Hour-ahead traffic forecasts can shape everything from traveler information to staffing decisions and advance corridor management, but only if the numbers behind them can be trusted. A new study from researchers at American International University-Bangladesh, Rajshahi University of Engineering and Technology, and Multimedia University in Malaysia tackles that trust problem head-on, using westbound Interstate I-94 in Minnesota as a single-roadway case study. The team built an auditable, condition-stratified forecasting framework around the Metro Interstate Traffic Volume dataset, then subjected a lineup of machine learning models to a battery of accuracy checks, statistical comparisons, and explainability tools.

The work began with a data audit rather than a model build. The researchers screened the hourly records for duplicates, repeated timestamps, physically implausible rainfall observations, and missing hourly intervals. One implausible rainfall event was excluded, and gap-aware 24-hour windowing was applied so that no forecasting sequence silently bridged a data gap. The cleaned dataset retained 40,574 usable hourly timestamps, from which 28,855 valid forecasting sequences were derived, each built on a 24-hour input window.

The benchmark lineup was deliberately broad. It included a persistence baseline, a 24-hour seasonal-naive reference, two gradient-boosted tree models (LightGBM and XGBoost, the latter tested both with and without weather covariates), two neural approaches (DLinear and a gated recurrent unit, or GRU), a compact iTransformer-style model, and a purpose-built architecture called the Temporal-Mean Residual Network, or TMRNet. Every trainable model was run across five prespecified random seeds so that reported performance reflected variability in training rather than a single lucky run.

The results were clear on one point: tree-based models led the field. XGBoost without weather variables achieved the lowest mean test mean absolute error, 141.67 ± 0.31 vehicles per hour, followed by LightGBM at 142.63 ± 0.41 and all-feature XGBoost at 142.89 ± 0.56. TMRNet, the neural entry, trailed with a mean absolute error of 220.34 ± 3.01 vehicles per hour. In other words, the deep learning alternative not only lost on accuracy but did so with wider seed-to-seed variability.

The weather question produced a subtler answer. Removing weather variables from XGBoost produced an ensemble-mean MAE difference of −0.98 vehicles per hour relative to the all-feature version, with a 95% moving-block-bootstrap confidence interval spanning −2.21 to 0.12 and a Diebold–Mariano p-value of 0.106. That interval crosses zero, so the study concludes that adding weather inputs did not produce a statistically significant accuracy gain for XGBoost. The authors caution that this is a statement about this dataset and configuration, not a universal verdict on weather-aware traffic forecasting.

Horizon sensitivity was examined in a separate single-run analysis using seed 11. XGBoost posted mean absolute errors of 143.64, 182.65, and 197.44 vehicles per hour at the 1-, 3-, and 6-hour horizons respectively, a pattern consistent with the expected degradation of forecast skill as the target moves further from the observed history.

One of the study’s most careful contributions is its treatment of rain. The test partition contained 5,371 zero-recorded, 353 light-rain, 45 moderate-rain, and just two heavy-rain target hours. Worse, all 400 positive-rain test targets fell within the final 909 valid test targets, meaning the rainfall-stratified comparison is temporally confounded with the late-summer portion of the test period. The authors therefore label moderate-rain results as exploratory and state plainly that heavy-rain robustness is unsupported by their data.

Uncertainty was quantified rather than assumed. At 90% nominal coverage, XGBoost achieved an empirical prediction-interval coverage of 0.904 with a mean interval width of 661.02 vehicles per hour, while TMRNet reached 0.903 coverage with a substantially wider 945.69 vehicles per hour. The near-identical coverage with narrower intervals gave XGBoost a clear edge in interval sharpness, a practical advantage when intervals are used for operational planning.

Explainability rounded out the evaluation. TreeSHAP, Integrated Gradients, and grouped permutation importance all pointed in the same direction: recent traffic history and temporal-cycle variables, such as hour-of-day patterns, contributed more strongly to model predictions than direct rainfall inputs. This convergent evidence helps explain why dropping weather covariates barely moved the accuracy needle, since the models were leaning primarily on the traffic signal itself.

The authors are explicit about the limits. A single roadway, limited heavy-rain observations, and temporally clustered rainfall restrict broader conclusions, and the results do not establish suitability for sub-hourly traffic control or generalization to other transportation networks. What the study does deliver is a template: an auditable pipeline, seed-averaged reporting, paired statistical tests, condition-stratified error analysis, calibrated prediction intervals, and multiple explainability methods, all applied to a realistic hour-ahead forecasting task. For practitioners deciding whether weather data belongs in a traffic model, the message is that the answer depends on evidence gathered under exactly this kind of scrutiny.

Subject of Research: Hour-ahead traffic-volume forecasting on Interstate I-94 using explainable, condition-stratified machine learning evaluation

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

Article References: Ahmed, F., Ahmed, N., Apurbo, T. H., Noor Asif, M. H., & Islam, M. S. (2026). Explainable hour-ahead traffic-volume forecasting with condition-stratified evaluation: an interstate I-94 case study. Frontiers in Future Transportation, 7, Article 1942649. https://doi.org/10.3389/ffutr.2026.1942649

Image Credits: AI Generated

DOI: 10.3389/ffutr.2026.1942649

Keywords: traffic forecasting, Interstate I-94, XGBoost, LightGBM, TMRNet, prediction intervals, TreeSHAP, Integrated Gradients, rainfall stratification, Metro Interstate Traffic Volume dataset, machine learning, explainable AI