关于「贝叶斯自适应调价:弹性、库存、冷却期、参考价格与止损线」,《冷启动测价与贝叶斯调价》给出的定量结论是什么?
结论:下半部分是 4000 次模拟里每周查一次的成绩单。门槛定在 0.90、要求至少 600 位访客才开始判:一个其实赚钱的臂(真实转化率 0.032)被误杀的比例是 0.013;一个其实亏钱的臂(真实转化率 0.018)六周都没被止住的比例是 0.155,被止住时平均在第 3.10 周。门槛提到 0.95,误杀降到 0.004,漏杀却升到 0.305。这个亏钱的臂若一直开着,六周烧掉的广告费比它带来的毛利多 RM 857。门槛怎样定,不是统计问题,而是生意问题:误杀一个好臂损失的是它未来的毛利,漏杀一个坏臂损失的是这 857。
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」。
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: 0.004, 0.013, 0.018, 0.032, 0.155, 0.305
| break-even conversion = 0.0250 | ||||
| P(conv < break-even) after n visitors at an observed rate | ||||
| n | rate 0.015 | rate 0.020 | rate 0.025 | rate 0.030 |
| 200 | 0.742 | 0.566 | 0.388 | 0.240 |
| 400 | 0.875 | 0.673 | 0.420 | 0.208 |
| 800 | 0.963 | 0.784 | 0.443 | 0.155 |
| 1600 | 0.996 | 0.888 | 0.460 | 0.090 |
| weekly checks over six weeks, 4000 launches per truth | ||||
| threshold | min_visitors | false_stop(0.032) | missed_stop(0.018) | mean_stop_week(0.018) |
| 0.80 | 0 | 0.075 | 0.067 | 2.14 |
| 0.80 | 600 | 0.044 | 0.073 | 2.62 |
| 0.90 | 0 | 0.022 | 0.148 | 2.75 |
| 0.90 | 600 | 0.013 | 0.155 | 3.10 |
| 0.95 | 0 | 0.008 | 0.298 | 3.19 |
| 0.95 | 600 | 0.004 | 0.305 | 3.39 |
| six weeks on the losing arm without a stop burns RM 857 of traffic beyond gross profit |
查看完整验算程序
# Chapter 6. p64: A stop-loss line on paid traffic: posterior chance that a price arm is below break-even.
# Fictional arm: RM 99, unit cost RM 31, paid traffic RM 1.70 per visitor, so break-even conversion is
# 1.70 / (99 - 31) = 0.025. Conversion has a flat Beta(1, 1) prior; after k sales in n visitors,
# P(conv < x) = P(Binomial(n + 1, x) >= k + 1), computed exactly in logs.
# Then 4,000 fictional launches per truth (seed 6464), 300 visitors a week for six weeks, weekly sales
# Poisson(300 * conv). A rule "stop when P(conv < break-even) >= t, once at least m visitors are in"
# is checked every week. Truths: 0.032 (profitable arm) and 0.018 (losing arm).
import math
import random
PRICE, COST, CPV = 99, 31, 1.70
BE = CPV / (PRICE - COST)
WEEKLY, WEEKS, RUNS = 300, 6, 4000
def p_below(k, n, x):
m = n + 1
lx, l1 = math.log(x), math.log(1.0 - x)
lg = math.lgamma
terms = [math.exp(lg(m + 1) - lg(j + 1) - lg(m - j + 1) + j * lx + (m - j) * l1)
for j in range(k + 1)]
return max(0.0, 1.0 - math.fsum(terms))
def poisson(rng, mu):
lim, k, t = math.exp(-mu), 0, rng.random()
while t > lim:
k += 1
t *= rng.random()
return k
if __name__ == "__main__":
print("break-even conversion = {:.4f}".format(BE))
print("P(conv < break-even) after n visitors at an observed rate")
print(" n rate 0.015 rate 0.020 rate 0.025 rate 0.030")
for n in (200, 400, 800, 1600):
cells = [p_below(int(n * r + 0.5), n, BE) for r in (0.015, 0.020, 0.025, 0.030)]
print(" {:5d} ".format(n) + " ".join("{:10.3f}".format(c) for c in cells))
paths = {}
for truth in (0.032, 0.018):
rng = random.Random(6464)
paths[truth] = [[poisson(rng, WEEKLY * truth) for _ in range(WEEKS)] for _ in range(RUNS)]
print("weekly checks over six weeks, {} launches per truth".format(RUNS))
print(" threshold min_visitors false_stop(0.032) missed_stop(0.018) mean_stop_week(0.018)")
for t in (0.80, 0.90, 0.95):
for m in (0, 600):
res = {}
for truth, runs in paths.items():
stops = []
for weeks in runs:
k, n, when = 0, 0, None
for w, s in enumerate(weeks, 1):
k, n = k + s, n + WEEKLY
if n >= m and p_below(k, n, BE) >= t:
when = w
break
stops.append(when)
res[truth] = stops
fs = sum(1 for x in res[0.032] if x is not None) / RUNS
hit = [x for x in res[0.018] if x is not None]
ms = 1.0 - len(hit) / RUNS
print(" {:9.2f} {:12d} {:17.3f} {:18.3f} {:21.2f}".format(t, m, fs, ms, sum(hit) / len(hit)))
burn = WEEKLY * WEEKS * (CPV - 0.018 * (PRICE - COST))
print("six weeks on the losing arm without a stop burns RM {:.0f} of traffic beyond gross profit"
.format(burn))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 6 章「贝叶斯自适应调价:弹性、库存、冷却期、参考价格与止损线」(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}
}