Hi @chododom, internally RNN has an output_chunk_length=1, meaning during prediction when n>output_chunk_length it will always perform auto-regression (consuming it's own predictions for future predictions) to forecast n points.

training_length just means how of those auto-regressive predictions it performs for one sample during training.

You can do two things:

• simply increase n from 4 to 8 when calling predict()
• multiple training:
  • train a model1 with the lower training_length
  • save the model with model1.save(...)
  • create a new model2 object with the same hyperparameters but different training_length
  • load the weights from model1 save into model2 with model2.load_weights(...) (since training_length changed, for this to work you need to set skip_checks=True, and load_encoders=False).
  • train model2

@dennisbader Thank you for the advice! Do you think this will also work with the TFT model? If I understand correctly, the TFT has a built-in prediction horizon length.

@adesso-dominik-chodounsky, any of Darts' torch models (neural networks) and regression models support auto-regressive predictions with n>output_chunk_length. The only requirement is that if you use past/future_covariates, you must know these far enough into the future to generate the n prediction points.