关于「夜间漏斗台账:实战工具与检查器规格、预注册检验与推翻条件」,《夜间自治漏斗》给出的定量结论是什么?
结论:第八章给出夜间漏斗台账、试算工具、按四件套分组的检查器与预注册检验:基准先验下半年累计净贡献提升中位数约 +0.082,仍有 0.093 的概率一无所获。书中参数为示意;全部计算由书中 23 段程序生成;本书是非个人化的经营分析,不构成经营、法律或财务建议,不宣称任何转化率提升的保证。
数据口径:书中参数为示意;全部计算由书中 23 段程序生成;本书是非个人化的经营分析,不构成经营、法律或财务建议,不宣称任何转化率提升的保证。
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: 0.082, 0.093
| prior | weeks | P5 | P50 | P95 | P(uplift <= 0) |
| base | 13 | +0.000 | +0.040 | +0.125 | 0.304 |
| base | 26 | +0.000 | +0.082 | +0.217 | 0.093 |
| lean | 13 | +0.000 | +0.000 | +0.030 | 0.721 |
| lean | 26 | +0.000 | +0.000 | +0.061 | 0.530 |
| base+fade | 13 | +0.000 | +0.040 | +0.082 | 0.497 |
| base+fade | 26 | +0.000 | +0.040 | +0.125 | 0.266 |
查看完整验算程序
# Chapter 8. p82: what "self-improving conversion" can honestly claim.
# One tested change per week (p72 policy "one", SE 0.032, bar 1.645).
# Prior over true effects on net contribution per visitor:
# base 0 (0.60), +4% (0.25), -4% (0.15)
# lean 0 (0.75), +3% (0.10), -3% (0.15)
# base+fade base, but half of the positive effects fade to 0 after
# 4 weeks (novelty), while harmful ones stay
# Output: quantiles of the cumulative uplift after 13 and 26 weeks and
# the probability that the loop ends at or below where it started.
import math
import random
SE, ZBAR, REPS = 0.032, 1.645, 4000
rng = random.Random(40582)
PRIORS = {"base": [(0.60, 0.0), (0.25, 0.04), (0.15, -0.04)],
"lean": [(0.75, 0.0), (0.10, 0.03), (0.15, -0.03)]}
def draw(prior):
u = rng.random()
for i in range(len(prior)):
if u < math.fsum(p for p, _ in prior[:i + 1]):
return prior[i][1]
return prior[-1][1]
def run(name, weeks):
prior = PRIORS["base" if name == "base+fade" else name]
live = []
for w in range(weeks):
f = draw(prior)
if (f + SE * rng.gauss(0, 1)) / SE > ZBAR:
fades = name == "base+fade" and f > 0 and rng.random() < 0.5
live.append((f, w + 4 if fades else None))
logs = []
for f, end in live:
if end is None or end > weeks:
logs.append(math.log1p(f))
return math.expm1(math.fsum(logs))
print("prior weeks P5 P50 P95 P(uplift <= 0)")
for name in ["base", "lean", "base+fade"]:
for weeks in (13, 26):
xs = sorted(run(name, weeks) for _ in range(REPS))
le0 = len([x for x in xs if x <= 0]) / REPS
print("{:<10} {:>6} {:>+7.3f} {:>+7.3f} {:>+7.3f} {:>11.3f}"
.format(name, weeks, xs[REPS // 20], xs[REPS // 2],
xs[19 * REPS // 20], le0))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《夜间自治漏斗微创调优、多臂赌博机流量自动倾斜与零人工转化率自举母典》第 8 章「夜间漏斗台账:实战工具与检查器规格、预注册检验与推翻条件」(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-405 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026405,
author = {Fan, Yafu},
title = {Autonomous Nightly Funnel Micro-Surgery, Bandit Traffic Allocation and Zero-Touch Conversion Self-Improvement},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}