Kelly Criterion

Kelly criterion (also known as the scientific gambling method, Kelly bet, or growth-optimal portfolio strategy) is a foundational mathematical rule in financial economics, probability theory, and business and economics theory for sizing sequential wagers and leveraged portfolio allocations under uncertainty. Formulated in 1956 by Bell Labs physicist John L. Kelly Jr. in his landmark Bell System Technical Journal paper "A New Interpretation of Information Rate", the criterion determines the optimal fraction f* of current capital that an investor or decision-maker should allocate to a favorable stochastic opportunity at each step of a repeated multiplicative process in order to maximize the expected geometric (logarithmic) growth rate of wealth. Directly bridging Shannon entropy and communication channel capacity with modern portfolio theory, Kelly proved that a gambler receiving side information over a noisy binary symmetric channel maximizes the long-run exponential rate of capital compounding by wagering a fraction of wealth equal to the theoretical edge divided by the payout odds—a rule subsequently generalized to continuous-time asset markets and applied famously in blackjack card counting and quantitative hedge fund management by mathematician Edward O. Thorp.

Why Maximizing Expected Value Leads to Ruin

Consider an investor starting with initial capital W₀ > 0 who faces an infinite sequence of independent and identically distributed (i.i.d.) favorable binary bets t = 1, 2, …, N. On each trial, wagering a fraction f ∈ [0, 1] of current wealth W_t−1 wins with known probability p ∈ (0, 1)—paying net odds of b > 0 dollars per dollar wagered—and loses the wagered stake with probability q = 1 − p, where the bet has a positive expected return (b p − q > 0).

If a naive risk-neutral bettor attempts to maximize the single-period arithmetic expected wealth E[W_t] = W₀(1 + f(bp − q))^t, the linear objective is maximized at the boundary f = 1 (wagering 100% of capital on every single trial). However, because wealth evolves multiplicatively rather than additively, a single loss on any trial t wipes out all accumulated capital (W_t = 0 forever). Because the probability of surviving N consecutive trials without a single loss is p^N → 0 as N → ∞, the f = 1 strategy results in almost-sure bankruptcy (lim_N→∞ W_N = 0 with probability 1), illustrating the profound divergence between ensemble arithmetic expectation and time-average geometric growth (a mechanism closely related to Daniel Bernoulli's 1738 logarithmic utility resolution of the St. Petersburg paradox).

Derivation of the Optimal Fraction f*

Suppose that out of N trials, the bettor experiences S wins and F = N − S losses, and reinvests a constant fraction f ∈ [0, 1) of current bankroll on every trial. Terminal wealth after N trials is:

Taking the N-th root and applying the natural logarithm yields the per-trial exponential compounding rate G_N(f) ≡ (1/N) ln(W_N / W₀). By Kolmogorov's Strong Law of Large Numbers, as N → ∞ the empirical win frequency S/N converges almost surely to p and (N − S)/N converges almost surely to q = 1 − p, so G_N(f) converges with probability 1 to the deterministic asymptotic expected logarithmic growth rate G(f) = E[ln(1 + f R)]:

Because G″(f) = −p b² / (1 + f b)² − q / (1 − f)² < 0, G(f) is strictly concave on [0, 1). Setting the first derivative G′(f) = 0 yields the canonical Kelly criterion formula:

In the special case of even-money odds (b = 1), the optimal Kelly fraction reduces to f* = p − q = 2p − 1. More generally, if a losing bet loses only a fraction a ∈ (0, 1] of the wagered amount while a winning bet gains fraction b > 0, the generalized binary Kelly fraction is:

Kelly's (1956) celebrated insight was that when b = 1 (fair even-money payout odds against a bookmaker) and a gambler receives noisy insider transmissions over a binary symmetric channel that transmits the true winner correctly with probability p > 1/2 and flips the bit with error probability q = 1 − p, substituting f* = p − q and 1 + f* = 2p, 1 − f* = 2q into the logarithmic growth formula (measured in base-2 logarithms) yields:

That is, the maximum exponential growth rate of the gambler's capital in doublings per trial equals the Shannon channel capacity C = 1 − H₂(p) (in bits per transmission) of the communication channel, establishing a rigorous mathematical isomorphism between capital compounding and Shannon entropy.

