Hey, thanks to the Nixtla team... we are using all of their time series in production, however we've been having a debate on the best way to structure our models from a "unique id" point of view: We are predicting hourly time series, and thus our "unique id" is the store combined with a given hour; however there are multiple ways to think of this problem. How does Nixtla deal with multiple time series? Does it create 1 model for all unique_ids, or is it training separate models for each time series? The basic issue is the formation of a unique_id formed with certain categorical variables versus just having everything as a feature. In our example, would the unique_id just be the store and then everything else just be a feature? Further, this unique_id determines whether you are predicting at an hourly level, daily level, etc. Would love to hear thoughts from the community on this topic! Thank you.

Hi Mark, I'll do my best to answer without seeing your data! The way to think of the unique_id is just a label for the series. In your case, it seems that a store identifier is enough. As for the features, this is something you have to experiment. For example, suppose you have stores in different citites. Your unique_id could be a store id plus the name of the city, or you could leave the city as a feature. For that, you have to test and see if the city feature actually helps. As for how our libraries handle multiple series, it basically trains a global model for all series. However, you can try multiple models (and you should) on your dataset, and then select the best model for a particular series. We show how to do this in section 6 of this guide. Finally, your unique_id doesn't determine the frequency of your data. This is a parameter that you provide when fitting the models. So make sure to set the right frequency or resample your data to the frequency that you want. I hope this helps!

@Marco incredibly useful. Thank you!

@Marco Isn’t statsforecast training individual models for each time series?