Distribution shift after a market event

A deployed model fails after a market-structure change. How would you distinguish covariate shift, target shift and a changed predictive relationship?

Audit pipeline integrity first. Then compare feature distributions, outcome distributions and conditional performance using labels as they mature. Feature-distribution change alone does not identify whether the conditional relationship changed. Evaluate calibration and residuals by relevant groups and periods, allowing for sampling uncertainty. Choose monitoring and retraining rules on development evidence; automatic retraining can amplify data errors or transient noise.