Drift

Distinguish data drift, a change in the input distribution, from concept drift, a change in the relationship between inputs and the target, which is the more damaging and harder to detect. Input drift is measurable without labels using distributional distances or a discriminator trained to separate training from live data; concept drift requires outcome labels, which usually arrive late or not at all. Practical responses are proxy monitoring on prediction distributions and downstream outcomes, scheduled retraining, and alerting on absence of expected signal rather than only on error.