data and concept drift
/ DAY-tuh and KON-sept drift /
Drift is the slow, often invisible way a deployed model goes stale because the world it was trained on has moved on. There are two flavors. Data drift means the inputs change: the model still sees the same kind of question, but the population it's asked about shifts — a model trained on last year's shoppers now sees younger ones with different habits. Concept drift is deeper: the relationship between input and correct answer itself changes — 'normal' spending patterns after a sudden economic shock, or what counts as 'spam' as scammers invent new tricks.
A clean way to feel the difference: imagine predicting house prices. Data drift is when the houses being sold change (suddenly more apartments than before) — same rules, new mix of inputs. Concept drift is when the rules change (a new policy makes location matter far more than size) — the very meaning of a good prediction has shifted. The first is a change in what you see; the second is a change in what the answer should be.
Why it matters: drift is the single most common reason a model that launched great gets quietly worse over time, and it never shows up as an error — predictions keep flowing, just less and less correctly. The honest truth is that drift is not a bug to be fixed once but a permanent condition of the real world; almost every production model degrades eventually. There is no cure, only management: monitor for it, and retrain on fresh data when it appears. A model is a snapshot of a moving world, and the world does not hold still.
A model that flagged fraudulent transactions worked well until scammers changed tactics. The inputs (transactions) still look superficially normal — no data drift alarm — but what makes one fraudulent has shifted. That's concept drift: the answer changed even though the questions look the same.
Concept drift is the sneaky one — the inputs look fine, but the right answer quietly changed.
Drift is not a defect you can engineer away once and for all; it is the default fate of any model facing a changing world. The realistic goal is not to prevent it but to detect it early and retrain — treating model maintenance as a permanent ongoing cost, not a one-time launch.