You've hit on a key insight. Predicting the future (even by a bit) turns out to be a very powerful learning signal for building models of the world.

It won't work on a traditional feedforward neural network but if you have feedback everywhere it appears to work.