Inputs Look The Same

A fraud model has run in production for two years. Input monitoring is clean: every feature histogram for 2026 matches 2023 within noise, and the drift test on $p(x)$ never fires. Accuracy has still fallen from 0.94 to 0.71. Auditing newly labeled cases shows that transactions with the SAME feature values are now fraudulent about three times as often as they were in 2023.

In Gama et al.'s taxonomy, which drift is this, and what follows from it?

Overfitting that only became visible in production; refit the same 2023 data with a stronger regularization penalty.

Real concept drift: $p(y \mid x)$ changed while $p(x)$ stood still. Reweighting the 2023 rows cannot repair it, because the old labels are now wrong for those inputs; the model needs freshly labeled data.

Virtual drift: $p(x)$ is unchanged, so the input-label relationship cannot have changed either, and the drop must come from a pipeline bug.

Covariate shift: reweight the 2023 training rows by the density ratio $p_{2026}(x)/p_{2023}(x)$ and refit.