5D Library › 慢崩盘流动性储备、实物资产期权估值与跨代财富保全动力学 · English

第五章 遍历性:平均数会骗人

Peters 的硬币

Peters 用一个简单的游戏说明「群体平均」与「时间平均」的差别 (Peters 2019):每一轮,财富以相同的机会乘以 1.5 或 0.6。

推导

# Chapter 5. p09: Peters' coin. Each round wealth is multiplied by 1.5 or
# by 0.6 with equal chance. The ensemble average grows 5% a round; the
# growth that a single player experiences over time is
# sqrt(1.5 x 0.6) - 1 per round. 10,000 players, 100 rounds, seed 88.
import random
from math import sqrt

if __name__ == "__main__":
    rng = random.Random(88)
    print("   expected gain per round (ensemble): {:+.3f}".format(
        0.5 * 1.5 + 0.5 * 0.6 - 1))
    print("   time-average growth per round:      {:+.4f}".format(
        sqrt(1.5 * 0.6) - 1))
    w = [1.0] * 10000
    print("   round   mean wealth   median wealth   share below start")
    for r in range(1, 101):
        w = [x * (1.5 if rng.random() < 0.5 else 0.6) for x in w]
        if r in (10, 25, 50, 100):
            s = sorted(w)
            print("   {:5d}   {:11.3f}   {:13.4f}   {:17.3f}".format(
                r, sum(w) / 10000, s[5000],
                sum(1 for x in w if x < 1) / 10000))
    print("   theory at round 100: mean {:.1f}, median {:.4f}".format(
        1.05 ** 100, 0.9 ** 50))
#    expected gain per round (ensemble): +0.050
#    time-average growth per round:      -0.0513
#    round   mean wealth   median wealth   share below start
#       10         1.613          0.5905               0.625
#       25         3.621          0.1695               0.654
#       50        11.763          0.0718               0.757
#      100        19.255          0.0052               0.867
#    theory at round 100: mean 131.5, median 0.0052

按群体平均计算,每一轮的期望收益是 +0.050,看起来是个好游戏;但一个人一轮一轮玩下去,他实际经历的增长率是每轮 -0.0513。一万名玩家玩到第 100 轮时,中位数财富只剩 0.0052,0.867 的人低于起点;样本平均是 19.255,远低于理论上的群体平均 131.5,因为理论平均靠的是一万人里几乎不会出现的极少数幸运者。一个家族只有一条时间路径,它关心的是时间平均,而不是群体平均。

杠杆、凯利与崩盘年

同样的道理决定了杠杆应该多高。凯利准则要求最大化财富对数的期望 (Kelly 1956)。p10 比较市场年回报服从正态分布,以及服从一个均值与标准差都相同、但每年有 0.03 机会出现崩盘年的混合分布。

推导

# Chapter 5. p10: Leverage and time-average growth. Yearly market return
# R has mean 6% and sd 18%; cash pays 2%. A fund holds L times its
# wealth in the market, so wealth grows by 1 + 0.02 + L (R - 0.02).
# Time-average growth is E[log] of that factor; a factor of zero or
# less is ruin, counted here as a loss of 99% (factor 0.01). R comes
# from a normal law, or from a crash mixture with the same mean and sd:
# with chance 0.03 a crash year (mean -40%, sd 10%), otherwise a normal
# year (mean 7.42%, sd 16.23%). 200,000 years per law, seed 88. Kelly
# under the normal law: L = 0.04 / 0.18^2.
import random
from math import log

def draws(rng, crash, n=200000):
    out = []
    for _ in range(n):
        if crash and rng.random() < 0.03:
            out.append(rng.gauss(-0.40, 0.10))
        elif crash:
            out.append(rng.gauss(0.0742, 0.1623))
        else:
            out.append(rng.gauss(0.06, 0.18))
    return out

def growth(rs, lev):
    g, ruin = 0.0, 0
    for r in rs:
        f = 1.02 + lev * (r - 0.02)
        if f <= 0:
            ruin += 1
            f = 0.01
        g += log(f)
    return g / len(rs), ruin / len(rs)

if __name__ == "__main__":
    rng = random.Random(88)
    normal, crash = draws(rng, False), draws(rng, True)
    m = sum(crash) / len(crash)
    sd = (sum((x - m) ** 2 for x in crash) / len(crash)) ** 0.5
    print("   crash mixture check: mean {:.3f}, sd {:.3f}".format(m, sd))
    print("   Kelly leverage under the normal law: {:.2f}".format(
        0.04 / 0.18 ** 2))
    print("   leverage   growth, normal   growth, crashes"
          "   ruin in 30 years")
    for lev in (0.5, 1.0, 1.5, 2.0, 2.5, 3.0):
        gn, _ = growth(normal, lev)
        gc, rc = growth(crash, lev)
        print("   {:8.1f}   {:14.4f}   {:15.4f}   {:16.3f}".format(
            lev, gn, gc, 1 - (1 - rc) ** 30))
#    crash mixture check: mean 0.060, sd 0.180
#    Kelly leverage under the normal law: 1.23
#    leverage   growth, normal   growth, crashes   ruin in 30 years
#         0.5           0.0355            0.0354              0.000
#         1.0           0.0433            0.0422              0.000
#         1.5           0.0419            0.0364              0.004
#         2.0           0.0253            0.0019              0.161
#         2.5          -0.0126           -0.0466              0.428
#         3.0          -0.0693           -0.0921              0.612

正态分布下的凯利杠杆是 1.23,杠杆 1.0 与 1.5 的时间平均增长率相近(0.0433 与 0.0419)。加入崩盘年之后,杠杆 1.5 的增长率降到 0.0364;杠杆 2.0 几乎为零(0.0019),三十年内破产的机会是 0.161;杠杆 2.5 时破产机会升到 0.428。均值与标准差都没变,变的只是尾部;以正态分布计算的「最优杠杆」,在有崩盘年的世界里已经太高。

假说 · 尚未验证

作者的诠释(尚未验证): 以几十年为期的家族资产,按正态分布计算的凯利杠杆至少应减半使用。什么会证明它错: 以包含至少两次大型崩盘的长期历史回报模拟,按正态凯利杠杆持有的组合,其时间平均增长率与破产概率都不比半凯利更差。

小结

群体平均会把一个对每个人都不利的游戏说成有利;家族只能活在一条时间路径上,应以时间平均增长率作决定。尾部的崩盘年会把以正态分布计算的最优杠杆变成危险的杠杆。

参考文献

Kelly, Jr., John L. 1956. “A New Interpretation of Information Rate.” Bell System Technical Journal 35 (4): 917 to 926.
Peters, Ole. 2019. “The Ergodicity Problem in Economics.” Nature Physics 15: 1216 to 1221.

Fan, Y. (2026). Slow-Crash Liquidity Reserves, Real-Asset Option Valuation and Intergenerational Wealth Preservation: Thirty-Month Drawdown Kinetics, Distressed Resource Acquisition and Ergodic Dynasty Trusts [慢崩盘流动性储备、实物资产期权估值与跨代财富保全动力学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090231 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).