It doesn’t predict next frame, it interpolates two frames. It takes the most recent frame, previous frame, the motion vectors associated with them and tries to produce approximation that is close enough so your brain is deceived.

All that at the price of annoying input lag that’s inherently one frame bigger than normally.