probcal

Universal post-hoc probability calibration for binary classifiers: methods, metrics, diagnostics, and auditable offsetting — numpy-only.

The package unifies the binary calibration literature — Platt, temperature, beta, isotonic, centered isotonic, histogram binning, scaling-binning, BBQ, ENIR, Venn–Abers, and spline calibration — with an extensive catalog of evaluation metrics and statistical tests, visualization on both probability and logit scales, an auditable logit-offset (central tendency) adjustment, automatic method selection under nested validation, and prefit / cross-validation data flows.

Status: released on PyPI, beta. Breaking changes bump the minor version until 1.0.

pip install probcal

Start with Getting started, read the Concepts chapters — the theoretical foundation of the package, written before the corresponding code — and walk the executed PD calibration tutorial. The How it works page maps the full pipeline in one view.