https://github.com/nixtla logo
#neural-forecast
Title
# neural-forecast
f

Farzad E

02/21/2023, 4:49 PM
Is there a way to pass exogenous features to AutoNHITS?
c

Cristian (Nixtla)

02/21/2023, 4:55 PM
Hi @Farzad E. Yes, you need to create your own hyperparameter grid (example here: https://nixtla.github.io/neuralforecast/examples/automatic_hyperparameter_tuning.html), and specify the exogenous variables using the appropriate hyperparameter.
f

Farzad E

02/21/2023, 5:35 PM
@Cristian (Nixtla) so let's say your exogenous feature is week_day. Instead of passing this as a futr_exog_list as you do with NHITS, you pass this as a hyperparameter to AutoNHITS?!
c

Cristian (Nixtla)

02/21/2023, 5:36 PM
you need to add it as one entry of the configuration dictionary. For example:
"futr_exog_list":your_list
and you can try models with/without exogenous with:
"futr_exog_list":tune.choice([your_list,None])
so its not directly to the AutoNHITS model, but to the configuration dictionary, which you then pass it to the AutoNHITS with the
config
parameter
👍 1
You also need to have those variables in the dataframe
Also, for the hyperparameters you DONT specify in the config the model will use their default value (from the
NHITS
model). You need to specify all the hyperparameters you want to optimize.
👍 1
f

Farzad E

02/21/2023, 5:43 PM
Thanks. I was confused since I knew config was for hyperparameters from the examples and I knew there were specific entries for exogenous features but I didn't know we could pass them to config. This is very helpful.
c

Cristian (Nixtla)

02/21/2023, 5:45 PM
sorry if it was not clear in the documentation. We will add this in the examples!
👍 1
f

Farzad E

02/21/2023, 5:50 PM
No worries. Documenting is very time consuming and your team has made great progress. Thanks for your contributions to open source.
62 Views