Feature: New group aware time-based cross validation

[jrosell]

There are situations when sliding_period or group_vfold_cv is not exactly what is needed for a good cross validation.

I've seen some libraries in python implementing GroupTimeSeriesSplit:

• mlxtend.evaluate.GroupTimeSeriesSplit
• sklego.model_selection.GroupTimeSeriesSplit

They are based on some kaggle notebooks and stackoverflow answers:

• Cross-validation for grouped time-series (panel) data
• Found the Holy Grail: GroupTimeSeriesSplit

I also see that this feature is currently being developed in scikit-learn:

• Feature request: Group aware Time-based cross validation

[hfrick]

Thank you for the issue @jrosell ! We'll probably not come around to this in the immediate future but I'll leave the issue open for later (and for people to upvote if appropriate).