Model Drift

No jargon, no math

A model is a frozen photograph of the patterns in its training data. Score it on tomorrow's data and the match is never perfect, but it usually holds. Drift is when it stops holding. There are two flavours. Data drift: the inputs change, like a product catalogue doubling after an acquisition, so the model meets items it never saw. Concept drift: the inputs look the same but the relationship changed, like customers reacting to a price rise. Either way the effect is identical: predictions quietly degrade while the pipeline keeps serving them. You catch it by monitoring two things in production: the statistical shape of incoming data, and outcome metrics once real answers arrive. The remedy is a new training round on recent data, or a rollback. Where it breaks: drift is invisible until you measure outcomes, and by then the damage is done, so monitoring must run continuously.