A production model's KS statistic drops from 0.65 to 0.41 over 6 months. PSI for input features remains below 0.10. What is the most likely explanation?
Select an answer to reveal the explanation.
Short Explanation
Here's the deal — you're seeing concept drift because stable PSI under 0.10 says input distributions haven't moved, but KS dropping from 0.65 to 0.41 says the model's discriminative power has. The relationship between features and target has changed.
Full Explanation
The correct explanation is concept drift: stable PSI below 0.10 for input features indicates the input distributions have not shifted, yet the model's KS statistic has declined from 0.65 to 0.41, showing reduced discriminative power. That pattern points to a change in P(Y|X), meaning the relationship between features and target has evolved. The feature distribution shift explanation is wrong because feature PSI is stable. The implementation drift explanation is wrong because modified scoring code would typically cause a sudden change, not a gradual decline. The label drift explanation is wrong because a changed positive-class definition would affect ground-truth labels in the monitoring evaluation.