following core dev discussion on slack, I propose we simply deprecate and remove the current forecasting module and replace it with a cleaner, lightweight version from September onwards. I have gone through it all here
1491
the code is nearly all just wrappers for other packages. It came with us from the bad place, and there has been no interest in it from developers or users. These wrappers are highly complex, in my opinion over engineered and generally very inefficient (e.g. "VectorizedDF"). Forecasting is well served for packages such as nixtla etc and I dont see what extra this adds to the current ecosystem.
Describe your proposed solution
we should deprecate with a long lead time, then simply remove the forecasting, datatypes and possibly a subset of the transformers. Then we can discuss a more intuitive stripped down base class for forecasting as a series transformer in line with the other series estimators: transformers, anomaly detection and segmentation
I think deprecate in the next release and take it out in the one after. We could consider going to 1.0.0, but i dont think that really matters.
Describe alternatives you've considered, if relevant
TonyBagnall
commented
2 months ago
Describe the feature or idea you want to propose
following core dev discussion on slack, I propose we simply deprecate and remove the current forecasting module and replace it with a cleaner, lightweight version from September onwards. I have gone through it all here
1491
the code is nearly all just wrappers for other packages. It came with us from the bad place, and there has been no interest in it from developers or users. These wrappers are highly complex, in my opinion over engineered and generally very inefficient (e.g. "VectorizedDF"). Forecasting is well served for packages such as nixtla etc and I dont see what extra this adds to the current ecosystem.
Describe your proposed solution
we should deprecate with a long lead time, then simply remove the forecasting, datatypes and possibly a subset of the transformers. Then we can discuss a more intuitive stripped down base class for forecasting as a series transformer in line with the other series estimators: transformers, anomaly detection and segmentation
I think deprecate in the next release and take it out in the one after. We could consider going to 1.0.0, but i dont think that really matters.
Describe alternatives you've considered, if relevant
No response
Additional context
No response