Please note that I'm aware of the fact of logic of forest-based average result of trees addressed here. Considering the size of my sub-data (8640, 1) and knowing consumption CPU and memory from this post, still with runtime is so high with even n_estimators=5, n_jobs=-1 with 5 trees in Google Colab medium! So I have to set n_estimators=1 because of I have many time data samples like df1, df2,..., df10000 needs to be forecasted.

can you renew the link

Don't expect much miracles from AI & predictive models. However, SARIMAX and gated-units models could work, but depending on the potential patterns, one can forecast a few steps ahead.

If the matter concerns out-of-sample forecasting in a certain time horizon and X, Y, and Z might be related, we need to see dependencies within each one and between pairs (if there are). Sometimes, there could be no pattern, and it's not predictable, or you need to see the problem from another side of the glass, which means unpredictability of data using dependency tests in the past daily resolution of observations.

This has already been commented on by @Nikos M. and looks like this answer is redundant.