Train in Peacetime, Test in Wartime
The first time most people split data into training and validation sets, they shuffle everything together and randomly hold out a portion for testing. For many tasks, that’s fine. But the moment your data carries a time dimension, especially when you hit a special period like the COVID pandemic that upends the entire environment, this approach can quietly let your model cheat, and the good-looking scores will only make you more confident about it. ...