5D Library › 分拆经济学、防套利身份锁定与多阶授权阶梯 · English
第六章 菜单的稳健性:需求设错、加购定价与大客户风险
需求设错时的遗憾
第五章的菜单是对一组参数的精确最优解。可是卖家并不知道真实参数。程序生成 300 个「另一种世界」:每一类买家的估值乘上一个随机因子,每一类的占比也乘上一个随机因子再归一化(因子按类型排序,以保持递增差分)。在每个世界中,比较第五章菜单的利润与该世界自己的最优利润。
推导
# Chapter 6. [p11] Regret of the p09 menu when the seller's demand model
# is wrong. 300 alternative worlds (seed 68): each type's values are
# scaled by a factor exp(N(0, 0.3)) and each type's share by
# exp(N(0, 0.4)) (renormalised). Increasing differences survive a
# per-type scaling only if the factors rise with type, so the factors
# are sorted upward. In each world we compare the profit of the fixed
# p09 prices with that world's own optimum (exact search as in p09),
# and do the same for the p09 prices scaled down by 0.85 and 0.7.
import math
import random
from itertools import combinations_with_replacement as cwr
SHARE = (0.60, 0.22, 0.10, 0.05, 0.025, 0.005)
V = ((12, 30, 40, 45, 50, 55),
(15, 120, 180, 210, 240, 260),
(18, 160, 330, 420, 520, 560),
(20, 180, 420, 700, 1100, 1300),
(20, 200, 500, 900, 12870, 16000),
(20, 200, 520, 1000, 16000, 38000))
COST = (1, 3, 8, 12, 1500, 6000)
P09 = {1: 12, 2: 117, 3: 287, 4: 567, 5: 12537, 6: 34537}
def value(v, share, price):
tot = 0.0
for t in range(6):
opts = [(round(v[t][j - 1] - price[j], 9), j) for j in price]
opts = [o for o in opts if o[0] >= 0]
if opts:
j = max(opts)[1]
tot += share[t] * (price[j] - COST[j - 1])
return tot
def optimum(v, share):
best = 0.0
for assign in cwr(range(7), 6):
price, prev = {}, None
for t, j in enumerate(assign):
if j == 0:
continue
if prev is None:
price[j] = v[t][j - 1]
elif j not in price:
price[j] = price[prev] + v[t][j - 1] - v[t][prev - 1]
prev = j
best = max(best, value(v, share, price))
return best
if __name__ == "__main__":
rng = random.Random(68)
ratios, scaled = [], {0.85: [], 0.7: []}
for _ in range(300):
f = sorted(math.exp(rng.gauss(0, 0.3)) for _ in range(6))
v = [[x * f[t] for x in V[t]] for t in range(6)]
s = [SHARE[t] * math.exp(rng.gauss(0, 0.4)) for t in range(6)]
tot = sum(s)
s = [x / tot for x in s]
best = optimum(v, s)
ratios.append(value(v, s, P09) / best)
for k in scaled:
pk = {j: k * p for j, p in P09.items()}
scaled[k].append(value(v, s, pk) / best)
ratios.sort()
q = lambda p: ratios[int(p * (len(ratios) - 1))]
print(" p09 prices / world optimum, 300 worlds:")
print(" min {:.3f} 5% {:.3f} median {:.3f} mean {:.3f}".format(
ratios[0], q(0.05), q(0.5), sum(ratios) / len(ratios)))
print(" worlds where p09 earns under 0.5 of the optimum: {:.3f}"
.format(sum(r < 0.5 for r in ratios) / len(ratios)))
for k, r in scaled.items():
r.sort()
print(" prices x {:.2f}: 5% {:.3f} median {:.3f} mean {:.3f}"
.format(k, r[int(0.05 * 299)], r[150], sum(r) / len(r)))
print("\nReading: a ladder fitted to one demand model is fragile at")
print("the top. When studio values fall a little, the licence price")
print("is above them and the top tiers sell nothing. Pricing 15%")
print("below the point estimate lifts the worst worlds but lowers")
print("the median: the insurance has a premium.")
