Robust Time Series Monitoring: Anomaly Detection Using Matrix Profile and Prophet

Robust time series monitoring, particularly for anomaly detection, is a complex task due to the inherent noisiness of system metrics and the evolving definition of "normal." At Sentry, the AI/ML team developed a hybrid anomaly detection system that combines Matrix Profile and Meta’s Prophet to address these challenges. Matrix Profile, an unsupervised technique, identifies anomalous shapes in time series data, while Prophet handles seasonality and trend forecasting to reduce false positives. The system uses Summary Statistics Subsequence (SuSS) for optimal window size selection, and an AnomalyScorer module that converts raw data into actionable scores, reducing false alerts. Designed to scale, the system processes data in batches and streams, providing fast predictions while offering users configurable sensitivity levels. This approach allows for precise, context-aware anomaly detection, improving alert reliability and system monitoring.