关于「一位客户还值多少钱:BG/NBD、Gamma-Gamma 与留存率的假象」,《算法营销与归因》给出的定量结论是什么?
结论:BG/NBD 用四个参数描述非契约客户的购买与流失,只依赖频率与最近一次购买;它的期望公式中的 \(_2F_1\) 在长期预测时级数收敛很慢,用普法夫变换加变量代换后的欧拉积分可以稳定计算,两种算法互相核对。Gamma-Gamma 模型给出收缩后的期望客单价,但假设金额与频率独立;在常客花得少的虚构例子中,违反这一假设使收入高估 49%。逐期留存率随时间上升,通常是异质性造成的分选效应:sBG 精确重现了费德与哈迪的参数,并显示「留存率连乘」把期望存续年数低估了约三分之一。辛普森悖论是另一件事:子群都变差而合并后变好。
数据口径:一笔成交该算在哪个广告渠道头上?一个 A/B 测试什么时候可以停?一位客户还值多少钱?本书是教学与研究用的数学模型,不是营销、投资或法律建议。 第八章给出归因、实验、生命周期价值与预算四步核查表,以及一项预注册的模拟基准研究:在独立第三方冻结、已知真实增量的合成多渠道数据生成器上,比较观测沙普利归因与末次点击对真实增量份额的误差(唯一主要结局为二者平均绝对误差之差,最小关注效应 2 个百分点,90% 把握度需要 79 个合成世界、计划 90 个,判定规则写成代码)。
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: 49
| A. 946 fictional repeat buyers: refitted p 7.72, q 3.50, gamma 11.88 | |
| population mean spend 36.60 (published parameters | 35.22) |
| correlation of x with average spend | -0.075 |
| B. 104-week CLV, 10 % a year, published BG/NBD and Gamma-Gamma parameters | |
| customer (x, t_x, zbar) E[spend] E[orders, 104 wk] CLV | |
| (1, 30.0, 60.0) | 52.45 1.48 71.41 |
| (6, 38.0, 25.0) | 25.70 8.40 198.96 |
| (6, 12.0, 25.0) | 25.70 0.20 4.73 |
| C. 1,500 fictional buyers: revenue per buyer E[x m] (x repeat orders, m mean | |
| spend per order) against E[x] E[m], the product used when spend is | |
| modelled separately from frequency | |
| independent E[x m] | 106.6 E[x] E[m] 109.4 (+3%) |
| heavy buyers spend less E[x m] | 85.3 E[x] E[m] 126.7 (+49%) |
核心公式
查看完整验算程序
# Gamma-Gamma spend model (Colombo & Jiang 1999; Fader, Hardie & Lee 2005b): each
# transaction value ~ Gamma(p, nu), nu ~ Gamma(q, gamma) across customers, ASSUMED
# independent of the purchase process. A: fictional repeat buyers simulated with the
# published CDNOW estimates (p 6.25, q 3.74, gamma 15.44), refitted by maximum
# likelihood. B: expected spend E(Z | x, zbar) and a 104-week discounted CLV from the
# BG/NBD conditional expectation (FHL 2005a, eq. 10) for three fictional customers.
# C: when heavy buyers spend less per order, multiplying expected transactions by
# expected spend (the independence assumption) overstates revenue.
import math, random, statistics
# --- helpers repeated from earlier programs, so that this one runs on its own ---
def f21_series(a, b, c, z, tol=1e-12, kmax=200_000):
term, s, k = 1.0, 1.0, 0
while abs(term) > tol * abs(s) and k < kmax:
term *= (a + k) * (b + k) / ((c + k) * (k + 1)) * z; s += term; k += 1
return s, k
def nelder_mead(f, x0, it=300):
n = len(x0)
pts = [list(x0)] + [[x * (1.2 if i == j else 1) for j, x in enumerate(x0)]
for i in range(n)]
for _ in range(it):
pts.sort(key=f); c = [sum(p[i] for p in pts[:-1]) / n for i in range(n)]
w = pts[-1]; r = [c[i] + (c[i] - w[i]) for i in range(n)]
if f(r) < f(pts[0]):
e = [c[i] + 2 * (c[i] - w[i]) for i in range(n)]
pts[-1] = e if f(e) < f(r) else r
elif f(r) < f(pts[-2]):
pts[-1] = r
else:
k = [c[i] + 0.5 * (w[i] - c[i]) for i in range(n)]
if f(k) < f(w):
pts[-1] = k
else:
pts = [pts[0]] + [[pts[0][i] + 0.5 * (p[i] - pts[0][i]) for i in range(n)]
for p in pts[1:]]
return min(pts, key=f)
# --- end of repeated helpers ---
BG = (0.243, 4.414, 0.793, 2.426)
def gg_ll(th, data):
p, q, g = th
if min(th) <= 0:
return -1e18
ll = 0.0
for x, zbar in data:
ll += (math.lgamma(p * x + q) - math.lgamma(p * x) - math.lgamma(q)
+ q * math.log(g) + (p * x - 1) * math.log(zbar) + p * x * math.log(x)
- (p * x + q) * math.log(g + x * zbar))
return ll
def e_spend(x, zbar, p, q, g):
return p * (g + x * zbar) / (p * x + q - 1)
def ey(t, x, tx, T, r, al, a, b):
z = t / (al + T + t)
num = (a + b + x - 1) / (a - 1) * (1 - ((al + T) / (al + T + t)) ** (r + x)
* f21_series(r + x, b + x, a + b + x - 1, z)[0])
den = 1 + (x > 0) * a / (b + x - 1) * ((al + T) / (al + tx)) ** (r + x)
return num / den
def clv(x, tx, T, m, weeks=104, annual=0.10):
d = (1 + annual) ** (1 / 52); prev, tot = 0.0, 0.0
for w in range(1, weeks + 1):
cur = ey(w, x, tx, T, *BG); tot += (cur - prev) / d ** w; prev = cur
return tot * m, prev
def spenders(rng, n, p, q, g, dep=0.0):
out = []
for _ in range(n):
x = 1 + min(int(rng.expovariate(0.4)), 20) # repeat transactions
nu = rng.gammavariate(q, 1 / g) * math.exp(dep * (x - 3))
out.append((x, statistics.fmean(rng.gammavariate(p, 1 / nu) for _ in range(x)),
p / nu)) # (x, zbar, true mean spend)
return out
def pearson(x, y):
"""Pearson correlation (statistics.correlation needs Python 3.10+)."""
mx, my = statistics.fmean(x), statistics.fmean(y)
sxy = math.fsum((a - mx) * (b - my) for a, b in zip(x, y))
sxx = math.fsum((a - mx) ** 2 for a in x); syy = math.fsum((b - my) ** 2 for b in y)
return sxy / math.sqrt(sxx * syy)
if __name__ == "__main__":
PUB = (6.25, 3.74, 15.44)
data = spenders(random.Random(1999), 946, *PUB)
th = nelder_mead(lambda t: -gg_ll(t, [d[:2] for d in data]), [3.0, 3.0, 10.0], 400)
mean_pop = th[0] * th[2] / (th[1] - 1)
print(f"A. 946 fictional repeat buyers: refitted p {th[0]:.2f}, q {th[1]:.2f},"
f" gamma {th[2]:.2f}")
print(f" population mean spend {mean_pop:.2f}"
f" (published parameters: {PUB[0] * PUB[2] / (PUB[1] - 1):.2f})")
print(f" correlation of x with average spend: "
f"{pearson([d[0] for d in data], [d[1] for d in data]):+.3f}")
print("B. 104-week CLV, 10 % a year, published BG/NBD and Gamma-Gamma parameters")
print(" customer (x, t_x, zbar) E[spend] E[orders, 104 wk] CLV")
for x, tx, zb in ((1, 30.0, 60.0), (6, 38.0, 25.0), (6, 12.0, 25.0)):
m = e_spend(x, zb, *PUB); v, n = clv(x, tx, 39.0, m)
print(f" ({x}, {tx:4.1f}, {zb:4.1f}) {m:6.2f} {n:6.2f}"
f" {v:7.2f}")
print("C. 1,500 fictional buyers: revenue per buyer E[x m] (x repeat orders, m mean")
print(" spend per order) against E[x] E[m], the product used when spend is")
print(" modelled separately from frequency")
for dep, lab in ((0.0, "independent"), (0.25, "heavy buyers spend less")):
d2 = spenders(random.Random(7), 1500, *PUB, dep)
xm = statistics.fmean(d[0] * d[2] for d in d2)
pm = statistics.fmean(d[0] for d in d2) * statistics.fmean(d[2] for d in d2)
print(f" {lab:24s} E[x m] {xm:6.1f} E[x] E[m] {pm:6.1f}"
f" ({pm / xm - 1:+.0%})")唯一数据源与引用
范雅福(Fan, Y.). (2026). 《算法营销、多触点沙普利归因与获客漏斗的随机控制》第 5 章「一位客户还值多少钱:BG/NBD、Gamma-Gamma 与留存率的假象」(版本 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23836264
DOI: 10.5281/zenodo.23836264 · 阅读本章所在专著 · 5DL-B-048 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026048,
author = {Fan, Yafu},
title = {Algorithmic Marketing, Multi-Touch Shapley Attribution and Stochastic Control of Acquisition Funnels},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
doi = {10.5281/zenodo.23836264},
url = {https://doi.org/10.5281/zenodo.23836264}
}