← All writing
Note · Quantitative finance

How much leverage the typical path can take

Leverage multiplies the drift but squares the volatility drag, so there is a best amount of it. For a volatile asset that amount is well below fully invested.

Nov 20265 min readKelly criterionStochastic calculus

The previous note ended on an uncomfortable fact: a price that drifts up 6% a year with 38% volatility has a median that shrinks, because volatility drag (σ2/2=7.2%\sigma^2/2 = 7.2\%) is larger than the drift. The obvious next question is practical. You don’t have to put all your money in that asset. If you can choose how much to hold, and even borrow to hold more, which amount makes the typical outcome grow fastest?

The answer has a name, the Kelly criterion, and a one-line derivation.

A coin first

Take a bet that pays even money and wins with probability p=0.55p = 0.55. You stake a fraction ff of your wealth each round. After nn rounds with kk wins, wealth is Wn=W0(1+f)k(1−f)n−kW_n = W_0 (1+f)^k (1-f)^{n-k}, so the growth rate per round is

G(f)=pln⁡(1+f)+(1−p)ln⁡(1−f).G(f) = p \ln(1+f) + (1-p)\ln(1-f).

Setting G′(f)=0G'(f) = 0 gives f∗=2p−1=10%f^* = 2p - 1 = 10\%, and a growth rate of about 0.50%0.50\% per round. Stake 20%20\% instead, twice as much, and the growth rate falls to almost exactly zero: twice the risk, none of the reward. The rule comes from a 1956 paper by John Kelly, then at Bell Labs, on the information rate of noisy communication channels [1]; the investing reading came later.

The continuous version

Now hold a fraction LL of wealth in an asset that follows geometric Brownian motion, keep the rest in cash earning rr, and rebalance continuously. L>1L > 1 means borrowing. Wealth itself follows a GBM, with drift r+L(μ−r)r + L(\mu - r) and volatility LσL\sigma, so the result from the previous note applies directly: the median grows at

g(L)=r+L(μ−r)−12L2σ2.g(L) = r + L(\mu - r) - \tfrac{1}{2} L^2 \sigma^2 .

The reward is linear in LL and the drag is quadratic. The parabola peaks at

L∗=μ−rσ2,g∗=r+(μ−r)22σ2.L^* = \frac{\mu - r}{\sigma^2}, \qquad g^* = r + \frac{(\mu - r)^2}{2\sigma^2}.

Two consequences fall out of the parabola. Holding a fraction cc of the Kelly amount earns (2c−c2)(2c - c^2) of the maximum excess growth, so half Kelly keeps three quarters of the growth with half the volatility. And at c=2c = 2 the excess growth is zero: you carry all the volatility for nothing.

Try it: choose the leverage
Kelly L*–
Median growth–
Paths below start–
Ever halves–
Cash rate 0. Left: median growth g(L) = Lμ − L²σ²/2. Right: 200 one-year paths rebalanced daily. "Ever halves" is the model's chance of ever falling to half, over any horizon.

What it says about the hero’s asset

With μ=6%\mu = 6\%, σ=38%\sigma = 38\% and r=0r = 0, Kelly says L∗=0.42L^* = 0.42: keep 58% of the money in cash. The best achievable median growth is a modest 1.25%1.25\% a year, while holding the asset outright (L=1L = 1) shrinks the median by 1.2%1.2\% a year.

A daily-rebalanced simulation (one year, 20,000 paths) shows what each choice does to the typical outcome:

Leverage LLMedian wealthMean wealthPaths below start
0.42 (Kelly)1.0151.02846%
10.9931.06851%
20.8541.14358%
30.6351.22966%

The mean keeps rising with leverage while the median collapses. A three-times leveraged, daily-rebalanced product on this asset loses about a third of its value on a typical year even though its expected value goes up. This is the “leveraged ETF decay” people complain about, and it needs no fees or tracking error to appear.

For a calmer asset the answer flips. A broad equity index with μ=8%\mu = 8\%, r=4%r = 4\% and σ=16%\sigma = 16\% gives L∗=1.56L^* = 1.56: Kelly says borrow. The criterion is not “use less leverage”, it is “size by μ/σ2\mu/\sigma^2”.

Why nobody bets full Kelly

The drift is the number you know worst. The standard error of an estimated annual drift is about σ/T\sigma/\sqrt{T} for TT years of data: with 38% volatility and ten years of history that is ±12%\pm 12\% around a 6%6\% estimate. L∗L^* is linear in μ\mu, so the “optimal” leverage is uncertain by a factor of two or more.

Drawdowns are deep even at the optimum. In this model log-wealth is a Brownian motion with drift, and the chance that wealth ever falls to a fraction xx of its starting value is

P(ever reach xW0)=x 2/c−1\mathbb{P}(\text{ever reach } x W_0) = x^{\,2/c - 1}

for a fraction cc of Kelly (with r=0r = 0). At full Kelly the chance of halving at some point is 1/21/2; at half Kelly it is 1/81/8. Most people discover they cannot sit through the full-Kelly path.

The left wall is a cliff. In the continuous model the growth curve is a symmetric parabola, so overbetting and underbetting by the same amount cost the same. Real returns jump. In the coin version G(f)→−∞G(f) \to -\infty as f→1f \to 1: a single bad outcome at high leverage is not a slower path, it is the end of the path.

So practitioners treat L∗L^* as a ceiling and run a fraction of it, often a half.

Checking it numerically

kelly.py
import numpy as np
def growth(L, mu, sigma, r=0.0):
return r + L * (mu - r) - 0.5 * L**2 * sigma**2
def leveraged_paths(L, mu, sigma, days=252, n=20_000, seed=1):
rng = np.random.default_rng(seed)
dt = 1 / 252
z = rng.standard_normal((n, days))
asset = np.exp((mu - 0.5 * sigma**2) * dt + sigma * np.sqrt(dt) * z) - 1
return np.prod(1 + L * asset, axis=1) # rebalanced to L every day
mu, sigma = 0.06, 0.38
L_star = mu / sigma**2
print(f"Kelly leverage {L_star:.2f}, max growth {growth(L_star, mu, sigma):.2%}")
for L in (L_star, 1, 2, 3):
w = leveraged_paths(L, mu, sigma)
print(f"L={L:.2f} median {np.median(w):.3f} mean {w.mean():.3f} below start {(w < 1).mean():.0%}")

Takeaway

The median path is the one you will most likely live through, and it grows at μ−σ2/2\mu - \sigma^2/2, not at μ\mu. Leverage scales the first term and squares the second, which is why the same rule that says “borrow” for an index fund says “keep most of it in cash” for a volatile coin.

References

  1. J. L. Kelly Jr., “A new interpretation of information rate,” Bell System Technical Journal 35(4), 917–926 (1956). doi:10.1002/j.1538-7305.1956.tb03809.x