static_features documentation

[marcozanotti]

Description

Hi everybody,

I am trying to understand how static_features are effectively used by neuralforecast models, but also in mlforecast.

Unfortunately, I don't see any general documentation on that. There are a few articles on Nixtla's website explaining how to use different covariates (i.e. static, historical, and future), but I was not able to find an explanation on how they are embedded in models (except for TFT model).

I think it would be very helpful to have a general documentation page explaining how static_features are embedded into both mlforecast and neuralforecast objects.

Thank you for your help!

Link

No response

[marcopeix]

Here's how it works, let me know if it makes sense:

• Static features are passed via a separate static_df DataFrame.
• Internally, they are stored in their own tensors when creating the TimeSeriesDataset object.
• If a model supports static features, we pull the features and repeat them over each sliding window (when the model is univariate). This ensures the data is aligned. If the model is multivariate, we keep the original shape [n_series, n_features].
• Then, a model consumes the static features. How it's done depends on the model's architecture, but what usually happens is that we concatenate past time steps and features and feed that to the forward method.

[marcozanotti]

Hi @marcopeix ,
Thank you for your reply. Sure, it makes sense. Just one more question: are the scalers applied also to the static features? For instance, I fit an MLP with scaler_type = 'robust'.
Thanks.

[marcopeix]

Yes, you can do that via the local_static_scaler_type when initializing your NeuralForecast object. See here.

Let me know if you have more questions, or feel free to close the issue if all is good! 😄