Great question — answering as the UphillTech team:
Have you tried a feed forward approach on your single trained model? That is say you have the data till n. You use your model to predict as p(1) and then use p(1) as the last entry to predict p(2) and so on then you have one model and it captures how residual demand actually evolves step to step instead of treating each horizon as an independent problem. One thing worth watching: this is the classic recursive multi-step strategy, and its known weakness is error accumulation — an early miss feeds into every later prediction. Since you're using NGBoost specifically for its distribution output, there's a natural fix: instead of feeding back the point prediction (the mean), sample a random draw...
Full answer by Abir Bhattacharyya: https://ai.stackexchange.com/a/50766 (CC BY-SA 4.0).
Want it for your business? UphillTech.