关于「金主评分:素材更新、比例风险与名单规模」,《广告库反查与进箱物理》给出的定量结论是什么?

结论:这份名录实际抽出 143 家买家,基准比例 0.071。按分数取前 50 家,命中 34 家,精度 0.680,是基准的 9.5 倍;每百封信预计回复 5.60 封,投诉率 0.098 %。取前 100 家,精度 0.590,投诉率升到 0.112 %。取前 200 家,精度 0.410,投诉率 0.139 %。

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。

English version

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: 0.071, 0.098, 0.112, 0.139, 0.410, 0.590

directory 2,000 firms, payers 143, base rate 0.071
top khitsprecisionliftreplies/100complaint %
50340.6809.55.600.098
100590.5908.34.920.112
200820.4105.73.580.139
4001100.2753.82.560.159
8001290.1612.31.710.176
2,0001430.0711.00.870.298
platform guidance: keep complaints under 0.1 %, never 0.3 %
查看完整验算程序
# Chapter 2. p21: an ad-library score versus blasting the whole directory.
# Synthetic directory of 2,000 firms (illustrative). 6 % are "payers"
# who would buy authorised creator ads; 95 % of them advertise, their
# longest live ad is about 55 days old on average and they refresh about
# 3 creatives a month. 40 % of non-payers advertise lightly (20 days,
# 1.5 creatives). Score = ln(1 + days live) + 0.5 * creatives refreshed.
# Per-recipient reply and complaint probabilities are assumptions.
import math
import random

RNG = random.Random(20261123)
REPLY = {"payer": 0.08, "light": 0.005, "none": 0.002}
COMPLAIN = {"payer": 0.0005, "light": 0.002, "none": 0.004}


def poisson(rng, lam):
    k, p, lim = 0, 1.0, math.exp(-lam)
    while True:
        p *= rng.random()
        if p <= lim:
            return k
        k += 1


def firm(rng):
    if rng.random() < 0.06:
        if rng.random() < 0.95:
            return "payer", min(120.0, rng.expovariate(1 / 55)), \
                poisson(rng, 3.0)
        return "payer", 0.0, 0
    if rng.random() < 0.40:
        return "light", min(120.0, rng.expovariate(1 / 20)), \
            poisson(rng, 1.5)
    return "none", 0.0, 0


def score(days, refresh):
    return math.log1p(days) + 0.5 * refresh


if __name__ == "__main__":
    firms = [firm(RNG) for _ in range(2000)]
    payers = len([f for f in firms if f[0] == "payer"])
    ranked = sorted(firms, key=lambda f: (-score(f[1], f[2]), f[0]))
    print("directory 2,000 firms, payers {}, base rate {:.3f}".format(
        payers, payers / 2000))
    print("  top k   hits  precision  lift  replies/100  complaint %")
    for k in (50, 100, 200, 400, 800, 2000):
        top = ranked[:k]
        hits = len([f for f in top if f[0] == "payer"])
        rep = math.fsum(REPLY[f[0]] for f in top) / k * 100
        com = math.fsum(COMPLAIN[f[0]] for f in top) / k * 100
        print("  {:>5,}  {:>5}  {:>9.3f}  {:>4.1f}  {:>11.2f}  {:>11.3f}"
              .format(k, hits, hits / k, (hits / k) / (payers / 2000),
                      rep, com))
    print("platform guidance: keep complaints under 0.1 %, never 0.3 %")

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 2 章「金主评分:素材更新、比例风险与名单规模」(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}
}

本书其他章节结论