Geometric Brownian Motion and Quadratic Approximation

In continuous-time financial markets (Thorp, 1971; Merton, 1969), suppose a risky asset price S_t follows geometric Brownian motion with instantaneous drift μ and volatility σ > 0 alongside a risk-free asset earning interest rate r. By Itô's lemma, allocating a constant leveraged portfolio weight f to the risky asset and 1 − f to the risk-free bond yields the instantaneous logarithmic wealth growth rate (geometric drift):

Notice the quadratic volatility drag term −(f²σ²)/2: variance directly erodes geometric compounding. Maximizing g(f) with respect to f yields the continuous-time Kelly–Merton fraction:

where SR ≡ (μ − r) / σ is the asset's Sharpe ratio. Expressing any arbitrary allocation as a multiple f = c · f* (where c > 0 is the Kelly multiplier) reveals the parabolic relationship g(c f*) − r = c(2 − c) · (SR² / 2). At Half-Kelly (c = 1/2), the investor captures c(2 − c) = (1/2)(3/2) = 75% of the maximum excess geometric growth rate while experiencing only (1/2)² = 25% of the portfolio variance (cutting standard deviation in half). Conversely, at Double-Kelly (c = 2), the excess geometric growth rate vanishes completely (g(2f*) = r), and any leverage c > 2 produces negative excess geometric growth.

Breiman's Optimality Theorems vs. Samuelson's Critique

In 1961, mathematician Leo Breiman proved two landmark asymptotic optimality properties of the Kelly strategy W_N*:

1. Asymptotic Dominance: For any other essentially different admissible strategy W_N, lim_N→∞ (W_N* / W_N) = ∞ with probability 1.
2. Minimum Expected Time to Goal: The Kelly strategy minimizes the asymptotic expected number of trials E[T(M)] required for wealth to reach any large pre-assigned target threshold M → ∞.

Despite these properties, Nobel laureate Paul Samuelson (1971, 1979) mounted a celebrated critique—culminating in his paper written in words of one syllable, "Why We Should Not Make Mean Log of Wealth Big Though Years to Act Are Long"—arguing that the Law of Large Numbers does not imply that an investor with non-logarithmic Constant Relative Risk Aversion (CRRA) utility U(W) = W^1−γ / (1 − γ) (with risk aversion γ ≠ 1) should maximize E[ln W]. Specifically, an investor with relative risk aversion γ > 1 optimally chooses the Merton fraction f_γ = (μ − r) / (γ σ²) = (1/γ) f*, providing an exact utility-theoretic microfoundation for Edward O. Thorp's Fractional Kelly strategy (c = 1/γ < 1), which simultaneously guards against parameter estimation error in μ and reduces the probability of large interim drawdowns.

Why do quantitative hedge funds and professional bettors usually use Half-Kelly instead of Full Kelly?

Full Kelly (f*) assumes that the investor knows the true drift μ (or win probability p) and variance σ² without error. In real financial markets, expected returns μ are estimated from noisy historical data and are prone to upward estimation bias. Because the geometric growth curve g(f) is asymmetric in estimation risk—under-betting by 50% (Half-Kelly) retains 75% of maximum growth with half the volatility, whereas over-betting by 100% (Double-Kelly) destroys 100% of excess growth while doubling volatility—practitioners use Fractional Kelly (f = 0.25*f** to 0.50*f**) to stay safely on the left side of the growth peak and limit maximum drawdowns.

Does the Kelly criterion guarantee that an investor will never lose money over a finite horizon?

No. Although the Kelly fraction f* < 1 guarantees zero probability of total bankruptcy in discrete binary trials (since 1 − f* > 0 always leaves positive residual capital to scale future bets), a Full-Kelly bettor still experiences substantial interim drawdowns: in the continuous diffusion limit, a Full-Kelly portfolio has a 1/2 probability of suffering a 50% drawdown from its high-water mark at some point in the future, and a 1/k probability of dropping to 1/k of its peak wealth. Reducing the allocation to Fractional Kelly f = c f* reduces the probability of dropping to fraction x ∈ (0, 1) of initial wealth from x down to x^(2/c)−1 (so at Half-Kelly c = 0.5, the probability of a 50% drawdown drops from 50% to (0.5)³ = 12.5%).

1. Kelly, J. L., Jr. (1956). "A New Interpretation of Information Rate." The Bell System Technical Journal, 35(4), 917–926. https://doi.org/10.1002/j.1538-7305.1956.tb03809.x
2. Breiman, L. (1961). "Optimal Gambling Systems for Favorable Games." In Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability (Vol. 1, pp. 65–78). University of California Press. https://projecteuclid.org/euclid.bsmsp/1200512159
3. Thorp, E. O. (1971). "Portfolio Choice and the Kelly Criterion." Proceedings of the Business and Economic Statistics Section, American Statistical Association, 215–224. https://doi.org/10.1142/9789814293501
4. Samuelson, P. A. (1979). "Why We Should Not Make Mean Log of Wealth Big Though Years to Act Are Long." Journal of Banking & Finance, 3(4), 305–307. https://doi.org/10.1016/0378-4266(79)90023-290023-2)
5. MacLean, L. C., Thorp, E. O., & Ziemba, W. T. (2010). "Good and Bad Properties of the Kelly Criterion." Quantitative Finance, 10(7), 681–687. https://doi.org/10.1080/14697688.2010.506108