A Short-Term Anomaly Forecasting Model for Univariate Time Series and Its Use in the Sfera.IAS Product

Artem Zorin24 May 2025

Problem

Operations teams watch business services and infrastructure through time series: requests per minute, tickets processed, CPU and disk utilization. A static threshold fires only after the metric has broken. Dynamic thresholds, built as an uncertainty interval around a forecast, fire earlier but still only when the actual value leaves the interval.

Example from the article: in the Sfera.IAS monitoring product, mean time to detection (MTTD) is 4 minutes for basic monitoring and 20 minutes for dynamic thresholds. A short-term forecast raises it to 2 hours.

Solution in brief

Build two forecasts of the same series: a long-term one, retrained rarely, and a 2-hour short-term one, retrained on fresh data and corrected by its recent error. If the short-term forecast leaves the uncertainty intervals or diverges from the long-term one, alert that a degradation is about to start. The base model is a seasonal regression trained without anomaly labels on CPUs.

Questions this case answers
  • How to alert on a degradation before a univariate metric leaves its dynamic thresholds by comparing a short-term forecast retrained on fresh data with a long-term forecast
  • How to monitor tens of thousands of time series in real time on CPU without labeled anomalies
  • How to detect service degradation early with a two-horizon short-term forecast on univariate time series
  • How to alert on an anomaly before the metric leaves its dynamic thresholds

Tools and methods

  • Two-horizon forecast comparison — alert trigger
  • Generalized additive regression (Stan) — base forecaster for seasonal series
  • Forecast error correction — shifts the short-term forecast by its recent error
  • VictoriaMetrics, ClickHouse — stores for real-time metrics and forecasts
  • NeuralProphet — planned base model
Continue reading — it's free

Google or a one-time email code. No password.