Anomaly Detection Basics: Z-Scores, IQR, Seasonality & Alert Precision
Flag fraud and outages the way monitoring and risk teams do: z-scores, robust z with median and MAD, IQR fences, rolling ±kσ bands, seasonal baselines, multivariate distance, precision/recall against labelled incidents and the cost of alert fatigue, in 8 graded labs.
The first lesson is free — no account needed.
About this course
An anomaly detection course that flags fraud and outages the way monitoring and risk teams do, on one fictional payments company's data in a spreadsheet, for data analysts, fraud and risk analysts, SRE and monitoring engineers and data science students. You standardise card payments and flag |z| ≥ 3 and see how one huge outlier masks another, catch the hidden one with a robust z-score built on the median and MAD, set Tukey's IQR fences on API latency, draw rolling mean ± 3σ bands around a traffic metric, compare a same-hour-last-weeks seasonal baseline with a flat rule, combine three standardised features into one distance score, score a detector against labelled incidents with precision, recall and F1 at each threshold and pick the threshold with the lowest cost of false alarms and missed incidents that an on-call team can handle. The first lab is free.
Lessons
- Z-Score Flags: How Far from Normal Is a Transaction?Free
- Robust Z-Scores with the Median & MAD Pro
- IQR Fences: Box-Plot Outliers in Server Latency Pro
- Rolling Mean ± k·σ Bands for a Live Metric (with a Chart) Pro
- Seasonal Baselines: Same Hour, Last Weeks Pro
- Multivariate Outliers: Standardised Distance (Mahalanobis-Lite) Pro
- Labelled Anomalies: Precision & Recall at Each Threshold (with a Chart) Pro
- Alert Fatigue & Cost: Choosing the Threshold Pro