"one-shot" RegressionModels

[hrzn]

Currently, our RegressionModels will create k internal models when output_chunk_length=k. This can scale poorly for larger values of k.

Another option, which (AFAIK) is often used in Kaggle competitions, is to train only one model, taking as input features that can all be computed at least k time steps in advance such as:

• Sufficiently old lags
• Sufficiently old window features
• Future covariates

Pros:

• Seems to work very well in many cases (often relying on good future covariates capturing time axis too)
• Scales very well

Cons:

• Cannot access recent lags, but this does not always matter

It'd be nice to provide such an option in Darts. It could be either an option in RegressionModel, or potentially a new kind of RegressionModel altogether.

It might be nice to couple it with adding encoders for RegressionModels, which could dynamically create the future covariates (e.g., time axis features) on the fly.

[benitocm]

Please, could you elaborate a bit more por provide a link to a kaggle example?