Kelly Criterion: Theory & Applications

• Kelly Rule is a wealth allocation strategy that maximizes expected logarithmic growth by optimizing the fraction of wealth allocated to repeated gambles and portfolios.
• It features both discrete and continuous-time formulations, with adaptations for drawdown constraints, finite-horizon objectives, and risk management.
• Modern extensions include robust optimization with ambiguity sets, including Wasserstein balls, and applications in prediction markets, machine learning, and beyond.

to=arxiv_search tool code: {"query":"Kelly criterion rule stop-loss distributionally robust Wasserstein Kelly prediction markets quantile Kelly arXiv", "max_results": 10} to=arxiv_search tool code: {"query":"Kelly criterion arXiv (Nielsen, 2013, Busseti et al., 2016, Sun et al., 2018, Li, 2023, Beygelzimer et al., 2012, Long, 19 Apr 2026)", "max_results": 10} The Kelly Rule, or Kelly criterion, is a prescription for allocating fractions of wealth across repeated gambles or portfolios so as to maximize expected logarithmic wealth and, under multiplicative wealth dynamics, the long-run growth rate. In the canonical binary gamble with win probability $p$, loss probability $1-p$, and net odds $b$, the optimal fraction is $f^*=\frac{bp-(1-p)}{b}$; in general portfolio form it is the optimization of $\mathbb{E}[\log(1+f^\top R)]$, and in continuous-time geometric Brownian motion with risky drift $\mu$, volatility $\sigma$, and risk-free rate $r$ it yields the constant risky weight $(\mu-r)/\sigma^2$ [(Hsieh et al., 2017); (Lillo et al., 26 Aug 2025); (Nielsen, 2013)]. Modern work treats the rule as a growth-optimal principle that extends to continuous-time control, drawdown constraints, ambiguity sets, finite-horizon quantiles, prediction markets, and several non-financial settings.

1. Formal definition and canonical formulations

In its most general single-period form, the Kelly problem is

$\max \mathbb{E}[\log W_{t+1}],$

with the feasible portfolio or betting rule chosen so that wealth remains nonnegative (Lillo et al., 26 Aug 2025). In repeated IID gambles, this is equivalent to maximizing the expected logarithmic growth rate per play, which is the standard long-run interpretation of the criterion (Hsieh et al., 2017). For a vector of random returns $1-p$0, allocation vector $1-p$1, and wealth recursion $1-p$2, the admissible set is determined by the survival condition $1-p$3 for every $1-p$4 in the support; the optimization becomes

$1-p$5

This support-aware formulation is central whenever rare outcomes can force bankruptcy (Hsieh et al., 2017).

Several equivalent parameterizations are standard. In the long-only gambling model of Busseti, Ryu, and Boyd, one chooses $1-p$6 with $1-p$7 to maximize $1-p$8, where $1-p$9 is a nonnegative return vector and one component is risk-free cash (Busseti et al., 2016). In discrete-time stock-market approximations with excess returns $b$0, a second-order expansion gives the moment system

$b$1

so that $b$2 in the unconstrained approximation; for Gaussian excess returns this becomes $b$3 (Byrnes et al., 2018).

For a single multinomial event with subjective probabilities $b$4, state prices $b$5, explicit stakes $b$6, and cash fraction $b$7, the full Kelly solution can be written in state-price form as

$b$8

with the activity threshold $b$9. In this representation, cash acts as an implicit position in every outcome, and the optimal terminal wealth satisfies $f^*=\frac{bp-(1-p)}{b}$0 (Long, 13 Mar 2026).

2. Continuous-time growth optimality and dynamic control

In continuous time, the standard Kelly problem is embedded in a market with one risky asset $f^*=\frac{bp-(1-p)}{b}$1 and one risk-free asset $f^*=\frac{bp-(1-p)}{b}$2, where

$f^*=\frac{bp-(1-p)}{b}$3

With discounted wealth $f^*=\frac{bp-(1-p)}{b}$4 and risky fraction $f^*=\frac{bp-(1-p)}{b}$5, the discounted wealth dynamics are

$f^*=\frac{bp-(1-p)}{b}$6

For log utility $f^*=\frac{bp-(1-p)}{b}$7, the Hamilton–Jacobi–Bellman equation yields the classical constant Kelly fraction

$f^*=\frac{bp-(1-p)}{b}$8

together with the value function

$f^*=\frac{bp-(1-p)}{b}$9

The paper "The Kelly growth optimal strategy with a stop-loss rule" shows that the optimal control itself satisfies a nonlinear PDE independent of the terminal utility,

$\mathbb{E}[\log(1+f^\top R)]$0

or, in terms of $\mathbb{E}[\log(1+f^\top R)]$1,

$\mathbb{E}[\log(1+f^\top R)]$2

This makes the unconstrained Kelly rule a trivial constant solution and permits direct numerical solution once boundary conditions for $\mathbb{E}[\log(1+f^\top R)]$3 are known (Nielsen, 2013).

