Quickstart

The five-line version:

from rocci import roc_band

band = roc_band(y_true, y_score, random_state=0)
band.plot()
print(band.summary())

y_true is anything label-like (bools, {0, 1}, {-1, 1}, or any two values plus pos_label=); y_score is anything score-like (higher = more positive). Arrays, Series, tensors, and lists all work — Using rocci with your data has the full matrix.

What you get back

roc_band returns a frozen RocBand carrying the band and its provenance:

band.fpr, band.tpr          # the empirical ROC on the FPR grid
band.lower, band.upper      # the simultaneous band arms
band.auc, band.auc_ci       # Mann-Whitney AUC (= sklearn) + bootstrap CI
band.confidence             # the simultaneous coverage target you asked for
band.at([0.05, 0.10])       # (lower, tpr, upper) at any FPR you care about
band.to_dataframe()         # pandas, if you have it

A typical summary():

rocci confidence band (envelope)
  samples: n_neg=212, n_pos=357
  coverage: 95% simultaneous
  AUC: 0.9974  (CI: 0.9926, 1.0000)
  band area (mean width): 0.0534
  backend: rust, n_boot=2000
  no distribution-free lower bound exists below FPR ~= 0.0125; increase the
  number of negatives to certify lower FPRs.
  please cite rocci (see CITATION.cff).

Two lines deserve a first-time explanation:

• "95% simultaneous" means the whole true curve stays inside the band in 95% of datasets. See Reading the band.
• The vacuous-region line is rocci being honest: below an FPR of about 1/n_neg-ish, no method can certify a distribution-free lower bound, so the lower band is 0 there rather than pretending otherwise.

From a fitted classifier

from rocci import from_estimator

band = from_estimator(clf, X_test, y_test, random_state=0)

Duck-typed like scikit-learn's RocCurveDisplay.from_estimator: uses predict_proba when available, else decision_function — but works with any object exposing either method, no sklearn import required.

Reproducibility

Pass random_state= to seed the bootstrap. Same seed + same backend + same rocci version ⇒ a bit-identical band, independent of thread count.

Run the scikit-learn vignette to see all of this executed on real data.