关于「预热:带反馈的速率控制」,《广告库反查与进箱物理》给出的定量结论是什么?
结论:先看干净名单。阶梯策略四十五天发出 11,324 封,进箱 4,948 封,投诉 15.4 次,最后一周进箱率只有 0.354。反馈策略只发 5,423 封,不到前者一半,进箱却有 4,964 封,略多于阶梯策略;投诉只有 3.2 次,最后一周进箱率 0.911。两者都在第 9 到第 11 天达到每天 150 封,差别在于阶梯策略不肯停,一路加到 300 封,把多出来的信全送给了匹配差的人。
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 标准库与结论吻合的打印值: 0.354, 0.911, 11324, 15.4, 150, 3.2
| list policy sent inboxed compl reach day inbox wk7 | |
| clean ladder | 11,324 4,948 15.4 1.00 9.0 0.354 |
| clean feedback | 5,423 4,964 3.2 1.00 11.0 0.911 |
| mixed ladder | 11,324 3,876 15.2 1.00 9.0 0.268 |
| mixed feedback | 3,605 2,669 5.4 0.70 11.0 0.687 |
| reach = share of runs whose volume reached the matched pool of | |
| 150 a day within 45 days, and the median day it happened |
查看完整验算程序
# Chapter 4. p41: warm-up as feedback rate control on a new subdomain.
# Illustrative model, 45 days, 100 seeded runs per row.
# matched pool 150 contacts a day; volume above it reaches poorly
# matched contacts (complaint 0.006 per inboxed message)
# matched complaint rate q: clean list 0.0006, mixed list 0.002
# inbox placement as in p33: p = 1 / (1 + (r / 0.002) ** 3), r an
# EWMA (weight 0.2) of the realised daily complaint rate
# Policy "ladder": start 20 a day, x1.3 daily up to 300, no feedback.
# Policy "feedback": start 20; every day read the Beta(1, 999) posterior
# of the last 14 days: grow x1.25 if P(q > 0.1 %) < 0.5, halve if
# P(q > 0.3 %) > 0.2, else hold; never above the matched pool.
import math
import random
import statistics
POOL, Q_IRR = 150, 0.006
def p_exceed(k, n, x):
# P(Beta(1 + k, 999 + n - k) > x) = P(Bin(n + 999, x) <= k)
big = n + 999
terms = [math.exp(math.lgamma(big + 1) - math.lgamma(j + 1)
- math.lgamma(big - j + 1) + j * math.log(x)
+ (big - j) * math.log1p(-x)) for j in range(k + 1)]
return min(1.0, math.fsum(terms))
def one_run(rng, q_rel, policy):
v, r = 20.0, 0.001
hist, sent, inboxed, comp, reach = [], 0, 0, 0, None
last = []
for day in range(45):
n = int(v)
p = 1.0 / (1.0 + (r / 0.002) ** 3)
good = min(n, POOL)
ib_g = len([1 for _ in range(good) if rng.random() < p])
ib_b = len([1 for _ in range(n - good) if rng.random() < p])
c = len([1 for _ in range(ib_g) if rng.random() < q_rel]) + \
len([1 for _ in range(ib_b) if rng.random() < Q_IRR])
sent, inboxed, comp = sent + n, inboxed + ib_g + ib_b, comp + c
r = 0.8 * r + 0.2 * (c / max(1, ib_g + ib_b))
hist.append((ib_g + ib_b, c))
if day >= 38:
last.append(p)
if reach is None and n >= POOL:
reach = day + 1
if policy == "ladder":
v = min(300.0, v * 1.3)
else:
w = hist[-14:]
kk = int(math.fsum(h[1] for h in w))
nn = int(math.fsum(h[0] for h in w))
if p_exceed(kk, nn, 0.003) > 0.2:
v = max(10.0, v / 2)
elif p_exceed(kk, nn, 0.001) < 0.5:
v = min(float(POOL), v * 1.25)
return sent, inboxed, comp, reach or 99, statistics.fmean(last)
if __name__ == "__main__":
rng = random.Random(20261126)
print(" list policy sent inboxed compl reach day inbox wk7")
for name, q in (("clean", 0.0006), ("mixed", 0.002)):
for pol in ("ladder", "feedback"):
runs = [one_run(rng, q, pol) for _ in range(100)]
m = [statistics.fmean(x[i] for x in runs) for i in (0, 1, 2, 4)]
hit = [x[3] for x in runs if x[3] < 99]
share = len(hit) / len(runs)
med = statistics.median(hit) if hit else 0
fmt = (" {:<5} {:<8} {:>6,.0f} {:>7,.0f} {:>5.1f}"
" {:>4.2f} {:>4} {:>8.3f}")
print(fmt.format(name, pol, m[0], m[1], m[2], share, med, m[3]))
print("reach = share of runs whose volume reached the matched pool of")
print("150 a day within 45 days, and the median day it happened")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 4 章「预热:带反馈的速率控制」(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}
}