[FEAT] Add support for historical and static exogenous in TimeXer

[marcopeix]

TimeXer now supports all types of exogenous features.

Example usage:

Y_train_df = AirPassengersPanel[AirPassengersPanel['ds'] < AirPassengersPanel['ds'].values[-12]].reset_index(drop=True)
Y_test_df = AirPassengersPanel[AirPassengersPanel['ds'] >= AirPassengersPanel['ds'].values[-12]].reset_index(drop=True)
futr_df = Y_test_df.drop(columns=["y", "y_[lag12]"])

models = [
    TimeXer(
        h=12,
        input_size=24,
        n_series=2,
        hist_exog_list=["y_[lag12]"],
        futr_exog_list=["trend"],
        stat_exog_list=['airline1'],
        max_steps=20,
        scaler_type="robust",
    ),
]

nf = NeuralForecast(
    models=models, 
    freq="ME", 
)

nf.fit(
    df=Y_train_df,
    static_df=AirPassengersStatic
)
fcsts = nf.predict(futr_df=futr_df)

[JQGoh]

A couple of questions and requests:

• Suggest to add a minimal example of test using exogenous variables for in tests/test_models/test_timexer.py. Consider examples like 1) X with static exog, 2) X without static exog
• I notice that you include pydantic dependency in this PR, is it because you ran into Ray-related failure and this helps to resolve it? If that is the case, do you still remember what kind of Ray error?

Otherwise, the PR LGTM

[marcopeix]

@JQGoh

1. Sure, I can add a test, good idea!
2. Right, I added pydantic because of ray issues. The error message was

ImportError: `ray.train.v2` requires the pydantic package, which is missing. Run the following command to fix this: `pip install pydantic`

So I added the dependency to fix the issue.

[elephaint]

Minor comment, the dependency we need is ray.train, that is missing. So, I'd add the ray.train dependency

[marcopeix]

For the tests, I agree we should have them, but I noticed that none of the models that support exogenous features have tests for exog features. I can make a separate PR to add a global test for models that support exog feat, something similar to the current test_model. And keep this PR only for TimeXer. Wdyt?

[JQGoh]

yes, I also notice that. Sounds good to me

For the tests, I agree we should have them, but I noticed that none of the models that support exogenous features have tests for exog features. I can make a separate PR to add a global test for models that support exog feat, something similar to the current test_model. And keep this PR only for TimeXer. Wdyt?

[elephaint]

LGTM, nice!

non-blocking - you can use "check_model" from .common._model_checks with "losses" to test models against each loss function, this typically uncovers bugs (I didn't run it on this PR, just "fyi" in case you didn't know)