https://github.com/nixtla logo
m

Manuel

03/21/2023, 5:15 PM
How does neuralforecast behave with sequences shorter than input_size? Are they padded or skipped/ignored or masked in some way?
c

Cristian (Nixtla)

03/22/2023, 3:16 PM
Hi @Manuel. Yes, the time series are padded with zeros at the beginning
m

Manuel

03/22/2023, 3:23 PM
Thanks @Cristian (Nixtla)! If short sequences have exogenous variables (e.g. futr_exog), how are they handled? Do they get filled with zero as well?
c

Cristian (Nixtla)

03/22/2023, 3:24 PM
all temporal variables are padded at the beginning of the time series
m

Manuel

03/22/2023, 3:45 PM
So, theoretically, instead of letting everything automatically pad with zero, might it be better to manually pad the time series with zero and create an exogenous binary variable such as "is_padding" to indicate to the model that that is a padding?