ETDP: a high-performance transportation delay prediction framework with explanations for supply chains

Purpose This work aims to develop a transportation delay prediction framework to improve prediction performance while providing explanations for its predictions. Design/methodology/approach An ETDP (high-performance Transportation Delay Prediction Framework with Explanations) for supply chains is proposed, in which we design an LGFT (Local-Global Fusion Transformer) model for high-performance prediction, and integrate LGFT with SHAP to explain its predictions. LGFT captures local feature interaction patterns via 1D CNN and global nonlinear relationships via an MLP, which are then fused through a Transformer Encoder to model long-range contextual dependencies. Integrated with SHAP, ETDP formalizes a three-level explanation mechanism that can analyze global feature importance, key feature dependency, and instance-level attribution. Findings The ETDP framework is evaluated on three public supply chain datasets covering binary delay prediction, last-mile logistics, and five-class supply chain risk classification. On DataCo, ETDP achieves the highest precision and remains competitive across the other evaluation indicators. On India, it obtains the best recall, F1-score, and MCC in the held-out test evaluation. On Mendeley, it reaches the best accuracy, recall, MCC, and lowest FNR, while remaining descriptively close to XGBoost in F1-score, though below it in precision and probability-ranking metrics. Repeated-run analyses show statistically supported improvements over several representative baselines, although ETDP does not uniformly dominate every model and metric. Originality/value ETDP is a useful combination of predictive modeling and post hoc explainability for transportation delay prediction. Fusing local and global features, it contributes a high-performance prediction model. Integrating with SHAP, it provides a supply-chain-oriented three-level explanation workflow. It can support both early delay-risk sensing and more targeted intervention prioritization, such as strategic optimization, context-specific risk mitigation and specific order diagnostics. Source code is available at: https://github.com/Tang12333/Transportation-Delay-Prediction-Framework-ETDP