Lightning Talk: Anomaly Detection - Robust Strategy
summary
Algorithm builds prediction bands from a median over one day and an approximate median absolute deviation (MAD) times 2, with an experimental mad_over_time for persistence. It adds a short-term smoothing signal called level = median over one hour, which requires roughly 30 minutes of sustained change to move. Result: short spikes are ignored, sustained spikes trigger anomalies with delayed onset and longer duration. Compared with the prior adaptive method, that one fires immediately on spikes but widens bands and can miss longer anomalies.
AI-written summary · about
related talks
34mAnomaly Detection in Time Series from Scratch Using...
45mAnomaly Detection in Time Series
18mCross-Industry Anomaly Detection Solutions with R and Shiny
30mRealtime Time Series Anomaly Detection in Production
29mREAL-TIME ANOMALY DETECTION MADE EASY
43mDetecting Anomalies in Streaming Data
39mMastering real-time anomaly detection with open source tools
38mWatch Everything, Watch Anything: Anomaly Detection