# p09 prices / world optimum, 300 worlds:
# min 0.137 5% 0.315 median 0.720 mean 0.687
# worlds where p09 earns under 0.5 of the optimum: 0.103
# prices x 0.85: 5% 0.493 median 0.645 mean 0.643
# prices x 0.70: 5% 0.404 median 0.530 mean 0.536
#
# Reading: a ladder fitted to one demand model is fragile at
# the top. When studio values fall a little, the licence price
# is above them and the top tiers sell nothing. Pricing 15%
# below the point estimate lifts the worst worlds but lowers
# the median: the insurance has a premium.在一半的世界中,第五章菜单能拿到最优利润的 0.720 以上;但在最差的 5% 的世界中只拿到 0.315 以下,0.103 的世界不到一半。失手几乎都发生在顶端:工作室的估值只要稍微下降,牌照价格就高于他们的估值,最上面两级什么都卖不出去。把全部价格压低 15%(乘以 0.85),最差 5% 的世界提高到 0.493,但中位数降到 0.645:稳健是要付保费的。
跨板块加购该不该打折
第五章的第四级是「跨板块加购」。直觉上,已经买了一个板块的人应当享受加购折扣。程序用高斯 Copula 让 C 与 A 两个板块的估值具有相关系数 ,比较 A 的单独最优价,与只向已买 C 的人出售 A 时的最优加购价 (Nelsen 2006)。
推导
# Chapter 6. [p12] The cross-module add-on (tier 4): what should a card
# holder of module C pay for module A? Values of C and A are uniform
# on 0..100 with a Gaussian-copula correlation rho (seed 68, 20,000
# buyers). The stand-alone optimal price of A is compared with the
# optimal add-on price of A offered only to buyers who bought C at its
# own optimal price.
import random
from statistics import NormalDist
N = 20000
PHI = NormalDist().cdf
def pairs(rho, rng):
out = []
for _ in range(N):
z1 = rng.gauss(0, 1)
z2 = rho * z1 + (1 - rho * rho) ** 0.5 * rng.gauss(0, 1)
out.append((100 * PHI(z1), 100 * PHI(z2)))
return out
def best(values):
s = sorted(values)
n = len(s)
return max((s[i] * (n - i), s[i]) for i in range(n))
if __name__ == "__main__":
print(" rho price C stand-alone A add-on A to C-holders"
" add-on/alone")
for rho in (-0.6, -0.3, 0.0, 0.3, 0.6, 0.9):
pop = pairs(rho, random.Random(68))
rc, pc = best([c for c, a in pop])
ra, pa = best([a for c, a in pop])
holders = [a for c, a in pop if c >= pc]
rh, ph = best(holders)
print(" {:5.1f} {:8.1f} {:14.1f} {:22.1f} {:13.3f}".format(
rho, pc, pa, ph, ph / pa))
print("\nReading: an add-on discount is right only when module values")
print("are negatively correlated. With independent or positively")
print("correlated values the add-on should cost the full price or")
print("more: a holder is a better-than-average buyer of the next one.")
# rho price C stand-alone A add-on A to C-holders add-on/alone
# -0.6 49.3 50.1 35.2 0.703
# -0.3 49.3 49.3 45.6 0.925
# 0.0 49.3 49.5 49.6 1.003
# 0.3 49.3 50.1 53.0 1.058
# 0.6 49.3 51.0 57.9 1.135
# 0.9 49.3 49.0 55.9 1.140
#
# Reading: an add-on discount is right only when module values
# are negatively correlated. With independent or positively
# correlated values the add-on should cost the full price or
# more: a holder is a better-than-average buyer of the next one.只有当两个板块的估值负相关时,加购才应打折: 时,最优加购价是单独价的 0.703。估值独立时两者几乎相等(1.003);正相关时,加购价应当高于单独价, 时为 1.135。原因是选择效应:买了 C 的人,在正相关之下,恰好是 A 的高估值买家。所以「已购用户加购八折」只在板块彼此替代时才对;在板块彼此加强时,它是把钱送回给最愿意付钱的人。
