Skip to content

Conditional inference, boosted

CTBoost is a gradient-boosting library built around conditional-inference trees. It combines alpha-stopped feature-test selection, optional per-node Bonferroni adjustment, a familiar Python and scikit-learn interface, native C++ training, and optional CUDA acceleration.

Install CTBoost See benchmark status

The current release, 0.1.61, is available on PyPI and GitHub. It reduces CPU inference overhead and fixes Huber/Quantile edge cases, NumPy conversion compatibility, and GNU FMA training-rounding differences. Saved-model formats, learning defaults, and conditional-inference feature selection are preserved.

Project status

CTBoost is an alpha project. Its API and model format are tested extensively, but it does not yet have the independent production history of CatBoost or XGBoost. Benchmark claims on this site state their dataset and protocol scope.

As of 9 September 2026, the fresh 0.1.61 Full HPO25 evaluation is prepared but has not started; the 0.1.60 run is retired. The latest downloadable author-run TabArena bundle measures 0.1.58. See benchmark status for the separate maintainer report, evidence, and limits.

  • :material-chart-tree:{ .lg .middle } A different tree learner

    Conditional tests select features before cut points are optimized. CTBoost preserves that ordering across its supported learning options.

  • :material-speedometer:{ .lg .middle } Native CPU and CUDA paths

    Histogram training, stochastic boosting, DART, early stopping, callbacks, ranking, survival objectives, and task-aware feature preprocessing.

  • :material-shield-check:{ .lg .middle } Auditable artifacts

    Versioned inference manifests, deterministic fingerprints, JSON/Python/C++/ONNX exports, exact empirical interventional TreeSHAP, and reproducible benchmark metadata.

  • :material-language-python:{ .lg .middle } Python ecosystem fit

    NumPy, pandas, SciPy sparse, Arrow, Polars, cuDF/CuPy/DLPack adapters, scikit-learn helpers, and Dask/Ray/Spark integration surfaces.

A small example

from ctboost import CTBoostClassifier

model = CTBoostClassifier(
    iterations=800,
    learning_rate=0.05,
    max_depth=6,
    ordered_ctr=True,
    random_seed=42,
)
model.fit(
    X_train,
    y_train,
    eval_set=(X_valid, y_valid),
    early_stopping_rounds=50,
)

probability = model.predict_proba(X_test)
explanation = model.predict_shap_values(X_test, X_train)

Where to go next