Dear Nixtla team! I have a question regarding the ...
# mlforecastt
thomas delaunait
05/18/2024, 8:06 PMDear Nixtla team! I have a question regarding the newly released AutoXgboost. Is it possible in the CV process to specify step_size as it is possible in statsforecasts? It seems only possible to specify n-windows and h. Thank you for your help
j
José Morales
05/20/2024, 5:56 PMHey. It's not currently possible, it currently uses a rolling window. Can you describe your use case for the step_size? Usually having the same samples repeated across CV folds biases the results towards those
t
thomas delaunait
05/20/2024, 8:51 PMThank you José for the answer. We would like to proceed as folowing :
thomas delaunait
05/20/2024, 8:51 PMthomas delaunait
05/20/2024, 8:53 PMFor example, in case of monthly forecast, we keep the last 12 months for CV. And we perform 12 folds of CV.
thomas delaunait
05/20/2024, 8:57 PMThe horizon is not always 12, but is rolling. Like this, we can proceed CV and it do not requires a lot of data. The issue with current CV in MLforecast is that it requires a lot of data. For example a n_windows=12 and h=12 requires 145 months. But if you have any best practices, thank you veru much for your help.
j
José Morales
05/20/2024, 9:10 PMDo you retrain the models?
José Morales
05/20/2024, 9:12 PMAlso, what exactly is your goal? Do you train the models and use those to predict for 12 months without retraining? That CV scheme is strange