This gets tricky. Once you look into your past, you're presumably looking at data that was used to generate your training corpus. So you would expect better acc

This gets tricky. Once you look into your past, you're presumably looking at data that was used to generate your training corpus. So you would expect better accuracy on that than you would find on present/future predictions.