Explainable hour-ahead traffic-volume forecasting with condition-stratified evaluation: an interstate I-94 case study

Introduction Hour-ahead traffic-volume forecasting can support traveler information, staffing, and advance corridor management. However, reliable evaluation requires attention to data quality, temporal variation, and forecasting uncertainty. This study evaluates an auditable, condition-stratified forecasting framework using traffic observations from westbound Interstate I-94 in Minnesota. Methods The Metro Interstate Traffic Volume dataset was audited for duplicate records, repeated timestamps, physically implausible rainfall observations, and missing hourly intervals. After preprocessing, 40,574 usable hourly timestamps yielded 28,855 valid gap-aware forecasting sequences with 24-hour input windows. The evaluated models included persistence, a 24-hour seasonal-naive baseline, LightGBM, XGBoost with and without weather variables, DLinear, GRU, a compact iTransformer-style model, and the Temporal-Mean Residual Network (TMRNet). Trainable models were evaluated using five prespecified random seeds. Performance was assessed through forecasting accuracy, paired statistical comparisons, direct forecasting horizons, rainfall-stratified errors, empirical prediction intervals, grouped permutation importance, TreeSHAP, and Integrated Gradients. Results XGBoost without weather achieved the lowest mean test MAE of 141.67 ± 0.31 vehicles/hour, followed by LightGBM (142.63 ± 0.41) and all-feature XGBoost (142.89 ± 0.56). TMRNet achieved a higher MAE of 220.34 ± 3.01 vehicles/hour. Removing weather variables from XGBoost produced an ensemble-mean MAE difference of −0.98 vehicles/hour relative to all-feature XGBoost (95% moving-block-bootstrap confidence interval: −2.21 to 0.12; Diebold–Mariano p = 0.106), indicating no statistically significant improvement. In a separate single-run sensitivity analysis, XGBoost achieved MAEs of 143.64, 182.65, and 197.44 vehicles/hour at the 1-, 3-, and 6-hour forecasting horizons, respectively. The test partition contained 5,371 zero-recorded, 353 light-rain, 45 moderate-rain, and two heavy-rain target hours. All 400 positive-rain test targets occurred within the final 909 valid test targets, making rainfall comparisons temporally confounded. At 90% nominal coverage, XGBoost achieved empirical coverage of 0.904 with a mean interval width of 661.02 vehicles/hour, compared with 0.903 and 945.69 vehicles/hour for TMRNet. Traffic history and temporal-cycle variables contributed more strongly to model predictions than direct rainfall inputs. Discussion The findings demonstrate the value of an auditable and condition-stratified approach to hourly traffic forecasting. Tree-based models generally outperformed the evaluated neural alternatives, while weather variables did not produce a statistically significant accuracy gain for XGBoost. However, the limited heavy-rain observations, temporal clustering of rainfall, and use of a single roadway restrict broader conclusions. The results do not establish heavy-rain robustness, suitability for sub-hourly traffic control, or generalization to other transportation networks.