A forecasting team evaluates a demand model by randomly mixing records from earlier and later weeks. What evaluation design better matches future prediction?

1. Sort the rows by target and use the highest targets for training only
2. Train on earlier periods and validate on subsequent periods
3. Put every row in both training and validation
4. Train on future weeks and validate only on earlier weeks