rocci 0.2.0

Distribution-free simultaneous confidence bands for ROC curves

rocci

Distribution-free simultaneous confidence bands for ROC curves.

rocci is a simple interface to easily add uncertainty estimates to your ROC curve that are very likely to be correct in nearly all use cases. It draws a simultaneous confidence band, which maintains the specified confidence of capturing the entire true (population) ROC.

rocci is designed to:

• just work. It should do the right thing off the shelf for nearly any data set.
• drop in to your workflow. It natively integrates with sklearn, torch, statsmodels, PyMC/arviz, and pandas/polars data.
• be fast. Core operations are implemented in rust for speed (invisible to the user).
• have a lightweight footprint. Minimal runtime dependencies (just numpy and scipy).
• support an open ecosystem. Permissive MIT license, easy extensibility.

By default, rocci uses a distribution-free method that provides informative bands without burdensome assumptions about the data. It maintains nominal coverage in a huge variety of contexts and the rare violations tend to be small misses — validated in a 2.25M-evaluation simulation study across Gaussian, heavy-tailed, skewed, and multimodal score distributions (see the validation repository and the docs' walkthrough of its conclusions). If you are comfortable adding a normality assumption to get tighter bands, rocci yields a "Working-Hotelling" band, but also carefully checks the normality assumption and warns you when it looks dicey.

Installation

pip install rocci            # prebuilt wheels — no Rust toolchain needed
pip install 'rocci[plot]'    # + matplotlib for band.plot() and diagnostics
uv add rocci                 # in uv-managed projects

Wheels cover Linux (glibc x86-64/aarch64 and musl), macOS (Intel and Apple silicon), and Windows, for every Python ≥ 3.10; runtime dependencies are numpy and scipy only. On any other platform pip falls back to the sdist (requires a Rust toolchain), and if no compiled kernel is present at runtime a pure-NumPy backend with identical statistical semantics takes over automatically. There is no conda package at present: the wheel is lightweight and self-contained, so pip install rocci works cleanly inside conda environments. Details: installation guide.

Quickstart

from rocci import roc_band

band = roc_band(y_true, y_score)
band.plot()
print(band.summary())

Docs: https://ndelaneybusch.github.io/rocci · Changelog: CHANGELOG.md · Contributing (including the release process): CONTRIBUTING.md