Hello everyone, I am fitting global models with neuralforecast and I am confused about the scaling. I see that it is possible to specify the scaler type in two ways: 1.

NeuralForecast(models:List[Any], freq:Union[str,int], local_scaler_type:Optional[str]=None)

2.

MLP(..., scaler_type:str='identity', ...)

What is the difference between the two (if there is any)? What happens if I specify both? Thank you for your help.

Hello! Using

scaler_type

defined in the model will scale windows individually. Using

local_scaler_type

scales your dataset entirely (including features) before making the windows and training. So, I would suggest you use one or the other, but not both. Usually, we define

scaler_type

only, it tends to make better predictions. But if your dataset has global information that is important, then

local_scaler_type

makes sense. I hope this helps!

Hi @Marco thank you for your answer. Now it is more clear. Is there any scaler transformation suggested in the context of global product demand forecasting?

I'm not very experienced with demand forecasting, but I would check the difference in performance when setting

scaler_type

or

local_scaler_type

, and then you can try different scalers from there. Does that make sense?

Yes, thanks!