Start with predictors selection (e.g. most 100 correlated predictors), then do cross validation is WRONG !

We should divide the data first.