Beta-distribution

[ChristianMichelsen]

Hi and thanks for a great package!

I have two small questions.

1. What would it require to implement a beta-distribution to fit fractions or probabilities?

2. I see that a very similar library called XGBoost-LSS also exists. What are the differences between the two libraries?

Cheers,

[CDonnerer]

Hi, In terms of adding new distributions, the ngboost docs offer info on what is in general required for a given distribution. However, it sounds like the use case here is a classification problem, for which ngboost (or this package) are not a good choice, see discussion here or here.

[ChristianMichelsen]

The use case is very much not classification, but rather regression, however, regression where the target is a continuous variable between 0 and 1. This also what's known as beta-regression in the statistics community, see e.g. this R package. This package, however, only allows for GLM-style models, and I had hoped to use XGBoost to model the beta-distribution parameters, similar as to how xgboost-distribution models e.g. n and p in the negative binomial model.

The specific usecase is similar to predicting how full individual train departures will be (ie. beta regression), and not just whether or not the train will be full (classification).