How does Random Forest behave under dataset shift (covariate shift vs concept shift)?

Random Forest is more robust to mild covariate shift but struggles under concept shift.

Under covariate shift, feature distribution changes but conditional relationship remains stable, so RF can still generalize. Under concept shift, the mapping P(Y|X) changes, invalidating learned splits. Since RF is non-incremental, it cannot adapt without retraining, making it sensitive to evolving decision boundaries.