5D Library › Regime-Switching Fractional Kelly, Maximum-Drawdown Constraints and Lifecycle Leverage Allocation · 中文

Chapter 2. Halving and Worse: First Passage of a Kelly Bettor

How often does Kelly wealth halve

A well-known property of the Kelly bettor is that it does not avoid deep falls; it only avoids ruin. In the continuous log-normal model the chance that wealth ever falls to a fraction xx of its starting value, betting cc times Kelly, is x2/c−1x^{2/c – 1} for c<2c < 2, and 1 for c≥2c \ge 2 (Thorp 2006; MacLean et al. 2011). Within a finite horizon the reflection principle gives the exact chance. p03 uses the Sharpe ratio 0.3125 of Chapter 1.

Calculation

# Chapter 2. p03: First passage of a Kelly bettor's wealth to a fraction
# x of its starting value, in continuous time (log-normal model, the
# Sharpe ratio S = 0.3125 of p01). Betting c times Kelly, log wealth has
# drift S^2 (c - c^2 / 2) and volatility c S. Ever reaching x has
# chance x^(2 / c - 1) while c < 2, and 1 when c >= 2. Within T years
# the reflection formula gives the chance exactly.
from math import exp, log, sqrt
from statistics import NormalDist

S = 0.05 / 0.16
PHI = NormalDist().cdf

def ever(x, c):
    return 1.0 if c >= 2 else x ** (2 / c - 1)

def within(x, c, t):
    nu, s, a = S * S * (c - c * c / 2), c * S, -log(x)
    st = s * sqrt(t)
    return PHI((-a - nu * t) / st) + exp(-2 * nu * a / (s * s)) * PHI(
        (-a + nu * t) / st)

if __name__ == "__main__":
    print("   multiple   ever: half  ever: 20%   10 yrs: half  10 yrs: 20%")
    for c in (0.25, 0.5, 0.75, 1.0, 1.5, 2.0, 3.0, 4.0):
        print("   {:8.2f}   {:10.3f}  {:9.3f}   {:12.3f}  {:11.3f}".format(
            c, ever(0.5, c), ever(0.2, c), within(0.5, c, 10),
            within(0.2, c, 10)))
#    multiple   ever: half  ever: 20%   10 yrs: half  10 yrs: 20%
#        0.25        0.008      0.000          0.000        0.000
#        0.50        0.125      0.008          0.048        0.000
#        0.75        0.315      0.068          0.178        0.007
#        1.00        0.500      0.200          0.325        0.043
#        1.50        0.794      0.585          0.565        0.209
#        2.00        1.000      1.000          0.726        0.415
#        3.00        1.000      1.000          0.897        0.737
#        4.00        1.000      1.000          0.965        0.902

At full Kelly wealth halves at some point with chance 0.500 and falls to a fifth with chance 0.200; within ten years the chances are 0.325 and 0.043. At half Kelly the lifetime chance of halving is only 0.125, and at a quarter 0.008. From twice Kelly upwards wealth eventually reaches every low level with certainty; within ten years twice Kelly halves with chance 0.726 and three times Kelly with 0.897.

So the claim that “full Kelly on an optimistic drift loses half or more almost surely” needs a qualification: it is true when the effective multiple cc is 2 or more, which is exactly what an estimate taken during a long bull run tends to produce (Chapter 1); at the true Kelly the chance of ever halving is one half, not one.

Ten years in a market that changes regime

The log-normal model has no bad years built in. p04 runs the two-regime market of Chapter 1 for ten years from a bull month, with leverages of half and full whole-history Kelly, twice Kelly, and 7.10, which is the full Kelly an investor computes from bull months alone.

