Hi everyone. What is the best for performing global transformation on exogenous features? For example, I want to try a MinMax scaler to every exogenous variable. Is there a way to make this easy, so that it is applied internally everytime we pass the future dataframe? Thank you very much!

Hey. You can probably do this with scikit learn estimators

Got it, thank you! I was wondering if there was something similar to the neuralforecast

local_scaler_type

, but I guess there is not.

i have seen people scaling using tree based models. to my knowledge there is no need for that, so unless you want to use a classical regression like ridge or so, i think you don't really need it

Thanks. Yes, I wanted to have a well-performing classical regression model as a "basecase" to compare against.

can you combine MLP with mlforecast? i would have thought it is part of neuralforecast

I was following this tutorial as a starting point: https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/prediction_intervals.html It uses MLP with mlforecast

cool, i was not aware 🙂 thx

@jan rathfelder you can use anything that follows a scikitlearn fit-predict pattern including creating your own class. And generally scaling isnt a thing for trees but for global time series forecasting you definitely want to at least have scaling as part of your hyperparameter tuning and view differencing as a type of scalar as well.

thx