The modelling pipeline was built and evaluated using TimeSeriesSplit cross-validation, a technique that respects the chronological order of traffic data, combined with autoregressive traffic-volume lags, meaning past traffic volumes feed forward as predictors of future volumes. Within the full-feature specification, Random Forest showed the most favourable validation profile among the candidate algorithms and was retained for detailed out-of-time (OOT) analysis, in which the model is assessed on data from a period separate from its training window. To examine the fitted associations between predictors and modelled traffic volume, the researchers applied SHapley Additive exPlanations (SHAP) and partial-dependence analysis, tools that attribute predictions to individual predictors and trace the average fitted relationship between a single predictor and the output, allowing the model’s use of its inputs to be inspected rather than simply trusting accuracy figures.

One of the study’s more sobering findings concerns weather. A separate ablation analysis, in which groups of predictors are removed to test their contribution, showed that concurrent meteorological variables provided only marginal incremental OOT improvement beyond what temporal variables and historical traffic lags already captured. For this dataset, the memory of past traffic and the rhythm of the day and week did most of the predictive work, a result that matters for any planner hoping to gain decision-relevant information from weather inputs in traffic-volume models.

The authors are explicit about the limits of what these predictions represent. Traffic-volume predictions do not directly measure route travel time, delivery delay, or site-production disruption. Moreover, the meteorological variables used in the full-feature analysis are concurrent observations in the archived dataset, meaning the OOT analysis represents retrospective traffic-volume modelling using temporal, lagged-traffic, and concurrent meteorological information — information that would not necessarily be available in advance in a real forecasting setting.

That gap is precisely what the proposed decision-translation framework is designed to address. Instead of claiming that the models are ready for operational use, the framework identifies the additional information and calibration needed before traffic-volume signals could inform construction-logistics decisions. The study notes that its analysis has not been evaluated against delivery windows, just-in-time disruption, cost, or emissions, and that such evaluation would require forecast-available weather information, route-specific travel-time data, delivery schedules, and site-operation data.

The methodological approach combines time-series cross-validation, out-of-time testing, ablation, and interpretable machine learning in a single workflow, offering a template for how traffic modelling research could be connected to practical decision-making without overstating what the models can deliver. The interpretability layer is central to this: by exposing which predictors drive modelled traffic volume and how, SHAP and partial-dependence analysis make the model’s fitted associations available for scrutiny before any downstream use is contemplated.

The research remains a proposed framework at this stage, and the authors caution that the bridge from predicted traffic volume to logistics outcomes has not yet been built or tested. Future work, as framed in the study, would need to pair the traffic models with the operational data that defines a real delivery decision, closing the loop between what a network model predicts and what a construction logistics manager needs to know. Until that evaluation is carried out with forecast-available weather, route-specific travel times, delivery schedules, and site-operation data, the framework stands as a structured account of what stands between hourly traffic-volume predictions and their use in construction delivery planning.

Subject of Research: Interpretable time-series machine learning for hourly traffic-volume modelling in construction logistics decision support

Article Title: Interpretable time-series traffic-volume modelling for construction logistics management: a proposed decision-translation framework

Article References: Thi Nguyen, T.-H., Nguyen, T. Q., & Bui, B.-N. G. (2026). Interpretable time-series traffic-volume modelling for construction logistics management: a proposed decision-translation framework. Frontiers in Built Environment, 12, Article 1954525. https://doi.org/10.3389/fbuil.2026.1954525

Image Credits: AI Generated

DOI: 10.3389/fbuil.2026.1954525

Keywords: construction logistics, traffic-volume modelling, time-series machine learning, Random Forest, SHAP, partial dependence, out-of-time validation, TimeSeriesSplit, ablation analysis, traffic-weather data, decision-translation framework, interpretable machine learning