I've been trying to get good references on how to solve a problem that's been bothering me regarding the modelling techniques I've used. I'm currently interested in making forecasts using ML for specific combinations, e.g. sales channel+sku. It is given that each series on this set for the given combinations will have different behavior and therefore will perform better with different lag/fourier features and different models for detrending/deseasonalizing/forecasting.

My point is, what would be a good pipeline to get the best possible options within a certain span, an in an automated way? And how would one define this span? I understand that the latter question depends heavily on the problem we're dealing with and that this question is far from trivial, but I'd like to get something better than pure intuition and trial&error.

I understand that a possibility would be to train a global model based on the data of all combinations but that, for now, is not on my plans due to time constraint issues. One strategy that I think is interesting is to create a synthetic feature and exclude features that are less relevant to the model than it, but I also have not found good references on how to do this, like how to define this synthetic feature according to the series.

If you guys have any ideas/references on how to deal with this issue it'd be very much appreciated.

  • $\begingroup$ sklearn has feature selection algorithms. You could use them as-is, or alternatively define your entire pipeline in an sklearn-compatible way to benefit from cross-validated feature selection (i.e. the selection will be more robust). $\endgroup$ Commented May 19 at 15:18


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.