关于「信噪比与序贯检验:小预算只能判明显坏」,《收费站买量精算》给出的定量结论是什么?
结论:\(\rho\) 为 3.0 时只要 6 单,按目标获客成本 26.51 折合 159 令吉,每天一百令吉的预算不到两天就够。\(\rho\) 为 2.0 时要 13 单、345 令吉。\(\rho\) 为 1.5 时要 38 单、1,007 令吉,大约十天。\(\rho\) 为 1.2 时要 186 单、4,931 令吉,大约 49.3 天。\(\rho\) 为 1.1 时要 681 单、18,053 令吉,每天一百令吉要跑半年。
数据口径:按示意参数,tarotmalaysia 一个破冰买家的期望净值是破冰价的 2.68 倍,按首单回本线止损,在四十条合成素材上只拿到可得利润的 0.222。 书中参数为示意,全部计算由书中程序生成。
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: 1.1, 1.2, 1.5, 1007, 13, 159
| A. 95 % interval of the CPA after n conversions (x point) | |||
| n | CPA low | CPA high | high / low |
| 1 | 0.179 | 39.498 | 220.07 |
| 2 | 0.277 | 8.257 | 29.83 |
| 3 | 0.342 | 4.849 | 14.17 |
| 5 | 0.429 | 3.080 | 7.19 |
| 10 | 0.544 | 2.085 | 3.84 |
| 20 | 0.647 | 1.637 | 2.53 |
| 50 | 0.759 | 1.347 | 1.78 |
| 100 | 0.822 | 1.229 | 1.49 |
| B. telling CPA ratio rho from 1 (z_a + z_b = 2.4865) | |||
| rho | conversions | spend RM | days at RM 100 |
| 1.1 | 681 | 18,053 | 180.5 |
| 1.2 | 186 | 4,931 | 49.3 |
| 1.5 | 38 | 1,007 | 10.1 |
| 2.0 | 13 | 345 | 3.4 |
| 3.0 | 6 | 159 | 1.6 |
查看完整验算程序
# Chapter 3. p31: why small budgets can only call "clearly bad".
# A: exact 95 % Poisson interval for the CPA after n conversions, shown
# as multiples of the point estimate S / n (Garwood interval found by
# bisection on the Poisson CDF).
# B: conversions needed to tell a CPA ratio rho from 1 (one-sided 5 %,
# power 80 %, log-rate normal approximation: n = ((z_a + z_b)/ln rho)^2)
# and the spend and days this costs at T = 26.51 RM, RM 100 a day.
import math
from statistics import NormalDist
T, DAILY = 26.51, 100.0
def pois_cdf(k, mu):
terms = [math.exp(-mu) * mu ** j / math.factorial(j)
for j in range(k + 1)]
return math.fsum(terms)
def solve(f, lo, hi):
for _ in range(200):
mid = 0.5 * (lo + hi)
if f(mid) > 0:
lo = mid
else:
hi = mid
return 0.5 * (lo + hi)
def garwood(n, conf=0.95):
a = (1 - conf) / 2
lo = 0.0
if n > 0:
lo = solve(lambda m: a - (1 - pois_cdf(n - 1, m)), 0.0,
10.0 * n + 10)
hi = solve(lambda m: pois_cdf(n, m) - a, 0.0, 10.0 * n + 20)
return lo, hi
if __name__ == "__main__":
print("A. 95 % interval of the CPA after n conversions (x point)")
print(" n CPA low CPA high high / low")
for n in (1, 2, 3, 5, 10, 20, 50, 100):
lo, hi = garwood(n)
print(" {:>3d} {:>7.3f} {:>8.3f} {:>9.2f}".format(
n, n / hi, n / lo, hi / lo))
z = NormalDist().inv_cdf(0.95) + NormalDist().inv_cdf(0.80)
print("B. telling CPA ratio rho from 1 (z_a + z_b = {:.4f})".format(z))
print(" rho conversions spend RM days at RM 100")
for rho in (1.1, 1.2, 1.5, 2.0, 3.0):
n = math.ceil((z / math.log(rho)) ** 2)
print(" {:>4.1f} {:>12,d} {:>10,.0f} {:>12.1f}".format(
rho, n, n * T, n * T / DAILY))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《测评与命理收费站买量精算、服务端转化回传校准与日入百单扩量母典》第 3 章「信噪比与序贯检验:小预算只能判明显坏」(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-345 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026345,
author = {Fan, Yafu},
title = {Paid-Ad Scaling for Diagnostic Paywalls: Meta and TikTok Stopping Rules, Server-Side CAPI Calibration and Hundred-Order Daily Unit Economics},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}