Forecast-based models are widely used for network traffic management and anomaly detection, but their reliability over long time horizons is rarely tested. A new preprint introduces CESNET-TimeSeries26, a dataset spanning multiple years of network traffic data, explicitly built to study annual seasonality, concept drift, and long-term anomaly detection. The authors argue that extended evaluation periods are essential to assess how well forecasting models adapt to changing network conditions.
The dataset is intended to fill a gap in publicly available benchmarks, which often cover only short windows and miss slow-evolving patterns. By providing a multi-year record, the dataset allows researchers to examine how traffic distributions shift over time and whether models remain accurate as those shifts occur. The paper's abstract emphasizes the need to evaluate models over extended periods, suggesting that the dataset is designed to stress-test both predictive accuracy and robustness to drift.
Because this is a single source, there are no conflicting viewpoints to compare. The contribution appears to be primarily a resource for the community, rather than a new algorithm or theoretical result. The authors' focus on annual seasonality and concept drift indicates that the dataset could be particularly useful for long-term capacity planning and for detecting anomalies that emerge gradually rather than as sudden spikes.
As with any new dataset, its value will depend on how well it represents real-world traffic and whether it becomes widely adopted. The preprint does not yet provide details on the exact time span or collection methodology, but the title suggests a 26-year coverage, which would be unusually long for a public network traffic dataset. If released, it could become a standard benchmark for longitudinal studies in network operations and machine learning.