Optimized Import Delay Prediction with Gradient Boosting and PSO: An Interpretable Machine Learning Approach Using SHAP

Abstract

Delays in import operations pose significant chal- lenges to global supply chains, particularly for trade-dependent countries. This study proposes a machine learning-based ap- proach to predict the risk of import delays by using operational timestamp data derived from real-world logistics activities. The data set was obtained from a logistics company in Indonesia, known as PT. XYZ and comprises import transaction records from January to June 2025. Three classification models were evaluated, Decision Tree (DT), Random Forest (RF), and Gra- dient Boosting Classifier (GBC), with hyperparameter tuning performed using Particle Swarm Optimization (PSO). The ex- perimental results indicate that the GBC model achieved the highest prediction accuracy of 93.9% following optimization. To improve model interpretability, SHAP (SHapley Additive exPlanations) was employed to analyze feature contributions, revealing that variables such as RcdDate, Delivery and Paid Duty Tax had a significant impact on delay outcomes. This integrated framework not only provides accurate predictions, but also offers actionable insights to support data-driven decision-making in logistics operations.