The same continuous-time logic appears in leverage problems. With broker funding cost $\mathbb{E}[\log(1+f^\top R)]$4, client borrowing rate $\mathbb{E}[\log(1+f^\top R)]$5, risky weight $\mathbb{E}[\log(1+f^\top R)]$6, and risky-asset GBM $\mathbb{E}[\log(1+f^\top R)]$7, the instantaneous expected log-growth is

$\mathbb{E}[\log(1+f^\top R)]$8

Garivaltis shows that if the broker and the Kelly gambler Nash-bargain jointly over leverage and the margin rate, then the investor chooses the risky position as if borrowing at the broker’s call rate,

$\mathbb{E}[\log(1+f^\top R)]$9

while the negotiated lending rate depends on the threat point; under zero margin loans as the disagreement point,

$\mu$0

This preserves growth-optimal sizing while reallocating surplus between borrower and broker (Garivaltis, 2019).

3. Constraints, drawdown control, and finite-horizon modifications

A large contemporary literature modifies the Kelly Rule because pure log-growth optimality can tolerate pathwise risk that is unacceptable operationally. Busseti, Ryu, and Boyd formulate a risk-constrained Kelly problem by adding a drawdown surrogate to the classical program. If $\mu$1 and the target is $\mu$2, they prove that, for $\mu$3,

$\mu$4

Replacing the original path-dependent constraint by this bound gives the convex Risk-Constrained Kelly program

$\mu$5

The same paper shows that a quadratic approximation links this formulation to a Markowitz-type mean-variance program (Busseti et al., 2016).

Periodic stop-loss rules alter the dynamic allocation even in the continuous-time GBM benchmark. If discounted wealth hits a boundary $\mu$6, the risky weight is set to zero until the next reset date. In scaled variables $\mu$7, $\mu$8, and $\mu$9 with $\sigma$0, the control PDE becomes

$\sigma$1

with boundary conditions $\sigma$2, $\sigma$3, and $\sigma$4. The long-horizon asymptotic is $\sigma$5, giving the CPPI-like limit $\sigma$6 (Nielsen, 2013).

Finite-horizon objectives can depart sharply from ordinary log-growth maximization. "Exact Finite-Horizon Quantile Kelly for Repeated Multi-Outcome Events" considers an $\sigma$7-outcome event repeated $\sigma$8 times in Arrow–Debreu coordinates $\sigma$9, where terminal wealth is the monomial $r$0 for multinomial count vector $r$1. For fixed upper quantile level $r$2, the terminal-wealth quantile is a positively homogeneous piecewise-monomial function on the closed wealth simplex; on each chamber of the multinomial count arrangement, the problem reduces exactly to a one-period Kelly problem for the shadow law $r$3. The unique unconstrained one-period Kelly point is

$r$4

but exact finite-horizon quantile maximizers can lie on support faces and need not equal $r$5; asymptotically, however,

$r$6

This suggests that finite-horizon quantile control is a genuine alternative objective rather than a perturbation of ordinary Kelly (Long, 19 Apr 2026).

A simpler departure from the constant-payoff model occurs when the payoff $r$7 on a winning round is itself random. Then the expected log-growth is

$r$8

and the Kelly fraction is characterized by the integral equation

$r$9

The paper proves that this $(\mu-r)/\sigma^2$0 is strictly smaller than the classical Kelly fraction computed using the same average payoff $(\mu-r)/\sigma^2$1, unless $(\mu-r)/\sigma^2$2 is almost surely constant (Pérez-Marco, 2014).

4. Estimation risk, ambiguity, and the problem of overconfidence

A recurrent criticism of the Kelly Rule is not its objective but its dependence on a fully specified return law. One line of work argues that under some parametric models the criterion is not too aggressive but too conservative. The "Restricted Betting Theorem" states that any Kelly optimizer $(\mu-r)/\sigma^2$3 under true support $(\mu-r)/\sigma^2$4 must satisfy

$(\mu-r)/\sigma^2$5

In the scalar case this confines the optimizer to

$(\mu-r)/\sigma^2$6

If the support is unbounded above and below, as with a normal distribution, then the theorem forces $(\mu-r)/\sigma^2$7. The paper’s argument is that the true-support solution can therefore be much smaller than the empirical Kelly solution $(\mu-r)/\sigma^2$8, because finite samples truncate tails and expand the empirical admissible set (Hsieh et al., 2017).

A second response is to replace the nominal distribution by an ambiguity set and maximize worst-case log-growth. Sun and Boyd define the distributionally robust Kelly problem

$(\mu-r)/\sigma^2$9

and show that it remains a convex optimization problem. They provide tractable formulations for polyhedral sets, ellipsoidal sets, $\max \mathbb{E}[\log W_{t+1}],$0-divergence balls, Wasserstein balls, and moment-based uncertainty sets, and extend Breiman-type asymptotic optimality to varying distributions constrained to lie in a fixed uncertainty set (Sun et al., 2018).

