Use XGBoost as base estimator

[fif911]

First of all thanks for such a great library and your continuous support. I am working on predictions with uncertainty using the NGBoost Regressor and want to improve the accuracy of predictions. My assumption is in the base estimator is better the whole algorithm may perform better.

Do you think it's possible to use a more advanced algorithm than the DecisionTreeRegressor e.g. XGBoost or even NN?

Best regards

[manningkyle304]

See xgboost distribution! I have found that ngboost tended outperform it slightly, but maybe you can find better results. it's also up to 10x faster

[alejandroschuler]

The usual understanding for boosting is that you want "weak" learners: decision trees are an ideal choice because they are fast to fit and are nonparametric if you allow depth to vary. Theoretically you can use a NN in NGboost but there's not much point, might as well do a generative model with deep learning. But you're welcome to try it, nothing stops you from seeing how it performs.