Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hey team! Is there a list, enum or something where I can check which `date_features` are implemented by default in mlforecast? And is there a overview about how they are treated? For instance is the `dayofweek` one-hot-encoded, ordinal, cyclic etc.?

Hey. We don't implement them ourselves, we extract the attributes from the datetime objects (pandas or polars), so you can see the lists here:
• <https://pandas.pydata.org/docs/reference/api/pandas.DatetimeIndex.html|pandas>: under the Attributes section
• <https://docs.pola.rs/py-polars/html/reference/expressions/temporal.html|polars>

You can also define a function that takes the datetime column and returns another column (<https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/custom_date_features.html|guide>)

We return the attributes as integers, so if you want to one hot encode them you can use a pipeline (<https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/sklearn_pipelines.html|guide>)

Thanks <@U03UX9KR75W>!

I am training models using LGBM, is it possible to use the build in `categorical_featuers` parameter in the library together with MLForecast?

The `categorical_features` argument defaults to `'auto'` , which means to treat every pandas categorical column as a categorical feature. So the easiest way would be to have those columns as categoricals. You can do that either with a <https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/sklearn_pipelines.html|scikit-learn pipeline> or with <https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/custom_date_features.html|custom date features>