The Wasserstein variant of this idea is developed further in "Wasserstein-Kelly Portfolios." Rather than robustifying directly over simple returns, that paper places the Wasserstein ball on log-returns $\max \mathbb{E}[\log W_{t+1}],$1, so that the robust objective becomes

$\max \mathbb{E}[\log W_{t+1}],$2

The modeling shift avoids the $\max \mathbb{E}[\log W_{t+1}],$3 pathologies that arise if the adversary can place mass on outcomes with nonpositive gross return. The resulting problem admits a finite-dimensional convex reformulation, and the reported S&P 500 experiments show greater stability and better out-of-sample performance than empirical Kelly across multiple metrics (Li, 2023).

Estimation risk can also be addressed by enlarging the investment menu. In a discrete-time binomial stock–bond market with a European put, the paper "Tackling estimation risk in Kelly investing using options" shows that fairly priced options are redundant under correct specification, but can be used to construct a class of Kelly portfolios that is asymptotically robust to misspecification of the up/down move. Two Kelly-with-option portfolios with opposite hedges can be convexly combined so that the long-run growth rate converges to that of the better constituent under the realized model (Lillo et al., 26 Aug 2025).

5. Aggregation, market applications, and institutional implementations

When traders themselves follow Kelly, the rule induces an aggregation mechanism. In a prediction market security priced at $\max \mathbb{E}[\log W_{t+1}],$4 and paying $\max \mathbb{E}[\log W_{t+1}],$5 on the event, an agent with belief $\max \mathbb{E}[\log W_{t+1}],$6 and wealth $\max \mathbb{E}[\log W_{t+1}],$7 chooses demand

$\max \mathbb{E}[\log W_{t+1}],$8

Market clearing implies

$\max \mathbb{E}[\log W_{t+1}],$9

so the equilibrium price is the wealth-weighted average of beliefs. After the outcome is realized, wealth updates exactly as if the market were applying Bayes’ Law to the index of traders, and the market’s cumulative log loss satisfies the worst-case regret bound

$1-p$00

With fractional Kelly, the same paper shows that agent $1-p$01 behaves like a full-Kelly trader with shrunk belief $1-p$02, and the market price converges to a time-discounted frequency rather than the raw empirical frequency (Beygelzimer et al., 2012).

Institutional implementations often combine Kelly sizing with additional state variables. In systematic short-dated SPXW put-writing, Kelly sizing is estimated by Monte Carlo under GBM using empirical win probabilities and average gain/loss magnitudes, then translated into contract counts through a margin model. The study compares three schemes—pure Kelly, VIX-rank scaling, and a multiplicative hybrid—and reports that ultra-short-dated, far out-of-the-money options deliver superior risk-adjusted returns, while the hybrid rule better balances return generation and drawdown control, particularly in low-volatility conditions such as 2024 (Wysocki, 9 Aug 2025).

The rule is also routinely approximated in feedback-control language. In the single-asset recursion

$1-p$03

the exact Kelly objective is $1-p$04. A second-order Taylor approximation gives

$1-p$05

together with closed forms for expected cumulative gain, its variance, and a survivability interval $1-p$06 that guarantees no bankruptcy. The paper emphasizes that the approximation is useful only when the implied $1-p$07 respects survivability or is explicitly saturated into the admissible interval (Hsieh, 2020).

6. Extensions, reinterpretations, and distinct uses of “Kelly”

The phrase “Kelly” is not unique to log-optimal gambling. In social choice theory, it refers to Kelly’s preference extension from alternatives to sets of alternatives: $1-p$08 Under this extension, Kelly-strategyproofness for set-valued social choice functions leads to strong indecisiveness results with weak preferences: every strategyproof rank-based SCF violates Pareto-optimality in the stated parameter ranges, every strategyproof support-based Pareto-optimal SCF returns at least one top choice of every voter, and every non-imposing strategyproof SCF returns a Condorcet loser somewhere (Brandt et al., 2021). This is a distinct technical meaning of “Kelly” and not a growth-optimal investment rule.

The language of Kelly has also been imported into machine learning and information theory. "Diffusion Models are Kelly Gamblers" interprets conditional diffusion as a Kelly-style use of side information: the additional information required to bind $1-p$09 to conditioning variable $1-p$10 is exactly

$1-p$11

and classifier-free guidance is described as boosting this binding at sampling time, albeit without corresponding to the reversal of any forward diffusion process (Premkumar, 28 Sep 2025). This suggests an information-theoretic rather than financial reading of Kelly optimality.

A further reinterpretation appears in epistemic decision theory. In the Sleeping Beauty problem, if bets are specified as fractions of wealth so that dynamics are multiplicative, then maximizing expected log-growth over the experiment yields an awakening bet fraction $1-p$12 at even odds, which implies the posterior $1-p$13 and therefore the “thirder” position. The same paper argues that a thirder using growth-based acceptance is immune to Dutch books of the relevant multiplicative kind, whereas a halfer is not (Abramowitz, 26 Sep 2025).

Taken together, these developments suggest that the Kelly Rule is best understood as a family resemblance rather than a single formula. Its core remains the maximization of expected logarithmic growth under multiplicative dynamics, but current research shows that the resulting optimizer depends sharply on support geometry, trading constraints, horizon, ambiguity set, market protocol, and even on whether “Kelly” names a betting criterion at all.