关于「多条规则并行:α 怎么分,转正的规则里几成是空的」,《规则的一生》给出的定量结论是什么?

结论:第六章显示平分 α 给 100 条规则时,无效规则的判决周数只增长 1.12 倍。

数据口径:书中的转化率、流量与每周计数均为示意;全部计算由书中程序生成;本书属非个人化经营方法,不构成经营、法律或投资建议,规则的转正、降级与退役须由主事者人工放行,紧迫感只能来自真实库存与真实截止。 书中以标注框区分已发表的证据、可运行的 Python 计算(20 段程序)、原典与作者的假说。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 1.12, 100

rules malpha/rulereal alphapowerwk H0wk H1fixed wk
10.050000.033220.87145.837.718.86
50.010000.006660.87096.2911.6914.19
100.005000.003300.87476.4513.5316.46
200.002500.001630.87456.4915.3018.70
400.001250.000840.87446.5116.9620.94
1000.000500.000330.87446.5219.3123.88
growth m = 1 to 100 (ratio): H0 1.12, H1 2.51, fixed 2.70
family false-promotion bound at m = 100: 0.0326 (100 x 0.00033)
查看完整验算程序
# Chapter 6. p61: splitting alpha across many parallel rules is cheap
# Model: m rules on trial at once; to keep the chance of any false
# promotion near 0.05 the simplest split gives each SPRT alpha/m
# (Bonferroni). We compute exactly, with weekly batches of 600 visitors
# as in p41 (p0 0.03, p1 0.036, beta 0.20), the realised per-rule
# alpha, the power and the mean weeks under H0 and H1, and compare the
# one-sided fixed-sample size at the same alpha and power. Wald's A
# grows like ln(1/alpha) while B barely moves, so null rules, the
# majority of a backlog, are cut almost as fast as before.
import math
from statistics import NormalDist

P0, P1, BETA, M = 0.03, 0.036, 0.20, 600
LA = math.log(P1 / P0)
LB = math.log((1 - P1) / (1 - P0))
N_MAX = 90000


def pmf(p):
    hi = min(M, int(M * p + 14 * math.sqrt(M * p * (1 - p))) + 2)
    return [math.exp(math.lgamma(M + 1) - math.lgamma(j + 1)
                     - math.lgamma(M - j + 1) + j * math.log(p)
                     + (M - j) * math.log(1 - p)) for j in range(hi + 1)]


def batched(p, alpha):
    a = math.log((1 - BETA) / alpha)
    b = math.log(BETA / (1 - alpha))
    step = pmf(p)
    open_k, keep, surv = {0: 1.0}, [], [1.0]
    for t in range(1, N_MAX // M + 1):
        n = t * M
        bins = {}
        for k, w in open_k.items():
            for j, q in enumerate(step):
                bins.setdefault(k + j, []).append(w * q)
        open_k = {}
        for k in sorted(bins):
            mass = math.fsum(bins[k])
            llr = k * LA + (n - k) * LB
            if llr >= a:
                keep.append(mass)
            elif llr > b:
                open_k[k] = mass
        surv.append(math.fsum(open_k.values()))
    return math.fsum(keep), math.fsum(surv[:-1])


def n_fixed(alpha):
    nd = NormalDist()
    za, zb = nd.inv_cdf(1 - alpha), nd.inv_cdf(1 - BETA)
    root = math.fsum([za * math.sqrt(P0 * (1 - P0)),
                      zb * math.sqrt(P1 * (1 - P1))])
    return math.ceil((root / (P1 - P0)) ** 2)


print("rules m  alpha/rule  real alpha  power   wk H0   wk H1  fixed wk")
base = None
for m in (1, 5, 10, 20, 40, 100):
    al = 0.05 / m
    a0, w0 = batched(P0, al)
    a1, w1 = batched(P1, al)
    nf = n_fixed(al) / M
    if base is None:
        base = (w0, w1, nf)
    print("%7d  %10.5f  %10.5f  %6.4f  %6.2f  %6.2f  %8.2f"
          % (m, al, a0, a1, w0, w1, nf))
print("growth m = 1 to 100 (ratio): H0 %.2f, H1 %.2f, fixed %.2f"
      % (w0 / base[0], w1 / base[1], nf / base[2]))
print("family false-promotion bound at m = 100: %.4f (100 x %.5f)"
      % (100 * a0, a0))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《线上规则实证淘汰、序贯概率比检验接入真实转化流与新文案规则转正降级母典》第 6 章「多条规则并行:α 怎么分,转正的规则里几成是空的」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-263 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026263,
  author    = {Fan, Yafu},
  title     = {Empirical Retirement of Live Rules: Wiring the Sequential Probability Ratio Test into Real Conversion Streams to Promote, Demote and Retire Copy Rules},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论