Feature request: Recursive multistep multivariate forecasting and direct multistep multi-series forecasting about skforecast HOT 4 OPEN

Recursive forecasting available for multivariate forecasting

This is quite challenging because this hybrid strategy requires each individual series to be predicted internally so that the predicted values can then be used as input features by a second model. We have skipped this so far because the idea of this approach being too computationally expensive, but we could do some experimentation to test its feasibility.

from skforecast.

Hello Kishan,

Thanks for your comments,

Direct forecasting available for multi-series forecasting.

This is something we have pending. Since we are still developing new features for ForecasterMultiseries, we will wait for a more stable version before implementing it.

Recursive forecasting available for multivariate forecasting.

As it is now, the ForecasterAutoregMultiVariate is only trained in one level (series 1 or series 2) and creates one model per step (direct approach). In the figure, let's say that if you specify level = 'Series 1' when creating the Forecaster, it will only use the training matrix at the top.

Since this forecaster needs the future values of the series to predict the following steps, it needs 2 models (one for each series) to make it recursive and call them in parallel. Predict step 1 for both series, predict step 2 for both series...

Is that what you have in mind?

A consistent naming approach

Completly agree. We have this figure in the repo page and documentation to help the users with this problem:

https://skforecast.org/latest/#forecasters
https://github.com/JoaquinAmatRodrigo/skforecast#forecasters

Best,

Javi

from skforecast.

The names of the Forecasters definitely need to be revisited, the family is growing and at some point we need to make the naming more intuitive. This may require some proper alias management to avoid breaking compatibility. ¡Maybe one day we are ready for version 1.0.0. ! 😄 😄

from skforecast.

Recursive forecasting available for multivariate forecasting

This is quite challenging because this hybrid strategy requires each individual series to be predicted internally so that the predicted values can then be used as input features by a second model. We have skipped this so far because the idea of this approach being too computationally expensive, but we could do some experimentation to test its feasibility.

Yes I agree. This would be computationally challenging. I think there are other features which are easier to add to skforecast which would add more value.

from skforecast.