics Company in Peru

Abstract

Last-mile logistics is a critical component of e-commerce distribution, particularly in urban environments where delivery failures, address inconsistencies, traffic variability, and operational inefficiencies affect service performance. This study evaluated a sequential Lean Logistics–Machine Learning (Lean–ML) framework in a last-mile logistics company in Huancayo, Peru. An applied quantitative, single-group quasi-experimental design was used, comprising baseline assessment (O1), implementation of 5S-based Lean Logistics (O2), and subsequent integration of a Random Forest (RF) predictive decision-support model (O3). The analytical dataset comprised 2900 historical delivery records, divided into 2320 training and 580 test observations. Across the sequential intervention, average delivery time decreased from 34.3 to 24.5 min (−28.5%), failed deliveries from 2.32% to 1.05% (−54.7%), operating cost from S/3.28/km to S/2.65/km (−19.2%), and address-registration errors from 27% to 2% (−92.6%), while on-time deliveries increased from 85% to 95% (+10 percentage points). The original model-development records reported 91.0% accuracy, 89.5% precision, 93.2% recall, and a 91.3% F1-score on the 580-record test set. A subsequent 30-day prospective field validation involving 3000 new orders recorded 67 incidents, of which 61 were correctly anticipated (91.0% incident detection). The results provide context-specific evidence that sequential process standardization and predictive decision support can support last-mile operational performance. However, because the study involved one company, no parallel control group, and a 30-day field-validation period, the findings should not be interpreted as isolated causal effects or as evidence of long-term model stability or broad generalizability.