In this review of categorical encoding, it states early on that
For regression tasks, Target and LeaveOneOut probably won’t work well
and later repeats that Target/LeaveOneOut (Owen Zhang's approach) encoders are best suited for classification tasks.
I could not seem to find other sources to support this claim, and I was wondering, why wouldn't LeaveOneOut encoding be a valid approach for regression tasks? Is it because it then causes overfitting to certain features? But then, how would this be different from a classification task?