Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hello <@U05EK0TTXM1>! There are two types of exogenous temporal covariates: historical and future.

If you have the data available for the future you need to specify these variables in the `futr_exog_list`. The model already uses all the timestamps from the exogenous covariates in the input window (given by `input_size`), even for this future variables. So you dont need to add lags :slightly_smiling_face:

The best way would be to train both versions of the model. Including the exogenous covariates and without including them. The NHITS model should take only minutes to train, are you observing very long training times?

hi, my situation is similar and i have those covariates in the historic df and future df.  i found them useful in algos like regression trees as these algos do not have automatic detection or generation of features like lags, mas and differencing.  i do notice auto nhits, nbeatsx, mlp and tft take longer time to train eg more than 120 mims on a 8 core vm.  i dont have access to gpu atm.