Time series anomaly detection finds unusual points or patterns.
Point anomalies: Single unusual values. Server CPU spike.
Contextual anomalies: Unusual given the context. High sales on a Tuesday (normal on Black Friday).
Collective anomalies: Unusual sequences. Gradually drifting sensor.
Methods:
- Statistical: Z-score, IQR after detrending
- Model-based: Forecast and flag large residuals
- Isolation Forest: Works well for multivariate
Interview question: "How detect anomalies in seasonal data?"
Remove seasonality first (STL decomposition), then detect anomalies in residuals.