5D Library › 分销网络幂律激活、网红矩阵组合优化与激励稳定边界 · English
第三章 网红矩阵的均值方差配置
「二成头部、八成腰尾部」能把波动降多少
马科维茨的组合理论指出,把资金分散到相关性不高的资产上,可以在期望收益相近时降低波动 (Markowitz 1952)。把它用在网红投放上,常见的说法是:二成预算给头部达人,八成分给大量中小达人,获客成本(CAC)的方差可以降低六成以上。p05 用每月十万令吉的预算比较两种方案:全部给两位头部达人,或一位头部达人加三十二位中小达人。所有达人共享一个共同因子(平台算法、季节、新闻), 是任意两位达人对数获客量之间的相关系数。
推导
# Chapter 3. p05: Head creators or many micro creators? A month's budget
# of RM 100,000. Customers per RM 10,000 of spend are lognormal: head
# creators mean 25 with coefficient of variation 1.2, micro creators
# mean 28 with CV 1.5. All creators share a common factor F (platform
# algorithm, season, news): log outcome = mu + sigma (sqrt(rho) F +
# sqrt(1 - rho) e), so rho is the correlation between any two creators'
# log outcomes. Plan A: two head creators at RM 50,000 each. Plan B: one
# head creator at RM 20,000 and 32 micro creators at RM 2,500 each.
# 20,000 simulated months per rho, seed 87.
import random
from math import exp, log, sqrt
def params(mean, cv):
s2 = log(1 + cv * cv)
return log(mean) - s2 / 2, sqrt(s2)
HEAD, MICRO = params(25.0, 1.2), params(28.0, 1.5)
def draw(rng, p, f, rho):
mu, s = p
return exp(mu + s * (sqrt(rho) * f + sqrt(1 - rho) * rng.gauss(0, 1)))
def month(rng, rho, plan):
f = rng.gauss(0, 1)
if plan == "A":
return sum(5.0 * draw(rng, HEAD, f, rho) for _ in range(2))
return 2.0 * draw(rng, HEAD, f, rho) + sum(
0.25 * draw(rng, MICRO, f, rho) for _ in range(32))
def stats(xs):
m = sum(xs) / len(xs)
v = sum((x - m) ** 2 for x in xs) / (len(xs) - 1)
cac = sorted(100000.0 / max(x, 1.0) for x in xs)
return m, v, cac[len(xs) // 2], cac[int(0.9 * len(xs))]
if __name__ == "__main__":
rng = random.Random(87)
print(" rho plan customers sd variance CAC median CAC 90th")
print(" mean cut (RM) (RM)")
for rho in (0.0, 0.1, 0.3, 0.5):
out = {}
for plan in ("A", "B"):
out[plan] = stats([month(rng, rho, plan) for _ in range(20000)])
for plan in ("A", "B"):
m, v, med, p90 = out[plan]
cut = 1 - v / out["A"][1]
print(" {:3.1f} {} {:9.0f} {:5.0f} {:8.3f} {:10,.0f}"
" {:8,.0f}".format(rho, plan, m, sqrt(v), cut, med, p90))
# rho plan customers sd variance CAC median CAC 90th
# mean cut (RM) (RM)
# 0.0 A 250 209 0.000 520 1,274
# 0.0 B 274 86 0.833 386 528
# 0.1 A 249 215 0.000 530 1,354
# 0.1 B 274 126 0.657 404 683
# 0.3 A 250 235 0.000 550 1,486
# 0.3 B 277 192 0.333 440 978
# 0.5 A 250 259 0.000 564 1,632
# 0.5 B 273 253 0.048 497 1,331当 = 0 时,分散方案的方差下降 0.833,确实超过六成,而且获客量的均值略高(274 对 250),CAC 的第九十百分位从 1,274 令吉降到 528 令吉。但 = 0.1 时下降 0.657, = 0.3 时只剩 0.333, = 0.5 时几乎没有效果(0.048)。分散只能消除各自的波动,消除不了共同因子;平台算法一改,三十二位达人会一起变差。
