Covariate Shift

$P(X)$ moves and $P(Y \mid X)$ holds. A new marketing channel brings younger customers, so you see a different mix of people, and everything you learned about how age relates to the outcome is still correct.

Which means it often needs nothing. The model remains right wherever it has support, and the real risk is that it's now extrapolating into regions your training data never covered — so the useful check is coverage, not the drift score.

When a correction is warranted, the principled one is importance weighting: reweight training rows so their distribution matches production. Worth distinguishing from its sibling, label shift, where the class balance moves while each class looks the same — fraud rates rising in a recession — which recalibration or a threshold change fixes rather than retraining.