Conformal k-NN Anomaly Detection for Univariate Data Streams

Anomaly detection in time-series data is an important applied task. Anomalies give essential and often actionable information in many applications. Examples can be found in such fields as healthcare, flight safety, intrusion detection and finance. In this report we consider conformal methods for describing and constructing non-parametric anomaly detection method for univariate time-series. Proposed method adapts to non-stationarity in considered data stream and provides probabilistic abnormality scores based on the conformal prediction paradigm. In current work we also propose a modification of the method based on signal pruning. This modification leads to significantly improve the quality of anomaly detection of the proposed method. The simplicity of the method allows its use in real-time applications. Testing of proposed anomaly detection method was performed on the Numenta Anomaly Detection benchmark and the Yahoo! S5 dataset, which contain both real-world and artificial time-series. Despite its relative simplicity the method performs on par with more complex prediction-based models.