This message was deleted.

Hey @Manuel, Thanks for using NeuralForecast. Can you open a github issue to have this in my TODO list? In the past we had

valid_loss=loss

, as a default for which we did not include the possibility to include

loss

, as a tunable hyperparameter. At this point we have the ability to determine the

valid_loss

independently but have not taken out the protections in the

AutoBase

class, here: https://github.com/Nixtla/neuralforecast/blob/main/neuralforecast/common/_base_auto.py#L67 In the meantime what I have observed with the number of components is that it follows a classic bias-variance u shape behavior like the ones in this plot.