大客户集中带来的波动
利润集中在少数 B2B 买家身上,还有一个代价:波动。设每年 4,000 位来访者,每一类的购买人数服从泊松分布,模拟 20,000 个年份与季度。
推导
# Chapter 6. [p13] Whale risk: the ladder's profit is concentrated in a
# few B2B buyers, so a year's profit is volatile. Visitors per year
# 4,000; each type's buyer count is Poisson with mean 4,000 x share
# (p09 shares); each buys the tier and margin of the p09 menu. 20,000
# simulated years and quarters (seed 68). Reported: plan, coefficient
# of variation, chance of missing 80% of plan, and the Herfindahl index
# of profit by tier.
import math
import random
SHARE = (0.60, 0.22, 0.10, 0.05, 0.025, 0.005)
MARGIN = (12 - 1, 117 - 3, 287 - 8, 567 - 12, 12537 - 1500,
34537 - 6000)
def poisson(lam, rng):
if lam > 60:
return max(0, int(round(rng.gauss(lam, math.sqrt(lam)))))
l, k, p = math.exp(-lam), 0, 1.0
while True:
p *= rng.random()
if p < l:
return k
k += 1
def simulate(visitors, runs, rng):
plan = sum(visitors * s * m for s, m in zip(SHARE, MARGIN))
out = []
for _ in range(runs):
out.append(sum(poisson(visitors * s, rng) * m
for s, m in zip(SHARE, MARGIN)))
mean = sum(out) / runs
sd = math.sqrt(sum((x - mean) ** 2 for x in out) / (runs - 1))
miss = sum(x < 0.8 * plan for x in out) / runs
return plan, sd / mean, miss
if __name__ == "__main__":
rng = random.Random(68)
contrib = [s * m for s, m in zip(SHARE, MARGIN)]
tot = sum(contrib)
hhi = sum((c / tot) ** 2 for c in contrib)
print(" profit share by tier: " + " ".join(
"{:.3f}".format(c / tot) for c in contrib))
print(" Herfindahl index {:.3f} (equal shares would give {:.3f})"
.format(hhi, 1 / 6))
for label, vis in (("year ", 4000), ("quarter", 1000)):
plan, cv, miss = simulate(vis, 20000, rng)
print(" {}: plan {:12,.0f} CV {:.3f} P(< 80% of plan) {:.3f}"
.format(label, plan, cv, miss))
print("\nReading: two tiers bought by 3% of buyers earn most of the")
print("profit, so a quarter misses its plan by a fifth far more")
print("often than the visitor count alone would suggest.")
# profit share by tier: 0.013 0.050 0.055 0.055 0.545 0.282
# Herfindahl index 0.386 (equal shares would give 0.167)
# year : plan 2,023,760 CV 0.083 P(< 80% of plan) 0.006
# quarter: plan 505,940 CV 0.168 P(< 80% of plan) 0.112
#
# Reading: two tiers bought by 3% of buyers earn most of the
# profit, so a quarter misses its plan by a fifth far more
# often than the visitor count alone would suggest.各级利润占比的 Herfindahl 指数为 0.386,而六级平均分配时为 0.167。按年计,利润的变异系数是 0.083,低于计划八成的概率是 0.006;按季度计,变异系数升到 0.168,低于计划八成的概率是 0.112。来访人数相同,波动却几乎全部来自那 3% 的 B2B 买家。
假说 · 尚未验证
作者的诠释(尚未验证): 一家以牌照为主要利润来源的测算公司,应当按季度准备至少一个季度的现金缓冲,并把消费者层级当作稳定现金流的来源,而不是利润来源。什么会证明这一判断错:若真实的 B2B 销售在季度之间远比泊松分布平稳(例如签约有固定的年度节奏),本章高估了季度波动。
小结
第五章的最优菜单在一半的世界中能拿到最优利润的七成以上,但在最差的世界中会跌到三成:失手发生在顶端。加购折扣只在板块估值负相关时才对,正相关时加购价应当高于单独价。利润集中在 3% 的买家身上,使季度利润低于计划八成的概率达到 0.112。
参考文献
Fan, Y. (2026). Unbundling Economics, Anti-Arbitrage Identity Locking and Multi-Tier Licensing Ladders: Orthogonal Demand Partitioning, Resale Cannibalization Bounds, Affiliate Margin Preservation and Five-Figure B2B Upsell Equilibria [分拆经济学、防套利身份锁定与多阶授权阶梯] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090184 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).