Dynamic Baseline (Anomaly Detection)

Concept

A dynamic baseline is the expected value of a metric for a given moment — same-weekday trailing median, trend-adjusted, season- and event-aware — plus a tolerance band. Alerts fire on deviation from expectation rather than on crossing a fixed line.

Layman Explanation & Analogy

Instead of "alert if CAC crosses ₹400" (wrong within a quarter as the business grows), the baseline asks "is today weird for a Tuesday, given the last four Tuesdays and the fact that it is Diwali week?". Only genuine surprises page anyone.

Worked real-world example

Yesterday's CAC vs the median of the same weekday over the prior 4 weeks: +38% deviation on a matured window with no calendar event fires an "investigate" alert — with the auto-decomposition (CPM flat, CVR down) attached so the human starts from a hypothesis.

What people get wrong & common traps

Never evaluate efficiency alerts on immature windows (conversion lag makes yesterday always look "down"), never alert humans about unactionable trivia, and review false positives monthly — an unreviewed channel trains itself to be ignored. Detection should be deterministic rules; let narratives explain, not decide severity.

Related terms