Spike Detection

Spike detection is identifying a sudden, abnormal rise or drop in a metric, distinct from ordinary fluctuation, so its cause can be reached sooner.

A spike only has meaning once normal is defined; a large number alone is not a spike unless measured against the average and dispersion of that same metric at that same time. Spike detection is therefore first of all a baseline problem: seasonality must be accounted for (nights and weekends have their own patterns) or every natural peak triggers a false alarm. The two directions of a spike carry two different meanings: an upward spike is usually news — a campaign landed, a link went viral, or a bot began crawling — while a sudden drop is often danger, the first sign of a technical error, an outage or a broken tracking script. The big trap is treating every spike as real growth; much of a sudden traffic spike comes from a bot or a spam referral rather than a human audience, so every spike should first be verified for its source and whether it is human.