Incremental Predictive Value of Spatial Graph Structure in Hourly Istanbul Traffic Forecasting

A Confirmatory Multi-Season Study

Abstract

I had long assumed that when forecasting traffic in a highly connected and complex city such as Istanbul, providing a model with the physical road network and the actual relationships between streets should offer a substantial predictive advantage. Yet one question kept returning to me: if a model already observes the previous 24 hours of traffic, the time of day, and the day of the week, might it already be learning much of this spatial information indirectly through temporal patterns? To avoid the temptation of unconsciously modifying the analysis or selecting a more favorable result, I constrained the study before examining the confirmatory outcomes. The main research question, confirmatory months, model conditions, random seeds, statistical procedures, and primary evaluation metric were specified in advance, so that I could not later change the data or models simply to produce a more attractive result. Using hourly traffic data from Istanbul Metropolitan Municipality (İBB) for May and November 2024, I studied 64 traffic locations per analyzable month and compared a Dynamic Graph Transformer (DGT) supplied with physical road-network structure against a temporal MLP that used no spatial adjacency information. Both models used the preceding 24 hours as input. At the +1 hour horizon, the graph model achieved a mean MAE of 3.6804 km/h compared with 3.6986 km/h for the temporal MLP. This corresponds to a paired difference of -0.0181 km/h, or an estimated improvement of approximately 0.49%. The difference was not statistically reliable: the preregistered primary test yielded p = 0.5000, and the 95% hierarchical bootstrap confidence interval [-0.1760, +0.1120] km/h crossed zero. A secondary +1 hour versus +6 hour contrast produced a raw p value of 0.0137, but the Holm-adjusted p value was 0.0684 and therefore did not meet the registered significance criterion. The conclusion is not that road graphs are useless for traffic forecasting. Rather, under this specific experimental design, these two confirmatory months, and the tested architecture, adding explicit physical road structure did not provide a reliable predictive advantage over a strong temporal model. One possible explanation is that temporal regularities or adaptive relationships learned by the graph model already capture part of the information represented by the physical network. The result illustrates how an engineering idea can appear compelling and intuitively correct while real data show that its incremental contribution is smaller than expected.