Forecasting models are essential for network traffic management and anomaly detection, yet assessing their long‑term adaptability requires multi‑year data that captures annual seasonality and behavioral drift. Existing public datasets are too short for such horizon‑spanning tasks.
We introduce the CESNET-TimeSeries26 dataset, which consists of two parts:
- 995 days of continuous traffic time series from the CESNET3 network;
- Approximately 4.5 years of data for a single subnet. The collection covers 278,466 IP addresses, 285 institutions, and 566 institutional subnets, and is provided at three aggregation intervals (e.g., 5 min, 1 h, 1 day).
This dataset offers a definitive longitudinal benchmark for evaluating forecasting models in both operational management and security monitoring. Our analysis shows that annual seasonality dominates the periodic patterns; a model can exploit this component only when its training window spans at least one full year, a condition unmet by sub‑year datasets.
The benchmark enables testing model robustness to concept drift, long‑term anomaly detection, and comparative studies of forecasting approaches in realistic network environments.
Review