关于「价格就是臂:离散价格梯、ε-贪心、UCB1 与按收入抽样」,《冷启动测价与贝叶斯调价》给出的定量结论是什么?

结论:把第二章的带子切成三到五条等比间距的臂,奖励写成每访客收入,是本章一切计算的前提;在 99、129、159、209 四档的虚构样本里,转化最高的是 99 元,收入最高的是 159 元。首发六千名访客远不够做一次分辨相邻两档的固定检验,赌博机的价值在于让学习少花钱:按收入的 Thompson 抽样平均后悔值 5,949 元,等分流量 9,074 元,按转化率的 Thompson 12,739 元。UCB1 的半宽在稀有成交场景里宽得没有意义,ε-贪心有锁死的赌性。照第三章的办法造出的先验能再省一千多元,前提是它不自信;按天批量更新几乎不花钱。下一章把市场、渠道与新老客这些上下文加进来,并划出分群定价不能越过的线。

数据口径:价格展示与分群定价的法规只写一般框架,不是法律或税务意见。 书中以标注框区分已发表的证据、可运行的计算、原典与作者的假说,每条假说都写明什么结果会推翻它;原典考辨取《管子·国蓄》的轻重敛散与「准平」。

English version

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: 129, 159, 209, 99

geometric ladders inside [96, 212] yuan, snapped to prices ending in 9
K3: step ratio 1.4860 raw 96.0 142.7 212.0 snapped [99, 139, 209]
K4: step ratio 1.3022 raw 96.0 125.0 162.8 212.0 snapped [99, 129, 159, 209]
K5: step ratio 1.2190 raw 96.0 117.0 142.7 173.9 212.0 snapped [99, 119, 139, 169, 209]
K6: step ratio 1.1717 raw 96.0 112.5 131.8 154.4 180.9 212.0 snapped [99, 109, 129, 159, 179, 209]
arithmetic K4: raw 96.0 134.7 173.3 212.0 snapped [99, 139, 169, 209]; step ratios 1.403 1.287 1.223
arm conv rev/visitor profit/visitor sd(rev)/visitor
990.100 9.900 6.100 29.70
1290.088 11.352 8.008 36.55
1590.080 12.720 9.680 43.14
2090.052 10.868 8.892 46.40
best conversion arm 99, best revenue arm 159, best profit arm 159
revenue gaps to the best arm99: 2.820, 129: 1.368, 159: 0.000, 209: 1.852
profit gaps to the best arm99: 3.580, 129: 1.672, 159: 0.000, 209: 0.788
best revenue arm keeps 4/50.800 of the top conversion at 1.606 x the price
visitors per arm a fixed A/B test needs to separate the best arm from each rival (revenue)
159 vs 99: gap 2.820 yuan/visitor, n2,708 per arm
159 vs 129: gap 1.368 yuan/visitor, n13,406 per arm
159 vs 209: gap 1.852 yuan/visitor, n9,186 per arm
six weeks x 1,000 visitors: horizon 6,000; even split costs 1.510/visitor, 9,060 total
best-revenue arm earns 76,320 over the horizon; the best-conversion arm earns 59,400
查看完整验算程序
# Chapter 4. p40: Turning a Van Westendorp band into a ladder of price arms, and what each arm earns.
# Band ends are the rounded chapter 2 sample bounds (PMC about 96, PME about 212 yuan, convention A).
# Arms are spaced geometrically (equal percentage steps) and snapped to the nearest price ending in 9.
# The conversion per product-page visitor at each arm is a fictional "true" table for the fictional
# Yanchi Studio card set; the algorithms in p42-p44 never see it, they only see simulated purchases.
# Unit cost (print, box, pack) is a fictional 38 yuan.
import math
from fractions import Fraction

LOW, HIGH = 96, 212
ARMS = [99, 129, 159, 209]
CONV = [0.100, 0.088, 0.080, 0.052]
COST = 38
Z = 1.959964 + 0.841621                       # two-sided alpha 0.05, power 0.80


def snap9(x):
    return 10 * round((x + 1) / 10) - 1


print("geometric ladders inside [{}, {}] yuan, snapped to prices ending in 9".format(LOW, HIGH))
for k in (3, 4, 5, 6):
    step = (HIGH / LOW) ** (1 / (k - 1))
    raw = [LOW * step ** i for i in range(k)]
    print("   K={}: step ratio {:.4f}  raw {}  snapped {}".format(
        k, step, " ".join("{:.1f}".format(r) for r in raw), [snap9(r) for r in raw]))
arith = [LOW + (HIGH - LOW) * i / 3 for i in range(4)]
print("   arithmetic K=4: raw {} snapped {}; step ratios {}".format(
    " ".join("{:.1f}".format(a) for a in arith), [snap9(a) for a in arith],
    " ".join("{:.3f}".format(arith[i + 1] / arith[i]) for i in range(3))))

print("arm  conv   rev/visitor  profit/visitor  sd(rev)/visitor")
rev = [p * c for p, c in zip(ARMS, CONV)]
prof = [(p - COST) * c for p, c in zip(ARMS, CONV)]
sd = [p * math.sqrt(c * (1 - c)) for p, c in zip(ARMS, CONV)]
for i, p in enumerate(ARMS):
    print("{:4d}  {:.3f}  {:11.3f}  {:14.3f}  {:15.2f}".format(p, CONV[i], rev[i], prof[i], sd[i]))
bc, br, bp = CONV.index(max(CONV)), rev.index(max(rev)), prof.index(max(prof))
print("best conversion arm {}, best revenue arm {}, best profit arm {}".format(ARMS[bc], ARMS[br], ARMS[bp]))
print("revenue gaps to the best arm: " + ", ".join(
    "{}: {:.3f}".format(p, rev[br] - r) for p, r in zip(ARMS, rev)))
print("profit gaps to the best arm:  " + ", ".join(
    "{}: {:.3f}".format(p, prof[bp] - r) for p, r in zip(ARMS, prof)))
drop = Fraction(round(CONV[br] * 1000), round(CONV[bc] * 1000))
print("best revenue arm keeps {} = {:.3f} of the top conversion at {:.3f} x the price".format(
    drop, float(drop), ARMS[br] / ARMS[bc]))

print("visitors per arm a fixed A/B test needs to separate the best arm from each rival (revenue)")
for i, p in enumerate(ARMS):
    if i == br:
        continue
    gap = rev[br] - rev[i]
    n = Z * Z * (sd[br] ** 2 + sd[i] ** 2) / gap ** 2
    print("   {} vs {}: gap {:.3f} yuan/visitor, n = {:,} per arm".format(ARMS[br], p, gap, math.ceil(n)))
T = 6000
loss_even = math.fsum(rev[br] - r for r in rev) / len(ARMS)
print("six weeks x 1,000 visitors: horizon {:,}; even split costs {:.3f}/visitor, {:,.0f} total"
      .format(T, loss_even, T * loss_even))
print("best-revenue arm earns {:,.0f} over the horizon; the best-conversion arm earns {:,.0f}".format(
    T * rev[br], T * rev[bc]))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《零历史数据冷启动测价、范氏价格敏感度边界与贝叶斯自适应调价母典》第 4 章「价格就是臂:离散价格梯、ε-贪心、UCB1 与按收入抽样」(版本 0.1.0). 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-133 · v0.1.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026133,
  author    = {Fan, Yafu},
  title     = {Zero-History Cold-Start Price Testing, Van Westendorp Sensitivity Bounds and Bayesian Adaptive Repricing},
  year      = {2026},
  publisher = {5D Library},
  edition   = {0.1.0},
  note      = {DOI pending}
}

本书其他章节结论