@davidgilbertson, thanks a lot for the various bug reports, improvement ideas, and code snippets!
May we kindly ask, if you were to set a priority from the entire space, would it be still this? #4776
That is, ensure that the following use case is supported:
- direct reduction with arbitrary
fh
- with exogeneous data
X
X features including categorical variables- user specified lags on
X, or no lags on X (choice)
- user specified feature set used in the
sklearn estimator
- global and local should both work
- optinally: option to apply further transformations to
X
requirements:
- variable names of transformed variables (or equivalent information) must be easily specifiable, optimally also inspectable
- performance must be decent (comparable to current
make_reduce with similar estimator without specifying categorical variables
get_fitted_params should be inspectable with clear, identifiable, reference to transformed variables, e.g., when LinearRegressor is used
The above being a combination of feature requests (e.g., variable name inspection) and bugfix (potential ways to treat this combination case don't fully work yet).
@davidgilbertson, thanks a lot for the various bug reports, improvement ideas, and code snippets!
May we kindly ask, if you were to set a priority from the entire space, would it be still this? #4776
That is, ensure that the following use case is supported:
fhXXfeatures including categorical variablesX, or no lags onX(choice)sklearnestimatorXrequirements:
make_reducewith similar estimator without specifying categorical variablesget_fitted_paramsshould be inspectable with clear, identifiable, reference to transformed variables, e.g., whenLinearRegressoris usedThe above being a combination of feature requests (e.g., variable name inspection) and bugfix (potential ways to treat this combination case don't fully work yet).