ngboost

Why we included this project

Most regression models return a single point estimate, which is fine until a bad guess costs you real money. NGBoost takes a different approach: instead of one forecast it fits a full probability distribution over each prediction, so you walk away with both a point value and a usable confidence interval around it. It does this by combining natural gradient descent with a gradient boosting ensemble, and it slots neatly into the scikit-learn workflow through estimators like NGBRegressor and NGBClassifier. That matters for demand forecasting, risk scoring, and other settings where the cost of an error depends on which direction it lands. The library is also modular, so you can swap in different parametric distributions and scoring rules to match your data.