A Data-Driven Framework for Predicting Truck Turnaround Times in Maritime Terminals: Flow-Aware Models

Truck Turnaround Time (TTT), defined as the total gate-to-gate time during a truck visit to a container terminal, is a critical performance indicator of landside operations. Although machine learning has been applied to TTT prediction, existing studies typically pool import and export movements into a single model, overlooking operational heterogeneity between flows. Moreover, little attention is paid to analyzing the importance of factors that enable TTT prediction. This study develops a flow-disaggregated predictive framework for a Chilean container terminal using 754,568 export and 1,056,351 import truck visits recorded between 2017 and 2023. Four tree-based ensembles and three neural network architectures are benchmarked under a chronological train/test split. Tree-based models tend to achieve marginally lower errors, though differences are small and not uniform across flows. A pronounced asymmetry emerges: import predictions are substantially more accurate (MAE = 6.79 min; WAPE = 34.36%) than export predictions (MAE = 32.49 min; WAPE = 52.53%), and this gap persists after normalizing for differences in mean TTT. TreeSHAP analysis identifies distinct predictive structures: export TTT is primarily associated with gate congestion and maritime service activity, while import TTT is more strongly associated with intra-terminal travel distance and crane operator experience. The higher Gini concentration and bidirectionality of dominant export predictors are consistent with unobserved drivers—such as the states of inspection queues (customs, sanitary, etc.) and the off-dock truck staging area—that limit predictive accuracy beyond process variability alone. In the integrated model, flow-identifying variables rank among the most influential features, providing empirical support for flow disaggregation. These findings indicate that flow-specific modeling improves both accuracy and interpretability in operationally heterogeneous terminal processes.