关于「钩子赛马:分批老虎机、多重比较与诚实停止」,《录屏钩子赛马与收费站导流》给出的定量结论是什么?

结论:第四章把钩子当作分批老虎机来赛马,汤普森抽样把损失的播放从 6,883 次降到 952 次;十二个候选同台时朴素规则每场至少一个假赢家的概率为 0.252,错误发现率控制把赢家中的假赢家比例压到 0.066;偷看把假阳性率推到 0.270,序贯检验压回 0.010;每臂最低样本 2,311 次。

数据口径:两种打法各跑一年的合成实验里,对口的稳定案例录屏日均付款人数是赌爆款的 16.51 倍。 这是目标,不是承诺:在示意基线下,它需要每天 280,112 次播放。

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: 0.066, 0.252, 2311

A. 1,000 tournaments12 hooks, 3 real lifts (0.60 -> 0.64), n = 2,000 each
rule false wins true wins false-discovery share any false win
p < 0.050.423 2.271 0.122 0.252
BH q 0.100.264 1.807 0.066 0.132
B. Single-test power for a 0.04 lift at n2,000: 0.741
plays per arm for power 0.82,311
查看完整验算程序
# Chapter 4. p42: twelve hooks against a control, with and without false-discovery control.
# Each hook and the control get n = 2,000 plays; the outcome is staying past second 3.
# Control rate 0.60. Three hooks carry a real lift to 0.64; nine are exactly null.
# Two-sided two-proportion z-tests. Compare "every p < 0.05 is a winner" with the
# Benjamini-Hochberg procedure at q = 0.10. 1,000 synthetic tournaments, fixed seed.
import math
import random

N, P0 = 2000, 0.60
TRUE = [0.64, 0.64, 0.64] + [0.60] * 9


def norm_sf(z):
    return 0.5 * math.erfc(z / math.sqrt(2.0))


def count(rng, n, p):
    x = int(round(rng.gauss(n * p, math.sqrt(n * p * (1 - p)))))
    return min(n, max(0, x))


def pval(k1, k0, n):
    p1, p0 = k1 / n, k0 / n
    pp = (k1 + k0) / (2.0 * n)
    se = math.sqrt(pp * (1 - pp) * 2.0 / n)
    return 2.0 * norm_sf(abs(p1 - p0) / se)


def bh(ps, q):
    m = len(ps)
    order = sorted(range(m), key=lambda i: ps[i])
    kmax = 0
    for rank, i in enumerate(order, start=1):
        if ps[i] <= q * rank / m:
            kmax = rank
    return set(order[:kmax])


if __name__ == "__main__":
    rng = random.Random(20261120)
    stats = {"naive": [], "bh": []}
    for _ in range(1000):
        k0 = count(rng, N, P0)
        ps = [pval(count(rng, N, p), k0, N) for p in TRUE]
        naive = {i for i, p in enumerate(ps) if p < 0.05}
        for name, rej in (("naive", naive), ("bh", bh(ps, 0.10))):
            false = len([i for i in rej if TRUE[i] == P0])
            true = len(rej) - false
            stats[name].append((false, true, false / len(rej) if rej else 0.0, false > 0))
    print("A. 1,000 tournaments: 12 hooks, 3 real lifts (0.60 -> 0.64), n = 2,000 each")
    print("  rule        false wins   true wins   false-discovery share   any false win")
    for name in ("naive", "bh"):
        s = stats[name]
        f = math.fsum([x[0] for x in s]) / len(s)
        t = math.fsum([x[1] for x in s]) / len(s)
        fdr = math.fsum([x[2] for x in s]) / len(s)
        fw = len([1 for x in s if x[3]]) / len(s)
        print("  {:<10} {:>10.3f} {:>11.3f} {:>18.3f} {:>16.3f}".format(
            "p < 0.05" if name == "naive" else "BH q 0.10", f, t, fdr, fw))
    print()
    z = (0.64 - 0.60) / math.sqrt(0.62 * 0.38 * 2.0 / N)
    power = norm_sf(1.959964 - z) + norm_sf(1.959964 + z)
    print("B. Single-test power for a 0.04 lift at n = 2,000: {:.3f}".format(power))
    za, zb = 1.959964, 0.841621
    n80 = 2.0 * (za + zb) ** 2 * 0.62 * 0.38 / 0.04 ** 2
    print("  plays per arm for power 0.8: {:,.0f}".format(n80))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《零露脸案例录屏短视频量产、前三秒痛点钩子赛马与私域收费站导流母典》第 4 章「钩子赛马:分批老虎机、多重比较与诚实停止」(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-350 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026350,
  author    = {Fan, Yafu},
  title     = {Faceless Case-Screen Short-Video Production, Three-Second Hook Tournaments and Direct-to-Paywall Traffic Funnels},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论