Integrations and Performance¶

pandas¶

Install with pip install "bootstrapx-lib[pandas]". Importing bootstrapx registers the accessor.

import numpy as np
import pandas as pd
import bootstrapx

s = pd.Series(np.random.default_rng(0).normal(size=300))
result = s.bootstrap.bca(np.mean, n_resamples=4999, random_state=42)
print(result.to_frame())

DataFrame.bootstrap.summary() evaluates columns independently. It is useful for a table of separate column estimates, but it does not estimate a difference, ratio, lift, paired effect, or p-value between columns. Use the underlying arrays with bootstrap_two_sample() instead:

from bootstrapx import bootstrap_two_sample

comparison = bootstrap_two_sample(
    frame.loc[frame["variant"] == "control", "metric"],
    frame.loc[frame["variant"] == "treatment", "metric"],
    np.mean,
    effect="difference",
    random_state=42,
)
print(comparison.to_frame())

scikit-learn¶

Install with pip install "bootstrapx-lib[sklearn]". BootstrapCV supplies bootstrap training indices and out-of-bag test indices to scikit-learn:

from bootstrapx import BootstrapCV
from sklearn.datasets import load_breast_cancer
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.model_selection import cross_val_score

X, y = load_breast_cancer(return_X_y=True)
cv = BootstrapCV(n_splits=200, random_state=42)
scores = cross_val_score(
    GradientBoostingClassifier(random_state=0),
    X,
    y,
    cv=cv,
    scoring="roc_auc",
)
print(scores.mean(), scores.std())

The score distribution is based on varying out-of-bag subsets. Its standard deviation is not automatically a confidence interval for future deployment performance.

BootstrapCV assumes independent rows and rejects non-None groups rather than silently splitting observations from the same entity across train and test. Use a group-aware cross-validator for repeated users/accounts and a time-aware splitter for temporal validation. Cluster bootstrap for an interval is a separate workflow, not a group-aware machine-learning splitter.

The approximately 63.2% unique training fraction is an expectation for large samples, not a per-split guarantee. cross_val_score(..., cv=BootstrapCV(...)) reports OOB scores; it does not automatically compute the .632 estimator.

When Numba helps¶

The core library does not require Numba. Installing pip install "bootstrapx-lib[numba]" JIT-compiles index generation for:

moving block bootstrap (mbb);
circular block bootstrap (cbb);
stationary bootstrap (stationary);
the block-index part of tapered bootstrap (tapered).

It does not accelerate IID/BCa methods, the user statistic, pandas, scikit-learn, sieve bootstrap, or wild bootstrap. Results and random-state behavior are tested with and without the optional dependency.

Use it for repeated block-bootstrap analysis or large series. Skip it when you want the smallest environment or only run IID methods. In the audited 0.4.4 Apple Silicon run, warm end-to-end np.mean calls with 500 resamples were 11–34× faster for MBB, CBB, and stationary bootstrap across n=100–10,000. The first process call took 0.007–0.121 seconds. The factor and startup cost vary with CPU, cache state, series length, method, and statistic; benchmark your own repeated workflow rather than treating these as cross-machine guarantees.

Reproduce the comparison on your machine:

pip install -e ".[numba]"
python benchmarks/bench_numba.py