Comparative analysis of ensemble machine learning models for predicting material delivery delays in construction procurement and logistics

Material delivery delays represent a critical challenge in construction supply management, driving schedule overruns and cost escalation. Traditional delay-prediction approaches rely on subjective judgement and struggle to capture complex patterns in procurement and logistics processes. This paper develops an ensemble machine-learning (ML) model framework to predict delay occurrence and forecast duration using a real-world material delivery records dataset. The ensemble ML models were tuned using five-fold cross-validation repeated across five random seeds on the training set of a stratified 80:20 train-test split and evaluated on the completely unseen test set, on a dataset encompassing procurement and logistics operations to improve supply chain robustness. CatBoost was the best-performing classifier overall and handled the class imbalance most reliably, achieving the highest ROC-AUC (0.958), balanced accuracy (0.903), specificity (0.883) and MCC (0.673) on the test set. However, XGBoost achieved the best prediction accuracy for the delay duration with an R2 of 0.720, RMSE of 34.6 days and MAE of 17.5 days, indicating better agreement with actual delay durations compared with other existing ML models. In addition, SHAP analysis identifies manufacturer, material characteristics, project size and source country as the most influential predictors. Finally, the framework improves early visibility of risks and provides early warnings at the purchase-order preparation stage, supporting supplier evaluation, procurement planning and proactive decision-making in the construction supply management.