SPB Git

Repository Public MIT

gen_bootstrap.py 2.6 KB 43 lines python

1#!/usr/bin/env python32# ORIX Calc3# File: Packages/ORIXStatistics/Reference/gen_bootstrap.py4# Description: Generates bootstrap-interval references: fixed replicates -> percentile / basic / BCa limits, plus scipy.stats.bootstrap end-to-end targets.5#6# Author: Simon-Pierre Boucher7# Contact: contact@spboucher.ai8import numpy as np9import scipy.stats as st10from reflib import write1112rng = np.random.default_rng(99)13cases = []14for name, data, stat in [("lognormalMean", rng.lognormal(1.0, 0.8, 30), np.mean), ("normalSD", rng.normal(5, 2, 45), lambda a, axis=-1: np.std(a, ddof=1, axis=axis)),15                         ("skewMedian", rng.exponential(3.0, 21), np.median)]:16    n = len(data); B = 200017    idx = rng.integers(0, n, size=(B, n))18    reps = stat(data[idx], axis=-1)19    theta = float(stat(data, axis=-1))20    jack = np.array([float(stat(np.delete(data, i), axis=-1)) for i in range(n)])21    for level in (0.95, 0.9):22        alpha = (1 - level) / 223        # BCa ingredients from Efron (1987), written out here; the end-to-end result is compared with scipy.stats.bootstrap below.24        z0 = st.norm.ppf(((reps < theta).sum() + (reps <= theta).sum()) / (2 * B))25        d = jack.mean() - jack26        a = (d ** 3).sum() / (6 * (d ** 2).sum() ** 1.5)27        zl, zu = st.norm.ppf(alpha), st.norm.ppf(1 - alpha)28        a1 = st.norm.cdf(z0 + (z0 + zl) / (1 - a * (z0 + zl)))29        a2 = st.norm.cdf(z0 + (z0 + zu) / (1 - a * (z0 + zu)))30        bca = [float(np.quantile(reps, float(a1))), float(np.quantile(reps, float(a2)))]31        pct = [float(np.quantile(reps, alpha)), float(np.quantile(reps, 1 - alpha))]32        basic = [2 * theta - pct[1], 2 * theta - pct[0]]33        cases.append({"name": name, "data": [float(v) for v in data], "replicates": [float(v) for v in reps], "jackknife": [float(v) for v in jack],34                      "estimate": theta, "level": level, "percentile": pct, "basic": basic, "bca": bca, "z0": float(z0), "acceleration": float(a)})35    # Loose end-to-end comparison target: scipy's own bootstrap with a large number of resamples (different random stream).36    for method in ("percentile", "basic", "BCa"):37        r = st.bootstrap((data,), stat, n_resamples=20000, confidence_level=0.95, method=method, rng=np.random.default_rng(5), vectorized=True)38        cases[-2]["scipy_" + method.lower()] = [float(r.confidence_interval.low), float(r.confidence_interval.high)]39        cases[-2]["scipy_se"] = float(r.standard_error)4041write("bootstrap", {"source": "Efron (1987) formulas on fixed numpy replicates; scipy %s scipy.stats.bootstrap for the end-to-end targets" % __import__("scipy").__version__,42                    "cases": cases})43
Files main
Menu

Theme