Prediction shift: the leakage problem your gradient boosting model has and your CV under-detects
Chapter 2.4 of Mastering CatBoost opens with a definition I want to start from, because it is the technically-correct one and it is also the one that hides what is actually happening:
