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
第二章 首单激活奖励与销售竞赛
首单奖励买到的是谁
为了让新加入的分销员尽快开出第一单,许多计划会发放「首单激活奖励」。奖励的问题在于它也发给本来就会成交的人;而被奖励「推过门槛」的,恰恰是销售能力较弱、日后贡献较少的人。p03 用一个五千人的新进分销员队列计算这两种效应。
推导
# Chapter 2. p03: A first-sale activation bonus. A cohort of 5,000 new
# affiliates; each has a latent drive z ~ N(0, 1). Without a bonus the
# chance of a first sale within 30 days is logistic(-1.5 + z); a bonus of
# B ringgit adds 0.01 B to the logit. The bonus is paid to everyone who
# makes a first sale, including those who would have sold anyway
# (deadweight). An activated affiliate brings the brand a further
# 12-month margin of 120 exp(0.8 z - 0.32) ringgit, so the people a bonus
# tips over are the weaker sellers. Common random numbers, seed 87.
import random
from math import exp
def logistic(x):
return 1.0 / (1.0 + exp(-x))
def cohort(rng, n=5000):
return [(rng.gauss(0, 1), rng.random()) for _ in range(n)]
def run(people, bonus):
act = base = 0
value = base_value = 0.0
for z, u in people:
v = 120.0 * exp(0.8 * z - 0.32)
if u < logistic(-1.5 + z):
base += 1
base_value += v
if u < logistic(-1.5 + z + 0.01 * bonus):
act += 1
value += v
return act, base, value, base_value
if __name__ == "__main__":
people = cohort(random.Random(87))
print(" bonus activated extra deadweight cost per net gain")
print(" (RM) share extra (RM) (RM)")
for bonus in (0, 20, 50, 100, 200):
act, base, value, bv = run(people, bonus)
extra = act - base
cost = bonus * act
dead = base / act if bonus else 0.0
per = cost / extra if extra else 0.0
gain = value - bv - cost
print(" {:5d} {:9,d} {:5,d} {:10.3f} {:9,.0f} {:9,.0f}".format(
bonus, act, extra, dead, per, gain))
# bonus activated extra deadweight cost per net gain
# (RM) share extra (RM) (RM)
# 0 1,098 0 0.000 0 0
# 20 1,266 168 0.867 151 -281
# 50 1,522 424 0.721 179 -10,256
# 100 1,994 896 0.551 223 -65,017
# 200 3,020 1,922 0.364 314 -354,275不发奖励时有 1,098 人在三十天内开出首单。奖励 20 令吉,多激活 168 人,但领到奖励的人中 0.867 本来就会成交,每多激活一人的成本是 151 令吉;把这些人日后带来的毛利算进来,净额仍略为负数。奖励加到 200 令吉,多激活 1,922 人,每多激活一人的成本升到 314 令吉,净亏损 354,275 令吉。奖励越高,被推过门槛的人越弱,而本来就会成交的人照样领钱。
假说 · 尚未验证
作者的诠释(尚未验证): 小额(约 20 令吉)的首单激活奖励大致不赚不赔,加大奖励金额只会增加亏损。什么会证明它错: 在随机分配奖励金额的实验中,较高奖励组的十二个月毛利(扣除奖励支出后)显著高于低奖励组或无奖励组。
销售竞赛还是按量计酬
另一种常见的激励是销售竞赛:设一个奖金池,按业绩排名或业绩份额分配。Lazear 与 Rosen 证明,在能力相同、风险中性的条件下,排名竞赛可以和计件工资一样有效 (Lazear and Rosen 1981)。但分销员的能力差异极大。p04 让五十位分销员的能力服从重尾分布,用同样一万令吉的预算比较按业绩份额分配奖金池的竞赛(Tullock 竞赛 (Tullock 1980))与按量计酬。
推导
# Chapter 2. p04: A sales contest or a pay rate? 50 affiliates with
# ability a = 0.2 + Lomax(1.5) (seed 87); effort e costs e^2 / 2 and
# yields output a e. The same budget W = 10,000 is spent two ways.
# Contest: one prize pool, each affiliate's share of it equal to her share
# of all output (a Tullock contest); the equilibrium is found by
# iterating each affiliate's best response, e = W a (Y - y) / Y^2.
# Rate: r per unit of output above a floor T, with r set so that the
# payout is exactly W; an affiliate works (e = r a) only if it pays.
import random
N, W = 50, 10000.0
def abilities():
rng = random.Random(87)
return sorted(0.2 + (1 - rng.random()) ** (-1 / 1.5) - 1
for _ in range(N))
def contest(a):
e = [1.0] * N
for _ in range(3000):
y = [ai * ei for ai, ei in zip(a, e)]
tot = sum(y)
new = [W * ai * (tot - yi) / tot ** 2 for ai, yi in zip(a, y)]
e = [0.5 * x + 0.5 * n for x, n in zip(e, new)]
return e
def rate(a, floor):
def pay(r):
return sum(r * (ai * r * ai - floor) for ai in a
if (r * ai) ** 2 / 2 > r * floor)
lo, hi = 0.0, 1000.0
for _ in range(200):
mid = (lo + hi) / 2
lo, hi = (mid, hi) if pay(mid) < W else (lo, mid)
r = lo
return [r * ai if (r * ai) ** 2 / 2 > r * floor else 0.0 for ai in a]
def report(name, a, e):
y = [ai * ei for ai, ei in zip(a, e)]
low = sum(y[:N // 2]) / sum(y)
working = sum(1 for ei in e if ei > 1.0)
print(" {:<18} {:9,.0f} {:10.3f} {:7d}".format(
name, sum(y), low, working))
if __name__ == "__main__":
a = abilities()
print(" scheme output lower half working")
print(" share (e > 1)")
report("contest", a, contest(a))
report("rate, no floor", a, rate(a, 0.0))
report("rate, floor 20", a, rate(a, 20.0))
print(" ability: median {:.2f}, top {:.2f}".format(
(a[24] + a[25]) / 2, a[-1]))
# scheme output lower half working
# share (e > 1)
# contest 2,570 0.009 47
# rate, no floor 3,023 0.006 44
# rate, floor 20 3,020 0.000 8
# ability: median 0.87, top 22.45竞赛的总产出是 2,570,按量计酬是 3,023:同样的钱,计件买到的产出多出约 18%。竞赛的好处是更多人「动起来」(47 人对 44 人),但能力较低的一半在两种方案下都只贡献总产出的不到百分之一。设一个 20 的门槛、只对超出部分计酬,产出几乎不变(3,020),但只有 8 人还在努力:在重尾能力下,钱几乎都流向顶尖的少数人,而其余的人被放弃了。
小结
首单奖励的大部分发给本来就会成交的人,被它推过门槛的人日后贡献较少;在示意参数下,小额奖励大致打平,大额奖励明显亏损。在能力重尾的网络中,按量计酬比奖金池竞赛更有效率;门槛式奖金不减少产出,却让大多数人退出,这是一项值得在实验中衡量的代价。
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).