rocci¶

Distribution-free simultaneous confidence bands for ROC curves.

Envelope band vs Working-Hotelling band on heavy-tailed scores

rocci is a simple interface for adding uncertainty estimates to your ROC curve. It draws a simultaneous confidence band, which maintains the specified confidence of capturing the entire true (population) ROC — not just each point one at a time. The default is a new nonparametric method that works with nearly all common data distributions.

from rocci import roc_band

band = roc_band(y_true, y_score)   # labels + scores, any common container
band.plot()
print(band.summary())

That is the whole quickstart. See it annotated in Getting started, or run the scikit-learn vignette end to end.

Why rocci¶

rocci is designed to:

Just work. It does the right thing off the shelf for nearly any data set: no normality assumption, safe under heavy ties and discrete scores, and honest where no distribution-free bound exists.
Drop in to your workflow. Native integration with scikit-learn, torch, statsmodels, PyMC/arviz, and pandas/polars data — ingestion is duck-typed, with zero hard dependencies on any of those libraries.
Be fast. The bootstrap kernel is compiled Rust (a pure-NumPy fallback keeps the package working everywhere): 2 000 bootstrap replicates on 100 000 samples in well under half a second.
Make a minimal footprint. The only hard dependency is numpy; plotting is an optional extra (rocci[plot]).
Clear an unreasonably high bar of rigor. The method is validated with millions of simulations across diverse data sets; the implementation is verified with exacting tests.
Support an open ecosystem. Permissive MIT license, easy extensibility.

If you are comfortable adding a normality assumption to get a tighter band, normal=True gives the parametric Working–Hotelling band — and rocci checks the assumption and warns when it looks doubtful, as in the figure above, where the true curve escapes the parametric band entirely. Which band should I use? explains the trade.

Installation¶

pip install rocci            # prebuilt wheels (no need for rust toolchain)

# optional plotting support
pip install 'rocci[plot]'

Details in the installation guide.

Where to next¶

Reading the band — what "simultaneous" buys you, and what the vacuous region at tiny FPR means.
The envelope method — how the band is built.
Simulations and validation — the evidence that the method works where the classical bands fail.
How rocci is verified — the case for trusting the numbers.
API reference — the full public surface (it's small).

Citing¶

If rocci contributes to a publication, please cite it — see CITATION.cff in the repository. band.summary() ends with the same pointer.