Also doing random not sequential training and evaluation. The proper way to eval time-series models.