Overview
Anomaly detection systems fail in one of two directions: they miss genuine anomalies (false negatives that allow fraud, failures, and quality issues to go undetected) or they generate so many false positives that the operations team stops trusting the alerts and ignores them. Both failures are design failures — the threshold was set without considering the false positive budget, and the method was chosen without considering the anomaly type.
Effective anomaly detection requires matching the method to the anomaly type: point anomalies (single observations that deviate from the norm), contextual anomalies (observations that are normal globally but anomalous in context), and collective anomalies (sequences of observations that are anomalous together). Each type requires a different detection approach.
The Anomaly Detection System Design Prompt generates a complete anomaly detection specification: anomaly type classification, method selection matrix, threshold calibration to the false positive budget, alert triage framework, and a feedback loop that improves detection accuracy over time.
What you get: - Anomaly type classification (point/contextual/collective) - Method selection matrix by anomaly type and data structure - Threshold calibration to false positive budget - Alert triage and prioritization framework - Feedback loop for continuous improvement
Built for: data scientists and ML engineers building anomaly detection systems for fraud, quality control, infrastructure monitoring, and operational alerting.