精确的前沿与估计出来的前沿
p06 用对数正态分布的协方差公式精确计算头部份额 从 0 到 1 时的均值与标准差,再假设一家公司只有十二个月的数据,用样本方差挑选最小方差的份额。
推导
# Chapter 3. p06: The mean-variance frontier of p05, worked exactly, and
# what happens when it is estimated from a year of data. A share w of
# RM 100,000 goes to two head creators, the rest to 32 micro creators.
# For lognormal outcomes with log correlation rho, cov(X, Y) =
# m_x m_y (exp(rho s_x s_y) - 1). Then a firm that sees only 12 months
# of the two totals picks the minimum-variance w from its sample
# covariance; 2,000 such firms, rho = 0.1, seed 87.
import importlib.util
import random
from math import exp, sqrt
def load(name):
spec = importlib.util.spec_from_file_location(name, name + ".py")
m = importlib.util.module_from_spec(spec)
spec.loader.exec_module(m)
return m
p05 = load("p05_kol_portfolio")
GRID = [i / 10 for i in range(11)]
def moments(w, rho):
(mh, sh), (mm, sm) = p05.HEAD, p05.MICRO
eh, em = exp(mh + sh * sh / 2), exp(mm + sm * sm / 2)
items = [(5.0 * w, eh, sh)] * 2 + [(10.0 * (1 - w) / 32, em, sm)] * 32
mean = sum(c * e for c, e, _ in items)
var = 0.0
for i, (ci, ei, si) in enumerate(items):
for j, (cj, ej, sj) in enumerate(items):
r = 1.0 if i == j else rho
var += ci * cj * ei * ej * (exp(r * si * sj) - 1)
return mean, sqrt(var)
def totals(rng, rho):
f = rng.gauss(0, 1)
h = sum(5.0 * p05.draw(rng, p05.HEAD, f, rho) for _ in range(2))
m = sum(10.0 / 32 * p05.draw(rng, p05.MICRO, f, rho)
for _ in range(32))
return h, m
if __name__ == "__main__":
print(" head share w mean sd (rho 0.1) sd (rho 0.3)")
for w in GRID:
m1, s1 = moments(w, 0.1)
_, s3 = moments(w, 0.3)
print(" {:12.1f} {:5.0f} {:12.0f} {:13.0f}".format(
w, m1, s1, s3))
best = min(GRID, key=lambda w: moments(w, 0.1)[1])
print(" true minimum-variance share at rho 0.1: {:.1f}".format(best))
rng = random.Random(87)
picks = []
for _ in range(2000):
data = [totals(rng, 0.1) for _ in range(12)]
def sd_hat(w):
xs = [w * h + (1 - w) * m for h, m in data]
mu = sum(xs) / 12
return sum((x - mu) ** 2 for x in xs)
picks.append(min(GRID, key=sd_hat))
picks.sort()
worse = sum(1 for w in picks
if moments(w, 0.1)[1] > 1.2 * moments(best, 0.1)[1])
print(" chosen from 12 months: median {:.1f}, 10% to 90% {:.1f} to"
" {:.1f}".format(picks[1000], picks[200], picks[1800]))
print(" firms whose pick has sd 20% above the best: {:.3f}"
.format(worse / 2000))
# head share w mean sd (rho 0.1) sd (rho 0.3)
# 0.0 280 123 194
# 0.1 277 118 189
# 0.2 274 118 185
# 0.3 271 122 184
# 0.4 268 129 186
# 0.5 265 140 189
# 0.6 262 153 194
# 0.7 259 167 202
# 0.8 256 184 211
# 0.9 253 201 222
# 1.0 250 219 234
# true minimum-variance share at rho 0.1: 0.2
# chosen from 12 months: median 0.2, 10% to 90% 0.0 to 0.6
# firms whose pick has sd 20% above the best: 0.158在 = 0.1 时,真正的最小方差份额恰好是 0.2,标准差 118;全部投头部时是 219。但只凭十二个月数据挑选,中位数虽然是 0.2,十分之八的公司落在 0 至 0.6 之间,有 0.158 的公司挑出的方案比最优方案的标准差高出两成以上。十二个月的重尾数据不足以可靠地估计方差。
假说 · 尚未验证
作者的诠释(尚未验证): 在东南亚的短视频与直播电商平台上,中小达人之间的月度获客量相关系数低于 0.1,因此「二成头部、八成中小」的配置能把月度获客量的方差降低六成以上。什么会证明它错: 在至少二十四个月的投放数据中,中小达人之间月度对数获客量的平均相关系数高于 0.2,或分散配置的方差降幅低于四成。
小结
「二成头部、八成中小、方差降六成」只在达人之间的相关系数约 0.1 以下时成立;共同因子越强,分散的好处越小。最小方差份额在示意参数下恰好约为二成,但用一年的数据很难把它估准。
参考文献
Fan, Y. (2026). Affiliate Network Power Laws, Influencer Portfolio Optimization and Incentive Stability Bounds: Heavy-Tailed Productivity, Markowitz Traffic Allocation and Single-Tier Commission Equilibria [分销网络幂律激活、网红矩阵组合优化与激励稳定边界] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090229 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).