Authors - Jose Alejandro Ascencio-Laguna, Armida Gonzalez-Lorence, Ana Lilia Mondragon-Solis, Victor Alberto Gomez-Perez Abstract - Machine Learning (ML) and geospatial clustering have traditionally been applied as independent approaches to urban freight transportation chal lenges, particularly arrival time prediction under "just-in-time" constraints. De spite their complementary nature, their integration remains underexplored, while distance-based methods relying on Euclidean metrics yield error margins of 18 35 minutes, insufficient for operational logistics. This study proposes a hierarchical framework combining geographic k-means clustering (k=14) as a spatial segmentation layer with an enhanced Random For est regressor incorporating temporal feature engineering. The architecture is com putationally efficient and robust to real-world uncertainty after training. The framework was validated across three metropolitan areas in Mexico using 306,847 records from June 2024, benchmarked against five algorithms through stratified temporal validation and Wilcoxon tests with Bonferroni correction. The proposed model achieved a Mean Absolute Error of 347.2 seconds (5.79 min), representing a 68.1% reduction relative to historical baselines (MAE: 1,089 s) and a 19.9% improvement over standalone Random Forest (MAE: 433 s). Eu clidean distance was the dominant predictor (43.7%), followed by geographic coordinates (32.8%). All improvements were statistically significant (p
