Ensemble with models from different library

[sigord]

Hello everyone!
I am new to time series forecasting and have been given an assignment in this area. Luckily my closed environment included the gluonts 0.8.1 and sktime 0.8.0 libraries. I want to build an ensemble in sktime with sktime models, some neural networks from gluonts, xgbregressor from sklearn/xgboost. I am looking for any guidance on how to do this.

Thanks for creating this useful lib.

[fkiraly]

hm, sktime 0.8.0 is ages ago, are you sure you don´t mean sktime 0.18.0?

Now this could be a too old version of sktime, but in general you can take the extension templates in order to create sktime compatible interfaces for any library's forecaster. The forecaster template is here: https://github.com/sktime/sktime/blob/main/extension_templates/forecasting.py

There is also an in-depth video tutorial, which is also linked in the guide here:
https://www.sktime.net/en/stable/developer_guide/add_estimators.html

xgbregressor should work together with make_reduction, as it already has an sklearn compatible interface, for that you need to use the sklearn API: https://xgboost.readthedocs.io/en/stable/python/python_api.html#module-xgboost.sklearn

[sigord]

Thanks a lot for the quick reply! Yes, unfortunately I meant 0.8.0. Updating and adding libraries is expensive in some large companies.

[fkiraly]

😱

[fkiraly]

anyway, if you are up for adding an interface to gluonts to sktime, that would be appreciated (if you are doing it anyway?)