关于「均值加三倍标准差的盲区:重尾下的千分之一分位与鲸鱼客单」,《极值尾部与退单准备金》给出的定量结论是什么?
结论:第一章显示,在形状约 0.4 的示意订单尾部上,真实千分之一分位是正态估计的 2.57 倍,万分之一分位是 5.56 倍;样本上的正态估计中位数只有真实值的 0.363,最大的 20 笔订单占去离差平方和的 0.353。立项蓝图的「大数倍」据此修正为千分之一层两到三倍、万分之一层五六倍。
数据口径:书中的订单、退单、请求量与参数均为示意;全部计算由书中程序生成;准备金与服务器冗余的讨论属于非个人化经营与财务方法,不构成理财、会计、税务或投资建议,实际金额与生产变更须由经营者人工决定。 第一章显示,在形状约 0.4 的示意订单尾部上,真实千分之一分位是正态估计的 2.57 倍,万分之一分位是 5.56 倍;样本上的正态估计中位数只有真实值的 0.363,最大的 20 笔订单占去离差平方和的 0.353。
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: 2.57, 5.56
| GPD(sigma=1, xi): true quantile / normal-style estimate | |||||
| xi | mean+3sd | q99.9 | ratio | q99.99 | ratio(z=3.719) |
| -0.20 | 2.946 | 3.744 | 1.24 | 4.208 | 1.22 |
| 0.00 | 4.000 | 6.908 | 1.69 | 9.210 | 1.95 |
| 0.10 | 4.838 | 9.953 | 2.01 | 15.119 | 2.64 |
| 0.20 | 6.091 | 14.905 | 2.39 | 26.548 | 3.66 |
| 0.30 | 8.205 | 23.144 | 2.75 | 49.496 | 5.04 |
| 0.40 | 12.847 | 37.122 | 2.82 | 97.027 | 6.25 |
| 0.45 | 19.067 | 47.527 | 2.43 | 137.991 | 5.95 |
| Burr XII order model (b=200, c=2, k=1.25), yuan: | |||||
| mean 239.6 | sd 320.3 | mean+3sd 1200.5 | |||
| p=0.9990 | true | 3163.5 | normal | 1229.4 | ratio 2.57 |
| p=0.9999 | true | 7959.6 | normal | 1430.7 | ratio 5.56 |
| P(order > mean+3sd) = 0.01095 vs normal 0.00135 (8.1 times) |
查看完整验算程序
# Chapter 1. p11: how far mean + 3 sd falls short of the 99.9% quantile.
# Model: generalised Pareto tails GPD(sigma = 1, xi) with closed-form
# mean, sd and quantiles (Pickands 1975), plus the book's illustrative
# order-value model, a Burr XII law with b = 200 yuan, c = 2, k = 1.25
# (tail index c k = 2.5, tail shape xi = 1 / (c k) = 0.4). All
# parameters are illustrative.
# The ratio q / (mean + z sd) is a dose-response curve in xi.
import math
Z3 = 3.0
Z999 = 3.090232306167813 # standard normal 0.999 quantile
Z9999 = 3.719016485455709 # standard normal 0.9999 quantile
def gpd_moments(xi, sig=1.0):
# GPD: mean sig/(1-xi), var sig^2/((1-xi)^2 (1-2xi))
mean = sig / (1.0 - xi)
var = sig * sig / ((1.0 - xi) ** 2 * (1.0 - 2.0 * xi))
return mean, math.sqrt(var)
def gpd_q(p, xi, sig=1.0):
if abs(xi) < 1e-12:
return -sig * math.log(1.0 - p)
return sig / xi * ((1.0 - p) ** (-xi) - 1.0)
B, C, K = 200.0, 2.0, 1.25 # order value, Burr XII (illustrative, yuan)
def burr_q(p):
return B * ((1.0 - p) ** (-1.0 / K) - 1.0) ** (1.0 / C)
def beta(a, b):
return math.exp(math.lgamma(a) + math.lgamma(b) - math.lgamma(a + b))
def burr_moments():
m1 = B * K * beta(K - 1.0 / C, 1.0 + 1.0 / C)
m2 = B * B * K * beta(K - 2.0 / C, 1.0 + 2.0 / C)
return m1, math.sqrt(m2 - m1 * m1)
print("GPD(sigma=1, xi): true quantile / normal-style estimate")
print(" xi mean+3sd q99.9 ratio q99.99 ratio(z=3.719)")
for xi in (-0.2, 0.0, 0.1, 0.2, 0.3, 0.4, 0.45):
m, s = gpd_moments(xi)
q3, q4 = gpd_q(0.999, xi), gpd_q(0.9999, xi)
print(" %5.2f %8.3f %7.3f %6.2f %7.3f %6.2f"
% (xi, m + Z3 * s, q3, q3 / (m + Z999 * s), q4,
q4 / (m + Z9999 * s)))
m, s = burr_moments()
print("Burr XII order model (b=200, c=2, k=1.25), yuan:")
print(" mean %.1f sd %.1f mean+3sd %.1f" % (m, s, m + Z3 * s))
for p, z in ((0.999, Z999), (0.9999, Z9999)):
print(" p=%.4f true %7.1f normal %7.1f ratio %.2f"
% (p, burr_q(p), m + z * s, burr_q(p) / (m + z * s)))
sf3 = (1.0 + ((m + Z3 * s) / B) ** C) ** (-K)
print(" P(order > mean+3sd) = %.5f vs normal 0.00135 (%.1f times)"
% (sf3, sf3 / 0.0013499))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《极值理论广义帕累托尾部、超级鲸鱼客单上限与极端退单准备金母典》第 1 章「均值加三倍标准差的盲区:重尾下的千分之一分位与鲸鱼客单」(版本 1.0.0). 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-244 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026244,
author = {Fan, Yafu},
title = {Extreme-Value Theory and Generalised Pareto Tails: Whale-Order Ceilings and Reserves for Extreme Chargebacks},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}