关于「公平博弈的守恒律:鞅、有界停时与可选停时定理」,《鞅、停时与商业试错的停损界》给出的定量结论是什么?
结论:鞅是不赚不亏的累计过程;停时是只看过去的停止规则;可选停时定理说,有界停时下鞅的期望守恒。程序 p11 在六条规则上精确展示了这一守恒,胜率从 0.300 到 0.824 不等,期望全部是零。程序 p12 显示,定理只在停时无界时失效,而失效的代价是极少数路径的亏损与平均等待同时趋于无穷,这在有限资本与有限时间里不可能兑现。那么,最常见、最有诱惑力的那条规则,也就是「亏了就加倍」,在有限资本下究竟是什么样的一笔账?这是下一章的问题。
数据口径:可选停时定理说,只要停止时刻不偷看未来且有上限,不赚不亏的事情在停下时的期望结果就等于起点;示意参数下,「第一次领先就停」的胜率是 0.824,期望却精确为零。 书中参数均为示意;全部计算由书中程序生成。
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: 0.300, 0.824
| fair walk, deadline N = 20 periods, exact enumeration | |||||||
| rule | E[S_T] | P(+) | P(-) | E[T] | sd | worst P(worst) | |
| hold | 0 | 0.412 | 0.412 | 20.00 | 4.47 | -20 | 0.0000 |
| first+1 | 0 | 0.824 | 0.160 | 6.22 | 2.49 | -20 | 0.0000 |
| +3/-3 | 0 | 0.481 | 0.481 | 8.47 | 2.91 | -3 | 0.4625 |
| +2/-8 | 0 | 0.664 | 0.281 | 10.44 | 3.23 | -8 | 0.0712 |
| +5/-2 | 0 | 0.300 | 0.657 | 8.77 | 2.96 | -2 | 0.6566 |
| trail3 | 0 | 0.322 | 0.572 | 10.47 | 3.24 | -3 | 0.2495 |
| every row: E[S_T] is exactly 0 (a Fraction, not a rounded float) |
核心公式
\[E[S_{t+1}\mid S_0,S_1,\ldots,S_t]=S_t.\]
查看完整验算程序
# Chapter 1. p11: Six stopping rules on a fair walk all keep mean zero.
# Model: a fair business trial. Each period the running result moves
# +1 or -1 unit with probability 1/2 each (a martingale). A stopping
# rule may look at the whole past but not the future; every rule here
# is also cut at the deadline N = 20, so the stopping time is bounded
# and Doob's optional stopping theorem says E[S_T] = S_0 = 0. Exact
# enumeration with fractions; the rules and N are illustrative.
from fractions import Fraction
import math
N = 20
HALF = Fraction(1, 2)
def rule_hold(s, m):
return False
def rule_first_up(s, m):
return s >= 1
def rule_3_3(s, m):
return s >= 3 or s <= -3
def rule_2_8(s, m):
return s >= 2 or s <= -8
def rule_5_2(s, m):
return s >= 5 or s <= -2
def rule_trail3(s, m):
return m - s >= 3
RULES = [("hold", rule_hold), ("first+1", rule_first_up),
("+3/-3", rule_3_3), ("+2/-8", rule_2_8),
("+5/-2", rule_5_2), ("trail3", rule_trail3)]
def run(rule):
live = {(0, 0): Fraction(1)}
ends = [] # (final value, time, probability)
for t in range(1, N + 1):
new = {}
for (s, m), pr in live.items():
for d in (1, -1):
ns = s + d
key = (ns, max(m, ns))
new[key] = new.get(key, Fraction(0)) + pr * HALF
live = {}
for key in sorted(new):
s, m = key
if rule(s, m) or t == N:
ends.append((s, t, new[key]))
else:
live[key] = new[key]
return ends
print("fair walk, deadline N = %d periods, exact enumeration" % N)
print("%-9s %6s %6s %6s %6s %6s %6s %8s" % (
"rule", "E[S_T]", "P(+)", "P(-)", "E[T]", "sd", "worst", "P(worst)"))
for name, rule in RULES:
ends = run(rule)
mean = Fraction(0)
for s, t, pr in ends:
mean += s * pr
pw = math.fsum(float(pr) for s, t, pr in ends if s > 0)
pl = math.fsum(float(pr) for s, t, pr in ends if s < 0)
et = math.fsum(t * float(pr) for s, t, pr in ends)
var = math.fsum(s * s * float(pr) for s, t, pr in ends)
worst = min(s for s, t, pr in ends)
pwo = math.fsum(float(pr) for s, t, pr in ends if s == worst)
print("%-9s %6s %6.3f %6.3f %6.2f %6.2f %6d %8.4f" % (
name, str(mean), pw, pl, et, math.sqrt(var), worst, pwo))
print("every row: E[S_T] is exactly 0 (a Fraction, not a rounded float)")唯一数据源与引用
范雅福(Fan, Y.). (2026). 《杜布鞅论可选停时定理、加倍追注谬误与商业试错见好就收停损界母典》第 1 章「公平博弈的守恒律:鞅、有界停时与可选停时定理」(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-249 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026249,
author = {Fan, Yafu},
title = {Doob's Martingales and the Optional Stopping Theorem: the Doubling Fallacy and Rigorous Stop-Loss Bounds for Business Trials},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}