Predicting Traffic Jams With AI: Why Better Accuracy Is Only Part of Real Challenge

Getting an early warning about a traffic jam could help commuters avoid delays and give transport teams time to respond before queues spread across a city. A systematic review published in Encyclopedia, titled 'Artificial Intelligence and Machine Learning for Road Traffic Congestion Prediction and Forecasting: A Systematic Review of Methods, Validation, Explainability, and Reproducibility,' examines how far the technology has progressed and how much evidence exists to support its use. Author Yasmany García-Ramírez of Ecuador's Universidad Técnica Particular de Loja reviewed 129 studies, finding a rapidly expanding research field whose results remain difficult to compare and translate into dependable everyday tools.

Traffic Forecasting Research Is Growing Fast

The review followed PRISMA 2020 reporting guidance and searched Scopus, Web of Science Core Collection and IEEE Xplore for English-language journal articles and full conference papers published between 2000 and 2026, with searches running through July 5, 2026. Two external reviewers independently screened records and assessed available full texts, with the author resolving disagreements against predefined criteria. The search identified 1,131 records, leaving 734 after duplicate removal and 129 studies in the final review.

The included studies were published between 2009 and 2026, with 111, or 86%, appearing from 2020 onwards. Urban road networks dominated the evidence, accounting for 98 studies, and network-level forecasting was the most common scale among the 105 studies with that information recorded. Every included investigation was classified as a methodological study, making the review a picture of how prediction systems are developed and evaluated rather than proof that they have already reduced congestion in daily use.

Researchers drew on sources such as traffic sensors, GPS traces, cameras, connected vehicles, weather records and incident information to capture the changing conditions that shape road travel. Models ranged from decision trees, Bayesian approaches and conventional neural networks to recurrent networks, convolutional networks, graph-based systems, transformers, reinforcement learning and combinations of several techniques. Systems increasingly brought together information about changes over time and connections between roads, reflecting the way a slowdown can spread through an entire network.

Predicting a Traffic Jam Means Different Things

The studies looked at different aspects of traffic congestion, such as how serious a jam might become, whether one would happen or how it could spread to nearby roads. Predicting the level of congestion was the most common task, but researchers did not all define a traffic jam in the same way. The review could clearly identify what was being predicted in 111 of the 129 studies.

Researchers most often measured traffic by counting vehicles or checking their speed. Among the 76 studies with clear information about these measurements, 42 used vehicle flow and 33 used speed, with some using more than one measure. Other studies examined journey times, how closely vehicles were packed together or how much road space they occupied. Accurately predicting the number of vehicles does not automatically show whether a traffic jam will form, so a system may need more information to give a useful warning.

The studies also differed in how far ahead they predicted traffic conditions, with 102 giving at least one clear forecast period. Predictions covered less than 15 minutes in 59 studies, 15–60 minutes in 60 studies, more than one hour and up to six hours in 52 studies, and more than six hours in 45 studies. These numbers overlap because 71 studies tested several forecast periods. Another 24 used general descriptions such as "short-term" or "long-term" without giving a clear duration, and three had no forecast period recorded.

Impressive Accuracy Claims Need Stronger Evidence

Individual papers reported that their models performed better than the alternatives they tested, but this does not show that the same models would work better in every city. One study could also use or compare several types of AI, causing it to appear in more than one model category. A frequently used approach is not necessarily the most accurate or the most useful for a particular road network.

The review recorded methods for explaining how an AI system reached its predictions in 23 studies, or 17.8%, and identified systems capable of working with incoming traffic data in 43 studies, or one-third of the total. It could not reliably establish how many studies tested their models in independent settings, checked whether they worked in other locations, measured uncertainty or shared their data and code. Missing information in the review's records does not prove that researchers skipped these steps.

Access to research papers was another major limitation, as the reviewer could not obtain 195 of the 339 full reports requested, or 57.5%. Those missing papers might have affected the findings, but their influence remains unknown. The collected information was also insufficient to compare results across countries or formally rank study quality, and automated data collection needed human checking to correct errors. Recording each paper in a single model row and a single results row could also leave out details from studies that tested several different approaches.

What Cities Need Before Relying on AI Forecasts

Transport agencies need to define the decision a forecast will support, whether that involves issuing a congestion warning, estimating travel time or anticipating disruption across connected roads. The required warning period, available data, computing resources and consequences of missed or false alerts should guide model choice. Pilot projects need to check response speed, resilience to missing data, maintenance demands, integration with existing systems and stability as traffic patterns change.

Trustworthy evaluation requires testing on independent periods, corridors, networks or cities, with forecasting tests respecting the order in which observations occur. Researchers need consistent definitions, clearly documented data preparation and comparisons using the same inputs, horizons and computing budgets, including strong simpler models. Sharing code, configurations and evaluation procedures can make findings easier to reproduce, with controlled access or executable examples helping when data cannot be released.