Comment by jprafael

I've found that it works well to add the prediction horizon as a numerical feature (e.g. # of days), and them replicate each row for many such horizons, while ensuring that all such rows go to the same training fold.