Should the dataset be split first and then normalized or vice-versa? If I have fresh data that I want to predict an outcome for, should I also normalize it before I fit the Predict Model?