This message was deleted.

Hi! So it depends on the feature. Temporal features are scaled, but static features are not. So, weekday or a category are not scaled.

How does neuralforecast know which are temporal? For instance gen_forecast and weekday are future exogenous features in the tutorial. How does neuralforecast know to scale gen_forecast but not week_day?

Good catch! So it is a known issue on our side. Temporal features are those passed as target, future exogenous and historical exogenous, so they will be scaled. The static features are not scaled. It's not the best way to handle it, but we are working on a solution.

Ok. So just to confirm, currently, all features except for the static ones are scaled? In a future release, you'll let us differentiate categorical from real features across future, historical, static, etc?

Sure, you can add issue! Thanks!

So the best would currently be to classify week days in

static_exog_list

(contrary to the tutorial)? Or in the

fut_exog_list

and accepting it will be subject to scaling)? Would the same approach also apply to cyclically encoded time features (e.g, day of week sin-cos transformed)?

Even if week days are scaled, they keep their order, meaning that the value of Tuesday is always higher than Monday. As mentioned, anything in static is not scaled, but everything in futr and hist is scaled.