Hi Nixtla team. I was wondering whether you have any concrete examples of using Nixtla for Zero Inflated Poisson? I am thinking of combining classification for zero/non-zero and then regression where non-zero is predicted. Since Nixtla does not have support for classifiers this starts to become confusing and a bit hard to achieve, especially when using recursive forecasting. Any ideas or pointers would be appreciated (I couldn't really find much apart from classic stuff like

Croston

etc in the docs of

statsforecast

)

I'd train a MF model like LightGBM with Tweedie loss or a NeuralForecast model with Poisson loss directly, that usually gives best performance on intermittent data. In the docs of NF there's a tutorial on intermittent data

I’ve had pretty crap results using LGBM and tweedie so far, that was my initial idea. I’ll test out NeuralForecast with poisson loss, thanks for the pointer.

https://www.timeseriesclassification.com has got it covered, no need to reinvent the bike classification is a well solved case.