Time series forecasting with non-temporal information (exogenous features)

Question

I'm reviewing many time-series algorithms and libraries, such as Prophet, darts, auto_ts, etc.

All libraries discuss univariant time-series (where the task is to forecast based on a single time-series) and multivariant (where multiple time-series are available, and learning all series together might help at inference), but no library deals with non-temporal information.

I.e., Given that we wish to forecast customer expenses in the store, and we know their previous purchases, as well as their age, address, job description, etc.

These non-temporal features might also help in the prediction, but I don't see any known platforms exposing the ability to fuse these insights inside the model training.

Am I missing something?

Answer 1

Because that isn't strictly time series anymore. For example: the address variable. In the time series context if they haven't moved during your time frame then that variable would be a constant. So you are comparing across entities, while just time series involves one entity across time.

Typically that can be solved with GAMs or simply trying to 'featurize' the time part with seasonal dummies and a simple trend variable (or piecewise trend) and passing everything to a regression or something.