FEAT: TimeXer

[marcopeix]

Add TimeXer to neuralforecast. This is a transformer-based model that supports future exogenous features, and seems to perform very well in long-horizon forecasting according to here.

Original code implementation: https://github.com/thuml/TimeXer/tree/main
Paper: https://arxiv.org/abs/2402.19072

Note: The encoder and embedding are specific to TimeXer, that's why I am not reusing the existing common module from other models.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[elephaint]

LGTM, maybe the refactoring in a separate PR?

We can also do it separately though, it's non-blocking imho.

Small detail re. mint.json.

Does it (somewhat) reproduce the paper results?

[elephaint]

LGTM!

[marcopeix]

LGTM, maybe the refactoring in a separate PR?

We can also do it separately though, it's non-blocking imho.

Small detail re. mint.json.

Does it (somewhat) reproduce the paper results?

Here are the results for ETTm1:

Paper (avg for 4 horizons)
MAE: 0.397
MSE: 0.382

neuralforecast (avg for 4 horizons)
MAE: 0.423
MSE: 0.445

Although I was training for 1000 step with early stopping on all horizons and they all reached 1000 steps, so maybe training for longer would give closer scores