Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Image from iOS.jpg

Hi <@U03DDFKMT0X> , 
The most important data split is the time series cross validation one:

sorry just to clarify the question, when you're running hyperopt, does it create new data loaders on every trial?

and you're probably anticipating the follow up question :slightly_smiling_face: are the splits identical for every  train / valid / test loader?

Some experiment settings depend on things like batch size or the sampling frequency

Yes the train/validation/test splits are the same as the one in the picture

Some preprocessing of the windows might be different depending on hyperparameters
As long as you maintain the ‘ds_in_test’ or the ‘mask_df’ parameters constant the split is respected

gotchya, so you fix the index range for each split (i think it's called `get_mask_dfs` )?

Cool, if you find anything weird in the splits tell us. It is super important

seems to work as intended, but i did notice in `model_fit_predict`, line 893 on tag `0.0.9` that `f_cols` doesn't get passed down to `fit`. looks like it's on main as well. i could be wrong but that might be blocking exogenous variables from being used in auto-tuned models

The ‘f_cols’ is a placeholder, it is intended to distinguish past exogenous variables from future exogenous (available at prediction time)
If you are using exogenous have in mind that they are used as future exogenous by default

ah, it's not used for anything at the moment?