Calculation

# Chapter 2. p04: Ten years in the two-regime market of p02, starting
# in a bull month, rebalanced monthly; cash and borrowing earn 0.02 a
# year. Leverage 0.73 (half the whole-history Kelly 1.46), 1.46, 2.92,
# and 7.10, the full Kelly an investor computes from bull months only
# (0.12 / 0.13^2). Maximum drawdown is measured from the running peak;
# wealth at or below zero is ruin. 10,000 paths, seed 98.
import random
from math import sqrt

BULL, BEAR = (0.12 / 12, 0.13 / sqrt(12)), (-0.25 / 12, 0.28 / sqrt(12))
STAY, RF = (0.98, 0.92), 0.02 / 12
LEVS = (0.73, 1.46, 2.92, 7.10)

def run(rng, months=120):
    s = 0
    w = [1.0] * len(LEVS)
    peak = [1.0] * len(LEVS)
    dd = [0.0] * len(LEVS)
    for _ in range(months):
        if rng.random() > STAY[s]:
            s = 1 - s
        m, sd = BULL if s == 0 else BEAR
        x = rng.gauss(m, sd)
        for k, lev in enumerate(LEVS):
            if w[k] > 0:
                w[k] = max(w[k] * (1 + RF + lev * x), 0.0)
                peak[k] = max(peak[k], w[k])
                dd[k] = max(dd[k], 1 - w[k] / peak[k])
    return w, dd

if __name__ == "__main__":
    rng = random.Random(98)
    res = [run(rng) for _ in range(10000)]
    print("   leverage  DD>=50%  DD>=80%   ruin  median wealth  10th pct")
    for k, lev in enumerate(LEVS):
        dd = [r[1][k] for r in res]
        w = sorted(r[0][k] for r in res)
        print("   {:8.2f}  {:7.3f}  {:7.3f}  {:5.3f}  {:13.2f}  {:8.2f}"
              .format(lev, sum(d >= 0.5 for d in dd) / 1e4,
                      sum(d >= 0.8 for d in dd) / 1e4,
                      sum(x <= 0 for x in w) / 1e4, w[5000], w[1000]))
#    leverage  DD>=50%  DD>=80%   ruin  median wealth  10th pct
#        0.73    0.181    0.009  0.000           1.77      0.77
#        1.46    0.564    0.173  0.000           2.21      0.38
#        2.92    0.878    0.584  0.001           2.05      0.03
#        7.10    1.000    0.956  0.640           0.00      0.00

Measured from the running peak, half Kelly has a drawdown of half or more with chance 0.181 and of 80% or more with 0.009. Full whole-history Kelly reaches half with 0.564 and 80% with 0.173; twice Kelly 0.878 and 0.584. The bull-month Kelly of 7.10 has a drawdown of half or more on every one of the 10,000 paths, of 80% or more on 0.956 of them, and is wiped out on 0.640: its median wealth after ten years is 0.00. The median wealth of twice Kelly (2.05) is below that of full Kelly (2.21), while its tenth percentile is 0.03 against 0.38.

Chance of a drawdown of half or more, and of 80% or more, within ten years, by leverage, in the two-regime market (p04).

Hypothesis · not established

The author’s interpretation (not yet tested): most of the deep drawdowns suffered by leveraged individual accounts come from leverage set at two or more times a long-run Kelly estimate, rather than from bad luck at moderate leverage. What would show it wrong: in account-level data, accounts whose leverage was at or below a long-run Kelly estimate account for as large a share of drawdowns of half or more as accounts above twice that estimate.

Summary

A Kelly bettor halves its wealth at some point with chance one half; beyond twice Kelly every low is eventually reached. In a market with a bear regime, the full Kelly computed from bull months alone (about seven times wealth) led to a drawdown of half or more on every path and to ruin on 0.640 of them within ten years.

References

MacLean, Leonard C., Edward O. Thorp, and William T. Ziemba, eds. 2011. The Kelly Capital Growth Investment Criterion: Theory and Practice. World Scientific.
Thorp, Edward O. 2006. “The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market.” In Handbook of Asset and Liability Management, Volume 1, edited by Stavros A. Zenios and William T. Ziemba. North-Holland.

Fan, Y. (2026). Regime-Switching Fractional Kelly, Maximum-Drawdown Constraints and Lifecycle Leverage Allocation: Cyclical Drift Estimation Errors, Ruin-Bounded Bet Sizing and Human-Capital Amortization [体制转换分数凯利、最大回撤约束与生命周期杠杆配置] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090256 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).