关于「七十二小时的序贯边界:沃尔德检验与偷看的代价」,《现金纪律与止血红绿灯》给出的定量结论是什么?

结论:七十二小时不是一个判决时刻,而是一道硬上限。真正的判决由沃尔德边界给出:本章参数下,明显的好广告与坏广告平均 16 小时左右就能判决,比固定样本检验省下近一半;加上七十二小时上限,检验力几乎不受影响,只有落在盈亏平衡附近的测试会撞上上限,按预设关停。反过来,每小时用固定门槛偷看一次,会把毫无优势的广告以 0.325 的比例误判为赢家;坚持跑满十四天,坏广告的期望亏损是七十二小时关停的近五倍。

数据口径:北极星指标换成每日落袋净现金,即当日到账减当日真实付出再减退款准备计提;固定种子合成账本里,营收涨了 4.05 倍,落袋净现金却一路走低。 书中参数为示意,全部计算由书中程序生成。

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: 0.325

plantrue pfalse-win ratemean hour of win
one look at 72 h0.0150.00172.0
hourly, 72 h0.0150.1134.5
hourly, 14 days0.0150.1184.4
one look at 72 h0.0200.06072.0
hourly, 72 h0.0200.32513.5
hourly, 14 days0.0200.41944.9
expected net of a p = 0.015 ad run 14 days = -2,520
expected net of the same ad cut at 72 hours = -540
查看完整验算程序
# Chapter 3. p33: peeking every hour inflates false wins.
# Naive rule: every hour, test orders against break-even conversion
# 0.02 with a one-sided z-test at the 5 % level, and call the ad a
# winner the first hour z > 1.645. Compare one look at 72 hours, hourly
# looks for 72 hours, and "give it time": hourly looks for 14 days.
# True p = 0.015 (a loser) and 0.020 (break-even: no real edge). 25
# clicks an hour, CPC 1.2, contribution 60 (illustrative). 10,000
# tests per row, fixed seed; orders arrive by geometric gaps.
import math
import random

P_BE, RATE, CPC, CONTRIB = 0.02, 25, 1.2, 60.0
Z = 1.6449
TRIALS = 10000


def hourly_orders(p, hours, rng):
    counts = [0] * hours
    lq = math.log(1.0 - p)
    pos = 0
    while True:
        pos += int(math.log(1.0 - rng.random()) / lq) + 1
        h = (pos - 1) // RATE
        if h >= hours:
            return counts
        counts[h] += 1


def first_win(counts, looks):
    k = 0
    for h in range(len(counts)):
        k += counts[h]
        if h + 1 in looks:
            n = (h + 1) * RATE
            z = (k - n * P_BE) / math.sqrt(n * P_BE * (1.0 - P_BE))
            if z > Z:
                return h + 1
    return None


if __name__ == "__main__":
    rng = random.Random(20261120)
    plans = [("one look at 72 h", 72, {72}),
             ("hourly, 72 h", 72, set(range(1, 73))),
             ("hourly, 14 days", 336, set(range(1, 337)))]
    print("plan               true p  false-win rate  mean hour of win")
    for p in (0.015, 0.020):
        for name, hours, looks in plans:
            wins, when = 0, []
            for _ in range(TRIALS):
                c = hourly_orders(p, hours, rng)
                w = first_win(c, looks)
                if w is not None:
                    wins += 1
                    when.append(w)
            mh = math.fsum(when) / len(when) if when else 0.0
            print("  {:<17} {:.3f}  {:>14.3f}  {:>16.1f}".format(
                name, p, wins / TRIALS, mh))
    loss = (0.015 * CONTRIB - CPC) * 336 * RATE
    print("expected net of a p = 0.015 ad run 14 days = {:,.0f}".format(
        loss))
    print("expected net of the same ad cut at 72 hours = {:,.0f}".format(
        (0.015 * CONTRIB - CPC) * 72 * RATE))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《翻身战现金流铁血纪律、七十二小时止血红绿灯与日进五千复利重投母典》第 3 章「七十二小时的序贯边界:沃尔德检验与偷看的代价」(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-372 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026372,
  author    = {Fan, Yafu},
  title     = {Bootstrap Cashflow Discipline, Seventy-Two-Hour Kill-Switch Dashboards and Reinvestment Compounding to Five Thousand Daily},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论