Istanbul traffic forecasting with graph neural networks and Dynamic Graph Transformer

Preregistered open urban AI research · Istanbul

Istanbul Traffic Forecasting with GNN and ST-GNN

İstanbul GraphTraffic AI · Registered Confirmatory v2 completed

Graph Neural Networks ST-GNN Dynamic Graph Transformer OSF Preregistered

Published engrXiv Preprint GitHub Repository OSF Preregistration Registered Study Article Confirmatory Results Zenodo DOI

Preregistered graph-based Istanbul traffic forecasting

İstanbul GraphTraffic AI is a reproducible research project using public İBB hourly traffic data to study temporal forecasting, directed road graphs, adaptive connectivity and Dynamic Graph Transformers. A registered confirmatory Protocol v2 analysis has now been completed and permanently archived.

Registered confirmatory result

Primary H1 · +1h −0.0181 km/h

Mean paired daily MAE difference, DGT Directed Road + Adaptive minus Temporal MLP.

Primary statistical test p = 0.5000

One-sided paired Wilcoxon; registered α = 0.05. H1 was not statistically supported.

Relative MAE difference −0.490%

95% hierarchical bootstrap CI for the absolute paired difference: [−0.1760, +0.1120] km/h.

Transparent registered inference: the primary graph-model superiority hypothesis was not supported. H3 showed a raw short-vs-long-horizon signal (raw p = 0.0137), but it did not remain significant after the preregistered Holm correction (adjusted p = 0.0684).

Read the article-style registered confirmatory study →

Open the full registered confirmatory results and provenance →

How to interpret the registered result

What the result supports

The frozen Protocol v2 ran to completion on the analyzable confirmatory months, with a small point estimate favoring the registered directed-road graph model.

What it does not support

The registered evidence does not demonstrate superiority of DGT Directed Road + Adaptive over the Temporal MLP at +1h. The confidence interval crosses zero and the primary p-value is 0.5000.

What should not be inferred

This is not proof that graph structure is useless, and it is not evidence of universal equivalence. The inference is bounded by the registered months, models, node-eligibility rules, horizons, seeds and tests.

Research-integrity point: the non-significant primary result is preserved as registered. No post-hoc replacement months, hyperparameter tuning, alternative model search, or substitute significance test is used to overturn H1.

Confirmatory data record

Used months

2024-05 and 2024-11 satisfied the frozen feasibility and training-coverage rules.

Coverage exclusions

2024-02 and 2024-08 were excluded under the preregistered 64-node / 98% training-coverage rule; exclusion records are archived.

Exploratory separation

January 2025 was used during exploratory development and remains explicitly excluded from confirmatory inference.

Istanbul road and traffic sensor network as a graph

Graph representation of Istanbul traffic sensors, road corridors and Bosphorus crossings

Conceptual visualization of Istanbul traffic locations and road dependencies as a graph.

Research workflow

Permanent research archive

Published engrXiv Preprint 10.31224/7986

Published preprint of the registered confirmatory multi-season study.

OSF Registration DOI 10.17605/OSF.IO/FM5R7

Public preregistration and Protocol v2 update record.

Zenodo Version DOI 10.5281/zenodo.21916357

Versioned software/research release v0.1.0.

Confirmatory evidence archive archive/confirmatory-v2

Permanent registered statistics, provenance, graph diagnostics and SHA-256 integrity records.