CutoffFhSplitter

Temporal train-test splitter, based on cutoff and forecasting horizon.

Train and test splits are determied as follows:

for each cutoff point k=cutoff[i], in split:

• training fold is all loc indices up to and including k

• if fh is not passed, test fold is all loc indices strictly after k

• if fh is passed, test fold is all loc indices in k + fh, if fh is

relative.

More precisely, fh.to_absolute_index(cutoff=k) If ``fh` is absolute, then the test window is fh itself.

It should be noted that, unlike in CutoffSplitter, test folds are not determined by a window length, but by indices of the forecasting horizon fh, i.e., test folds can be non-contiguous, even if the data index is regular.