I have a panel dataset (multiple time series indexed on IDs and time) in which I want to perform multi-step forecasts (e.g. 5-steps forecasts). An example of the dataset (pandas dataframe) is the following:
IDs, time, f1, f2, ...
1 0 4.1 50 ...
1 1 3.3 44 ...
1 2 2.6 11
2 0 2.1 79
2 1 4.9 56 ...
2 2 0.1 11
... ... ... ... ...
However, i don't know how to train my models and specifically how should I split my data (train/test set). The end goal is to perform 5-steps forecasts given as inputs to the trained model x-length windows.
I was thinking to split the data as follows: 80% of the IDs would be in the train set and 20% on the test set and then to use sliding window for cross validation (e.g. using sktime's SlidingWindowSplitter).
Therefore, what would be a good training strategy to follow?