Online Time Series Detection: Streams, Latency, and Drift

00:00 Online detection under streaming constraints 00:59 Why future context is unavailable online 01:44 Offline detection vs online prediction 02:30 Streaming evaluation and chronological discipline 03:12 What defines online event detection 03:54 Online detection loop and state updates 04:37 Temporal streams and visibility boundaries 05:19 Early detection vs false alarms 06:06 Why offline methods fail in streams 06:52 Static vs dynamic streaming models 07:35 Windowed memory for online detectors 08:16 Retraining with recent batches 08:57 Incremental learning under drift 09:39 Non-stationarity and concept drift 10:25 Stability-plasticity dilemma in streams 11:13 Active adaptation after drift alarms 12:01 Passive adaptation and forgetting rates 12:43 Online prediction for early warnings 13:28 Predicting events with classification 14:18 Predicting events with forecasting models 15:03 Two-stage predicted-event detection 15:44 Multivariate online anomaly detection 16:32 Distance-based stream anomaly detection 17:15 Tree-based online anomaly detection 18:03 Projection-based online anomaly detection 18:54 Latency metrics for online alerts 19:38 Hybrid detectors for streaming events 20:28 Early uncertain alerts vs late reliable alerts Video description Online time series detection identifies events in a stream when future observations are unavailable and decisions must arrive fast enough to matter. How should a detector balance causality, latency, memory, adaptation, false alarms, and drift while still producing useful alerts? This video explains why online detection is not just offline event detection made faster. It covers streaming evaluation, chronological discipline, online detection loops, state updates, visibility boundaries, early detection, static and dynamic models, windowed memory, retraining with recent batches, incremental learning, non-stationarity, concept drift, and the stability-plasticity dilemma. The lesson also introduces online prediction, classification-based early warning, regression-based forecasting for future events, multivariate stream anomalies, distance-based detectors such as COD and MCOD, tree-based methods such as Half-Space Trees and Isolation Forest variants, projection-based methods such as LODA and RS Hash, latency metrics, and hybrid detector architectures. The goal is to design detectors that are accurate, timely, adaptive, and operationally useful.