关于「快的团队本来就好:混杂、随机分派与断点」,《五分钟线索与防飞单》给出的定量结论是什么?
结论:快团队本来就好,所以快团队与慢团队的比较会放大延迟的效应。在程序 p21 的已知世界里,混合比较得出 3.206 倍,同样区间的真实倍数是 1.441,团队内分层把估计拉回 1.417。随机分派的三千条线索给出 1.235 倍,区间罩住了真实值并排除了混合倍数。交班断点在重复抽样下几乎无偏,单次估计却波动明显。
数据口径:「一百倍」出自 2007 年一份未经同行评审的行业研究,比的是几率而不是概率;2011 年《哈佛商业评论》短文报告的倍数较小,而且该刊是编辑审稿的管理杂志,作者之一与研究所用软件公司有利益关联,这是本书的蓝图修正。 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: 1.417, 1.441, 3.206
| leads simulated | 20,000 in 40 teams |
| fast leads (t < | 5 min): 4,269, contact 0.5020 |
| slow leads (t > | 30 min): 5,339, contact 0.1566 |
| naive pooled multiple 5 v 30 min | 3.206 |
| true causal multiple at t | 5 v t = 30: 1.195 |
| true multiple for the same two bins | 1.441 |
| within-team (Mantel-Haenszel) estimate | 1.417 |
| share of the naive multiple that is confounding | 0.686 |
查看完整验算程序
# Chapter 2. p21: fast teams are better teams; the naive multiple lies.
# Model (illustrative, seed fixed): 40 sales teams with quality q drawn
# from N(0, 1). A team's typical first-touch delay is 12 exp(-1.1 q)
# minutes; each lead's delay is that times a lognormal draw. Contact
# probability is c(q) * (w + (1 - w) exp(-t / tau)) with c(q) = 0.75
# / (1 + exp(-1.2 q)), w = 0.5 and tau = 60 minutes: the true causal
# effect of delay is mild. The naive comparison pools all teams; the
# stratified estimate compares fast and slow leads inside each team
# (Mantel-Haenszel risk ratio) and so holds team quality fixed.
import math
import random
TEAMS = 40
LEADS_PER_TEAM = 500
W, TAU = 0.5, 60.0
def contact_p(q, t):
c = 0.75 / (1.0 + math.exp(-1.2 * q))
return c * (W + (1.0 - W) * math.exp(-t / TAU))
def simulate(rng):
rows = []
for team in range(TEAMS):
q = rng.gauss(0.0, 1.0)
typical = 12.0 * math.exp(-1.1 * q)
for _ in range(LEADS_PER_TEAM):
t = typical * math.exp(rng.gauss(0.0, 0.9))
y = 1 if rng.random() < contact_p(q, t) else 0
rows.append((team, t, y))
return rows
def rate(rows):
n = len(rows)
k = 0
for r in rows:
k += r[2]
return k / n if n else float("nan"), n
def mantel_haenszel(rows, fast, slow):
num, den = [], []
for team in range(TEAMS):
a = [r for r in rows if r[0] == team and fast(r[1])]
b = [r for r in rows if r[0] == team and slow(r[1])]
if not a or not b:
continue
ya = 0
for r in a:
ya += r[2]
yb = 0
for r in b:
yb += r[2]
n = len(a) + len(b)
num.append(ya * len(b) / n)
den.append(yb * len(a) / n)
return math.fsum(num) / math.fsum(den)
if __name__ == "__main__":
rng = random.Random(20261120)
rows = simulate(rng)
fast = lambda t: t <= 5.0
slow = lambda t: t >= 30.0
pf, nf = rate([r for r in rows if fast(r[1])])
ps, ns = rate([r for r in rows if slow(r[1])])
true_ratio = (W + (1 - W) * math.exp(-5 / TAU)) / (
W + (1 - W) * math.exp(-30 / TAU))
print("leads simulated: {:,} in {} teams".format(len(rows), TEAMS))
print("fast leads (t <= 5 min): {:,}, contact {:.4f}".format(nf, pf))
print("slow leads (t >= 30 min): {:,}, contact {:.4f}".format(ns, ps))
print("naive pooled multiple 5 v 30 min: {:.3f}".format(pf / ps))
print("true causal multiple at t = 5 v t = 30: {:.3f}".format(
true_ratio))
dec = lambda t: W + (1 - W) * math.exp(-t / TAU)
bf = math.fsum(dec(r[1]) for r in rows if fast(r[1])) / nf
bs = math.fsum(dec(r[1]) for r in rows if slow(r[1])) / ns
print("true multiple for the same two bins: {:.3f}".format(bf / bs))
print("within-team (Mantel-Haenszel) estimate: {:.3f}".format(
mantel_haenszel(rows, fast, slow)))
print("share of the naive multiple that is confounding: {:.3f}".format(
1.0 - math.log(bf / bs) / math.log(pf / ps)))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《高意向线索黄金五分钟转化、线下派单防飞单审计与转介绍佣金锁死母典》第 2 章「快的团队本来就好:混杂、随机分派与断点」(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-348 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026348,
author = {Fan, Yafu},
title = {Five-Minute High-Intent Lead Conversion, Offline Deal-Leakage Auditing and Referral Commission Lock-In},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}