The current version of MLForecast does not support applying custom weights during cross-validation training. The custom training feature described in the documentation is only available for the final model training, not for the cross-validation process.
This is a limitation of the current implementation, as the cross-validation in MLForecast is designed to use the default training procedure without any customization. The cross-validation is used for model selection and hyperparameter tuning, where the ability to apply custom weights may be beneficial for certain use cases.
Use case
If I want to apply some weight on samples during training, in order to improve my forecasts on a given set of dates/periods, I need to use the custom training feature you explained in this guide https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/custom_training.html.
However, in the case I want to perform model selection over cross validation scores of multiple models, I need to be able to reproduce the same customed training during cross validation. This is where I am currently blocked, as the current version of MLForecast does not support applying custom weights during the cross-validation process.
Description
The current version of MLForecast does not support applying custom weights during cross-validation training. The custom training feature described in the documentation is only available for the final model training, not for the cross-validation process.
This is a limitation of the current implementation, as the cross-validation in MLForecast is designed to use the default training procedure without any customization. The cross-validation is used for model selection and hyperparameter tuning, where the ability to apply custom weights may be beneficial for certain use cases.
Use case
If I want to apply some weight on samples during training, in order to improve my forecasts on a given set of dates/periods, I need to use the custom training feature you explained in this guide https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/custom_training.html. However, in the case I want to perform model selection over cross validation scores of multiple models, I need to be able to reproduce the same customed training during cross validation. This is where I am currently blocked, as the current version of MLForecast does not support applying custom weights during the cross-validation process.