NGBoost

0.5.10 verified Mon Apr 27 auth: no python

NGBoost is a library for probabilistic prediction via gradient boosting that estimates full probability distributions rather than point predictions. Current version is 0.5.10, supporting Python 3.9–3.14. It uses natural gradient boosting to learn parameters of a parametric distribution. Release cadence is periodic, with recent versions adding distribution support (Weibull, HalfNormal, Beta, etc.), numpy2 compatibility, and sklearn compatibility. Maintained by the Stanford ML Group.

pip install ngboost

Common errors

error ImportError: cannot import name 'NGBRegressor' from 'ngboost' ↓

cause Incorrect installation or trying to import from a different environment.

fix

Run pip install ngboost and ensure you are in the correct Python environment. Then use from ngboost import NGBRegressor.

error ModuleNotFoundError: No module named 'sklearn.metrics._pairwise_distances' ↓

cause Incompatible scikit-learn version (likely <0.22) with newer ngboost.

fix

Upgrade scikit-learn: pip install --upgrade scikit-learn

error AttributeError: 'NGBRegressor' object has no attribute 'pred_dist' ↓

cause Using an older version of ngboost where the method is called `predict_dist` instead of `pred_dist`.

fix

Use ngb.pred_dist(X) for ngboost>=0.5.0; for older versions use ngb.predict_dist(X).

error ValueError: When `natural_gradient` is True, `Score` must be an instance of `LogScore` or a class that implements `natural_gradient`. ↓

cause Attempting to use natural_gradient with a custom Score that doesn't support it.

fix

Either set natural_gradient=False or use a built-in Score like LogScore.

Warnings

breaking Version 0.5.5 upgraded sklearn dependency to >1.6. If using older sklearn, upgrade or pin ngboost<0.5.5. ↓

fix Update scikit-learn to >=1.6 or use ngboost<0.5.5

breaking Version 0.5.2 added support for Numpy 2.0. Code relying on numpy<2 behavior (e.g., object dtype in arrays) may break with ngboost>=0.5.2. ↓

fix Pin numpy<2 if using older scikit-learn or custom code that is not numpy2-compatible.

deprecated The 'natural_gradient' parameter in some internal methods was deprecated in v0.5.8 and may be removed in future. ↓

fix Avoid using 'natural_gradient' parameter; default behavior is correct.

gotcha Default distribution is Normal with LogScore. If you use a different distribution, ensure the Score matches the distribution (e.g., LogScore for exponential family). Mismatch leads to silent poor performance. ↓

fix Always specify both Dist and Score parameters explicitly.

gotcha Models saved with sklearn <1.3 may not load with sklearn >=1.3 due to pickle format. ngboost>=0.5.10 includes a compatibility loader `load_ngboost_model`. ↓

fix Use `from ngboost import load_ngboost_model` and call it instead of joblib.load.

Imports

NGBRegressor
```
from ngboost import NGBRegressor
```
NGBClassifier
wrong
```
from ngboost.ngboost import NGBClassifier
```
correct
```
from ngboost import NGBClassifier
```
Common error: trying to import from submodule instead of top-level
Normal
wrong
```
from ngboost import Normal
```
correct
```
from ngboost.distns import Normal
```
Distributions are in ngboost.distns, not top-level.

LogScore

wrong

from ngboost import LogScore

correct

from ngboost.scores import LogScore

Scores are in ngboost.scores submodule.

Quickstart

Basic usage of NGBoost for regression with a Normal distribution and LogScore.

import numpy as np
from ngboost import NGBRegressor
from ngboost.distns import Normal
from ngboost.scores import LogScore

# Generate sample data
np.random.seed(42)
X = np.random.randn(100, 1)
y = X[:, 0] + 0.1*np.random.randn(100)

# Train probabilistic model
ngb = NGBRegressor(Dist=Normal, Score=LogScore, n_estimators=100, learning_rate=0.01, verbose=False)
ngb.fit(X, y)

# Predict: mean and variance (or scale) of Normal distribution
y_preds = ngb.predict(X)
# Use ngb.pred_dist(X) to get distribution object