Split by Time, Not by Luck

Random train/test splits look rigorous until your model meets next month’s data — temporal splits mirror how predictions actually get used.