Changepoint detection is the task of spotting the moment a metric's behavior genuinely shifts — as opposed to normal noise wobbling around.
Also known as: change detection
Changepoint detection is the task of finding the point in time where a data stream stopped behaving the way it used to. A model's accuracy starts sliding, a latency distribution creeps up, an eval score quietly regresses. The hard part is not noticing that today looks different — noise makes every day look a little different — it is knowing when the difference is real.
Done naïvely, this is the peeking problem in disguise: keep watching for 'the point it changed' and pure noise will eventually hand you a false one. That is why a trustworthy detector needs an error guarantee that survives continuous watching, not just a rule that trips on the first large wobble.
ValidAnytime is changepoint detection built on anytime-valid evidence: an e-process accumulates support for 'something shifted' point by point, and the alarm fires the moment that evidence is decisive — with a false-alarm rate you can actually name, not one that erodes every time you look.
ValidAnytime turns these ideas into a live alarm you can trust — valid no matter how often you look. Prove it on your own data, free.