Anomaly Detection in Time Series

Technical note · Time series · Work in progress

These notes summarize basic concepts in anomaly detection for time-series data, with a focus on machine learning methods.

1. Problem setting

Anomalies are observations or temporal patterns that deviate from expected behavior. In time series, anomalies can appear as point anomalies, contextual anomalies, collective anomalies, or distributional drifts.

2. Common approaches

Reconstruction-based methods: learn to reconstruct normal data and use reconstruction error as an anomaly score.
Forecasting-based methods: predict future values and detect anomalies through prediction errors.
Representation-based methods: learn latent representations that separate normal and abnormal behavior.

3. Evaluation challenges

Evaluation is difficult because anomalies are rare, labels may be incomplete, and the temporal localization of events can be ambiguous. Metrics should therefore be selected carefully depending on the application.