关于「后台归因,给不了授权价」,《暗投授权与钩子归因》给出的定量结论是什么?
结论:先看 A 段。每组 1,000,000 人时,回报率标准误为 0.354,置信区间宽 138.6 个百分点。换句话说,真实回报率是零,还是翻倍,这场实验分不清。每组 100,000 人时,区间宽到 438.3;每组 10,000,000 人时,仍有 43.8。
数据口径:书中参数均为示意,全部计算由书中程序生成;本书是非个人化经营分析,不构成经营或法律建议。 第六章用二叉树为跳单期权定价,合理预估 20.82 已拦住示意模型中全部跳单。
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: 0.354, 100000, 1000000, 10000000, 138.6, 43.8
| A. ROI precision of a two-arm test (cv 10, illustrative) | ||
| n per arm | SE of ROI | 95% CI width, ROI points |
| 100,000 | 1.118 | 438.3 |
| 500,000 | 0.500 | 196.0 |
| 1,000,000 | 0.354 | 138.6 |
| 5,000,000 | 0.158 | 62.0 |
| 10,000,000 | 0.112 | 43.8 |
| B. n per arm to tell ROI gap d from zero (alpha 0.05, power 0.8) | ||
| d | n per arm | person-weeks, both arms |
| 0.25 | 15,697,760 | 31,395,520 |
| 0.50 | 3,924,440 | 7,848,880 |
| 1.00 | 981,110 | 1,962,220 |
| 2.00 | 245,278 | 490,556 |
| C. a campaign with true ROI 0.5 tested on 1,000,000 per arm | ||
| t statistic 1.41; chance its estimate shows a loss 0.0786 |
查看完整验算程序
# Chapter 1. p11: why one campaign's ROI is so hard to read (Lewis-Rao).
# Model: per-person weekly sales have mean mu and standard deviation
# sigma = cv * mu; the ad costs c per exposed person; gross margin m.
# ROI = (m * delta - c) / c, delta = incremental sales per person.
# A two-arm test with n people per arm estimates delta with standard
# error sigma * sqrt(2 / n), so ROI has standard error
# (m / c) * sigma * sqrt(2 / n). cv = 10 is the order of magnitude
# Lewis and Rao (2015) call common; mu, c and m are illustrative.
import math
from statistics import NormalDist
MU, CV, COST, MARGIN = 7.0, 10.0, 0.14, 0.5
SIGMA = CV * MU
Z = NormalDist()
def roi_se(n):
return MARGIN / COST * SIGMA * math.sqrt(2.0 / n)
def n_per_arm(d_roi, alpha=0.05, power=0.80):
za = Z.inv_cdf(1 - alpha / 2)
zb = Z.inv_cdf(power)
delta = d_roi * COST / MARGIN
return math.ceil(2 * (za + zb) ** 2 * SIGMA ** 2 / delta ** 2)
if __name__ == "__main__":
z = Z.inv_cdf(0.975)
print("A. ROI precision of a two-arm test (cv 10, illustrative)")
print(" n per arm SE of ROI 95% CI width, ROI points")
for n in (100000, 500000, 1000000, 5000000, 10000000):
se = roi_se(n)
print(" {:>12,} {:>9.3f} {:>10.1f}".format(
n, se, 2 * z * se * 100))
print("B. n per arm to tell ROI gap d from zero (alpha 0.05, power 0.8)")
print(" d n per arm person-weeks, both arms")
for d in (0.25, 0.5, 1.0, 2.0):
n = n_per_arm(d)
print(" {:>5.2f} {:>12,} {:>14,}".format(d, n, 2 * n))
se = roi_se(1000000)
t = 0.5 / se
print("C. a campaign with true ROI 0.5 tested on 1,000,000 per arm")
print(" t statistic {:.2f}; chance its estimate shows a loss {:.4f}"
.format(t, Z.cdf(-t)))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《暗投白名单授权经济学、三秒钩子生存分析与单层星探归因博弈母典》第 1 章「后台归因,给不了授权价」(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-333 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026333,
author = {Fan, Yafu},
title = {Dark-Ads Whitelist Licensing Economics, Three-Second Hook Survival Analysis and Single-Tier Scout Attribution Games},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}