search

simonpf / quantnn

Quantile regression neural networks
Other
15 stars 7 forks source link

Question. Quantile crossing problem? #3

Open FlashPlayer13 opened 2 years ago

FlashPlayer13 commented 2 years ago

Hi, thank you for this project!

I am trying to figure out - can we do smth to solve crossing problem? I mean when for example 49th quantile is more than 50th.

simonpf commented 2 years ago

My goto solution would be to regress on the PDF using a density regression neural network (DRNN). There's an implementation available in quantnn.drnn. Except that you need to provide bins instead of quantiles, it implements the same interface as the QRNN class so you should be able to just swith them out.

@adriaat has mentioned he wants to look into isotonic regression but AFAIK there's no implementation of that available.

adriaat commented 2 years ago

Following up on @simonpf 's comment, linear isotonic regression is available in sklearn.isotonic but nothing is implemented in quantnn. How to avoid quantile crossing is an open question in itself and there are many papers with suggestions on how to deal with it.