From XXLTraffic to EvoXXLTraffic: Scaling Traffic Forecasting to Sensor-Evolving Networks

Researchers introduce XXLTraffic and EvoXXLTraffic, new datasets spanning 27 years of California and Australian traffic sensor data that account for real-world network growth. Unlike existing benchmarks assuming fixed sensor networks, these datasets expose the challenge of forecasting across dynamically evolving road infrastructure with sensor growth rates exceeding 10,000%, and reveal that current state-of-the-art models fail to generalize under such conditions.

Traffic forecasting research has traditionally operated under an unrealistic constraint: the assumption of static sensor networks. Real infrastructure continuously evolves as cities expand, roads are added, and monitoring systems are upgraded. This disconnect between academic benchmarks and operational reality has allowed researchers to optimize for stable conditions that rarely exist in practice. The introduction of XXLTraffic and EvoXXLTraffic addresses this critical gap by exposing multi-year evolution patterns across California's PeMS network and Transport for NSW data, with some districts experiencing sensor proliferation exceeding 10,000%.

The significance lies not merely in dataset size but in its realistic portrayal of continuous network change. The fixed-sensor subsets enable traditional multi-year forecasting studies, while EvoXXLTraffic's streaming protocol treats each calendar year as a distinct task within an evolving infrastructure context. This design exposes fundamental limitations in current spatio-temporal graph neural networks and other SOTA methods, which were trained and validated under static assumptions.

For the transportation intelligence and smart-city sectors, this work has immediate practical implications. Organizations deploying forecasting systems must now contend with the reality that models trained on historical fixed-network data will degrade as infrastructure changes. The benchmark's revelation that established methods fail on evolving networks suggests significant opportunities for developing robust continual learning approaches.

Looking forward, research should focus on adaptation mechanisms that gracefully handle sensor network evolution without complete model retraining. The dataset serves as a foundation for developing truly deployable traffic forecasting systems that account for infrastructure dynamics, making it valuable for both academic advancement and real-world transportation management applications.

• →XXLTraffic dataset spans 27 years of traffic data, enabling research on multi-year forecasting gaps with realistic sensor network evolution
• →EvoXXLTraffic exposes dramatic sensor growth rates (305% to 10,000+%) across nine districts, revealing previously hidden forecasting challenges
• →State-of-the-art static spatio-temporal models fail significantly when evaluated under dynamic sensor-evolving conditions
• →Yearly streaming protocol reframes traffic forecasting as a continual learning problem rather than a static prediction task
• →Dataset bridges academic benchmarking and operational reality, exposing gaps in current approaches to infrastructure-scale forecasting