关于「在投时长:广告库里的生存曲线」,《广告库反查与进箱物理》给出的定量结论是什么?
结论:第一、二章用生存分析与比例风险模型给品牌打分,并说明撒网到整份名录时投诉率升到 0.298 %。
数据口径:公开广告库里仍在投的广告是右删失数据,须用卡普兰与迈耶的乘积极限估计去读;「仍在投且已投满 30 天」这条规则在合成广告库里把持续投放者的比例从 0.313 抬到 0.902。 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
| A. 600 ads, 129 still running on snapshot day (0.215) | |
| day KM S(t) naive S(t) true S(t) | |
| 7 | 0.511 0.478 0.510 |
| 14 | 0.403 0.352 0.405 |
| 30 | 0.269 0.198 0.267 |
| 60 | 0.165 0.082 0.167 |
| 90 | 0.104 0.025 0.098 |
| median days: KM 7.5 naive 6.3 true 7.5 | |
| B. by advertiser type (KM) | |
| steady n | 188 running= 96 S(14)=0.796 S(30)=0.664 |
| trial n | 412 running= 33 S(14)=0.223 S(30)=0.089 |
| C. running and > | 30 days old: 61 ads, 55 steady (0.902) |
| base rate of steady among all ads | 0.313 |
| log check: ln S_KM(30) | -1.3123 |
查看完整验算程序
# Chapter 1. p11: Kaplan-Meier survival of ads seen in a public ad library.
# Model (synthetic, illustrative): 600 ads start uniformly over the last
# 120 days. 30 % come from "steady" advertisers whose ads run a Weibull
# time (shape 0.9, scale 80 days); 70 % from "trial" advertisers (shape
# 0.6, scale 7 days). An ad still running on the snapshot day is
# right-censored at its age. We compare the Kaplan-Meier estimate with
# the naive reading that treats every ad as if it had already stopped.
import math
import random
RNG = random.Random(20261120)
N, WINDOW = 600, 120.0
STEADY_SHARE = 0.30
def draw_ads(rng):
ads = []
for _ in range(N):
steady = rng.random() < STEADY_SHARE
shape, scale = (0.9, 80.0) if steady else (0.6, 7.0)
dur = rng.weibullvariate(scale, shape)
start = rng.uniform(0.0, WINDOW)
age = WINDOW - start
if dur > age:
ads.append((age, 0, steady, dur)) # censored
else:
ads.append((dur, 1, steady, dur)) # stopped
return ads
def kaplan_meier(rows):
rows = sorted(rows, key=lambda r: (r[0], -r[1]))
at_risk, s, curve = len(rows), 1.0, []
i = 0
while i < len(rows):
t = rows[i][0]
d = c = 0
while i < len(rows) and rows[i][0] == t:
if rows[i][1]:
d += 1
else:
c += 1
i += 1
if d:
s *= 1.0 - d / at_risk
curve.append((t, s))
at_risk -= d + c
return curve
def surv_at(curve, t):
s = 1.0
for ti, si in curve:
if ti <= t:
s = si
else:
break
return s
def km_median(curve):
for t, s in curve:
if s <= 0.5:
return t
return float("nan")
if __name__ == "__main__":
ads = draw_ads(RNG)
cens = len([a for a in ads if a[1] == 0])
print("A. {} ads, {} still running on snapshot day ({:.3f})".format(
N, cens, cens / N))
km = kaplan_meier([(a[0], a[1]) for a in ads])
naive = kaplan_meier([(a[0], 1) for a in ads])
true = kaplan_meier([(a[3], 1) for a in ads])
print(" day KM S(t) naive S(t) true S(t)")
for t in (7, 14, 30, 60, 90):
print(" {:>3} {:>7.3f} {:>10.3f} {:>9.3f}".format(
t, surv_at(km, t), surv_at(naive, t), surv_at(true, t)))
print(" median days: KM {:.1f} naive {:.1f} true {:.1f}".format(
km_median(km), km_median(naive), km_median(true)))
print("B. by advertiser type (KM)")
for flag, name in ((True, "steady"), (False, "trial")):
sub = [(a[0], a[1]) for a in ads if a[2] == flag]
c = kaplan_meier(sub)
run = len([1 for a in ads if a[2] == flag and a[1] == 0])
print(" {:<6} n={:>3} running={:>3} S(14)={:.3f} S(30)={:.3f}"
.format(name, len(sub), run, surv_at(c, 14), surv_at(c, 30)))
alive = [a for a in ads if a[1] == 0 and a[0] >= 30]
st = len([a for a in alive if a[2]])
print("C. running and >= 30 days old: {} ads, {} steady ({:.3f})".format(
len(alive), st, st / len(alive)))
print(" base rate of steady among all ads: {:.3f}".format(
len([a for a in ads if a[2]]) / N))
print(" log check: ln S_KM(30) = {:.4f}".format(
math.log(surv_at(km, 30))))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《公开广告库金主反查、B2B冷触达进箱物理与高客单暗投签单母典》第 1 章「在投时长:广告库里的生存曲线」(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-347 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026347,
author = {Fan, Yafu},
title = {Active Ad-Library Prospecting, B2B Outbound Deliverability Physics and High-Ticket Whitelist Deal Closing},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}