Hi birinhos,
I agree with you on the fact that, making predictions based on the previous prediction it self in “long-term” is a bad idea. But I do it only for 50 steps, and uses true data to build up the memory as well, which usually works well in practice.
I’m not sure how predicting the variance would help? With the variance they can only know if the predictions going to fluctuate or not (which it usually does) in the future (not even the direction it’s going to go in — up/down). Can you elaborate what you meant?