PREDICTING LATE DELIVERY RISK IN GLOBAL SUPPLY CHAIN OPERATIONS USING MACHINE LEARNING

Abstract

This study develops a supervised machine-learning system for predicting late-delivery risk in a global supplychain and deploys it as an interactive analytics dashboard. Using the DataCo Global logistics dataset – 180,519 daily order records spanning 2015–2019 with 53 operational attributes – a gradient-boosted decision-tree classifier (XGBoost) is trained to predict the binary target Late delivery risk (0 = On-Time, 1 = Late). Categorical fields are label-encoded and numerical fields scaled prior to an 80/20 train-test split. On the held-out test set (n = 36,104) the model achieves Accuracy = 68.78%, Precision = 80.48%, Recall = 56.86%, F1 Score = 66.64%, and ROC-AUC = 0.7253, comfortably outperforming a majority-class baseline (54.83% accuracy). The trained model, together with its encoder and scaler artefacts, is served through a five-page Streamlit dashboard providing KPI reporting, dataset exploration, visual analytics, model-performance interpretation, and prediction support. The contribution of this paper is a compact, interpretable, and fully deployed late- delivery risk pipeline that links model outputs to concrete operational actions in shipping-mode selection, regional riskmonitoring, and proactive customer communication.