For retail demand with obvious seasonality, the validation window should preferably cover at least one complete seasonal cycle; otherwise, the error estimate will be distorted by differences in seasonal phase.