search

StatMixedML / XGBoostLSS

An extension of XGBoost to probabilistic modelling
https://statmixedml.github.io/XGBoostLSS/
Apache License 2.0
554 stars 58 forks source link

Statistical test for distribution fitting #75

Closed hadarshavit closed 4 months ago

hadarshavit commented 4 months ago

Thanks for the great package! Currently, when fitting data to a distribution only the NLLH loss is returned. I would like to have some statistical test result as well, like the Kolmogorov–Smirnov test. Any hint how I can get that?

Thanks!

StatMixedML commented 4 months ago

@hadarshavit Thanks for your interest in the project.

I would like to have some statistical test result as well, like the Kolmogorov–Smirnov test.

Can you please provide an example of what you want to do exactly. I suppose you want to use the KS for comparing the fitted vs. the empirical distribution?

hadarshavit commented 4 months ago

Yes, given a fitted distribution, we currently get the NLLH loss. I would like to have the option to the the KS pvalue result as well.

StatMixedML commented 4 months ago

Yes, given a fitted distribution, we currently get the NLLH loss. I would like to have the option to the the KS pvalue result as well.

Do you refer to an unconditional fit or to a fitted model including features? I suppose it is the former one, since the parameters for a fitted model are functions of the input features. So each observation has its own distribution.