https://github.com/nixtla logo
m

Mark Raphael

04/12/2022, 5:54 PM
Yes, I had some questions about using the neuralforecast package. First question I had was on which model to use. I have a time series of daily frequency with yearly and weekly seasonality, and there are also holiday effects I would need to include as exogenous variables. I was thinking of using NBEATSx or NHITS to model the series, but given that I'm only forecasting 7 days out would there be any benefit to using NHITS over NBEATSx? In my understanding the benefits of NHITS are mainly for longer horizon forecasting? Thank you
c

Cristian (Nixtla)

04/12/2022, 5:56 PM
Hi Mark, thanks for using our package! Yes, for short horizons the benefits of NHITS over NBEATSx diminish. For a forecasting horizon of 7 I suggest using the NBEATSx first.
m

Mark Raphael

04/12/2022, 6:02 PM
Awesome, thanks for developing the package and the quick reply! If I could ask a follow up, in order to capture the yearly seasonality, would you recommend using a long n_time_in (365) or including a representation of the day of year as another exogenous variable? (I was thinking that the latter would work better)
c

Cristian (Nixtla)

04/12/2022, 6:04 PM
no problem! yes, I suggest adding the day of year as an exogenous variable
m

Mark Raphael

04/12/2022, 6:04 PM
Great, appreciate the answers here
🙌 1