Data-Driven Prediction of Processing Times and Delays Across Supply Chain Operational Stages Using Machine Learning
Abstract
This study proposes an ensemble learning framework to predict personnel data entry durations across six operational stages using transaction records collected from an operational supply chain information system. To accurately represent operational performance, processing times were calculated in working minutes by excluding nonworking periods, including nights and weekends. The dataset comprised 110,695 unique shipment records collected between 2022 and 2026 and underwent stage-specific data quality assessment and preprocessing. A median-based baseline predictor and three ensemble learning algorithms, HistGradientBoosting, XGBoost, and LightGBM, were evaluated for regression tasks. All models were trained on a log-transformed target (log1p), and predictions were back-transformed to working minutes using the inverse transformation (expm1). To prevent data leakage and better reflect real-world deployment, a year-based temporal split was adopted, with 2026 records reserved for testing. Model performance was evaluated using mean absolute error, root mean square error, coefficient of determination, accuracy, precision, recall, and F1-score. A single regression model was trained per stage, and binary labels (normal/slow) were derived by thresholding the predicted durations at a stage-specific cut-off optimized on the validation set. The results showed that predictive performance varied substantially across operational stages, with LightGBM achieving the best overall performance for the Loading Arrival stage, while lower performance in the Transport Entry and unloading stages suggested the influence of external operational factors not represented in the available data. For the Loading Arrival stage, LightGBM achieved an MAE of 285.989 working minutes and an R² of 0.371 on the test set. Overall, the findings demonstrate that ensemble learning provides an effective and computationally efficient approach for predicting personnel data entry durations and supports real-time decision-making, personnel performance monitoring, and process optimization in data-driven supply chain operations.