Open mitchelloharawild opened 4 years ago
Croston's method is not based on a stochastic model. See https://robjhyndman.com/publications/croston/
Consequently it is not possible to have a generate.croston
function which is consistent with CROSTON forecasts.
Instead, it would be better to use a stochastic model for intermittent demand. See the tscount package for some examples. It would be nice to create an interface between fable and the tscount models.
I went over the tscount
paper and it looks very interesting for my application (demand planning for quite sporadically sold inventory - luxury perfumes and such).
I'll try it outside of fable to assess how useful it would be for us but since the entire codebase was now rewritten to use fable I'll probably write some bridge. I'll be happy to contribute it back if the quality will be on par with your expectations.
A fable.tscount
would be great and I'm happy to help you write this if it is useful for you.
This will be extremly useful for generating prediction intervals using simulation as described here https://otexts.com/fpp3/aggregates.html
How would you judge the difficulty of implementing this?