[BUG] make_reduction with direct strategy fits global models on the wrong data for forecasting horizon>1 about sktime HOT 4 OPEN

Thanks for reporting, the reasoning seems sound.
I'll need to check again in defail how this relates to _cut_df logic.

FYI @benHeid, is this related to some of the issues you looked at recently?

from sktime.

Thank you. I have one more thing: in the make_reduction documentation it's written that pooling only works for Recursive Forecaster

pooling: str {“local”, “global”}, optional
Specifies whether separate models will be fit at the level of each instance (local) of if you wish to fit a single model to all instances (“global”). Currently only works for RecursiveTimeSeriesRegressionForecaster.

In reality "global" is also implemented with Direct Forecaster (with this above issue). Is this docstring outdated?

from sktime.

In reality "global" is also implemented with Direct Forecaster (with this above issue). Is this docstring outdated?

Possibly - I thought it was functional.
@benHeid looked into this recently, so probably can say more.

from sktime.

I took a look on this Issue and I can confirm that this is a Bug. Might probably be fixed by calculating the n_timesteps twice. Before and after the transformation. Before the transformation to being still able to perform the checks that happen before transform. Afterwards, to have the correct length of the transformed time series, so that cut_df is working correctly. Perhaps we should also add a check to _cut_df that checks if n_obs is greater than the time series.

If it is greater, it just returns the input time series. See:

pd.DataFrame([0,1,2]).tail(4)
# Out:
#    0
# 0  0
# 1  1
# 2  2

pd.DataFrame([0,1,2]).tail(6)
# Out:
#    0
# 0  0
# 1  1
# 2  2

pd.DataFrame([0,1,2]).head(6)
# Out:
#    0
# 0  0
# 1  1
# 2  2

from sktime.