Kelly Criterion TradingOctober 5, 202618 min read

Turn Kelly Criterion Into Validated Position Sizes for Traders

Trader-focused, implementation-first guide to turn Kelly Criterion math into validated position sizes. Learn fractional Kelly, portfolio constraints, and...

!Geometric illustration of measured position sizing

The Kelly Criterion calculates the fraction of capital that maximizes long-run geometric growth, and traders use it as a mathematical ceiling rather than a literal order ticket. Compute full Kelly from your win rate and payoff ratio, then size positions with a fractional Kelly, typically half or less, after validating the inputs on out-of-sample data. The sections below walk through the formula, worked examples, portfolio math, and a complete validation workflow.

TL;DR:

Using full Kelly often results in large drawdowns, so most traders scale down to half or less to balance growth and risk.

Validating Kelly-based sizing with out-of-sample backtests helps ensure the estimated edge and payoff ratios hold in real market conditions.

Kelly sizing can recommend leverage above 100% when edge estimates are inflated, which signals potential overestimation or small sample bias.

Incorporating constraints such as maximum position size and diversification reduces portfolio risk beyond pure Kelly calculations.

Regular re-optimization and stress-testing with realistic costs and heavy-tailed scenarios improve the robustness of Kelly-based position sizing.

What the Kelly Criterion is and the core formulas traders need
Worked examples: computing Kelly from trade stats and sizing positions
Fractional Kelly and the growth versus drawdown trade-off
Multivariate Kelly: portfolios, covariance, and constraints
A practical workflow for sizing and validating Kelly-based positions
Common pitfalls and robust adjustments to Kelly sizing
When Kelly is the right benchmark and when it isn't
Validate your Kelly sizing rules with Backtestify
FAQ
Sources

What the Kelly Criterion is and the core formulas traders need

The Kelly criterion was derived as the betting fraction that maximizes the expected logarithm of wealth across repeated bets, which is mathematically equivalent to maximizing long-run compound growth. For a single binary bet, the formula is simple: f* = p − (q / b), where p is the probability of winning, q is the probability of losing (1 − p), and b is the ratio of the amount won to the amount risked on a win.

In trading terms, the same relationship shows up as f* = W − [(1 − W) / R], where W is your historical win rate and R is your average win divided by your average loss. This is the formula traders generally mean when they talk about "the Kelly formula" for a strategy with a fixed win rate and a fixed payoff structure per trade.

For continuous return streams, such as a single asset traded with varying position sizes over time rather than discrete win/loss bets, the approximation most commonly cited is:

f* = (μ − r) / σ²

Here μ is the expected return of the asset, r is the risk-free rate, and σ² is the variance of returns. This version treats the position as a continuously rebalanced allocation rather than a string of discrete trades, and it comes from the same log-utility maximization that produces the binary formula. Both versions answer the same question: what fraction of capital, bet repeatedly under the same statistical conditions, grows your account fastest over time.

The derivation rests on a few assumptions that matter more in markets than in a casino. It assumes the odds and probabilities (or the mean and variance) are known and stable across the betting sequence, that outcomes are independent or at least identically distributed, and that you can reinvest gains and losses continuously, resizing each bet as a fixed fraction of current capital. Trading violates all three to some degree: your edge drifts, consecutive trades are often correlated through shared market exposure, and execution happens in discrete lots rather than infinitely divisible fractions. The formula is still useful, but only as the ceiling on a sizing decision, not the decision itself.

It's worth separating two related ideas that get conflated in casual descriptions of Kelly. One is the growth-optimal property: Kelly sizing produces the highest expected compound growth rate over a long enough series of bets. The other is risk tolerance: nothing in the derivation says that growth-optimal is comfortable. A gambler or trader using full Kelly experiences wild equity swings on the way to that optimal growth, which is the central tension the rest of this guide addresses.

!What the Kelly Criterion is and the core formulas traders need — overview diagram

Worked examples: computing Kelly from trade stats and sizing positions

Start with a trading strategy that has been tested over a representative sample of trades: say it wins 55% of the time (W = 0.55), with an average winning trade of $300 and an average losing trade of $200, giving a payoff ratio R = 1.5.

Plugging into f* = W − [(1 − W) / R]:

f* = 0.55 − (0.45 / 1.5) = 0.55 − 0.30 = 0.25

Full Kelly says to risk a significant fraction of equity on each trade under these statistics. That number is an illustrative example built from the inputs above, not a market fact, and it is deliberately aggressive to show how the formula behaves.

To translate that fraction into an actual position size, you need the stop-loss distance, because Kelly gives you a fraction of equity to risk, not a number of shares or contracts directly.

Determine account equity (say $50,000).
Multiply by the chosen Kelly fraction to get dollar risk per trade (0.25 × $50,000 = $12,500 at full Kelly).
Divide dollar risk by the per-unit stop distance to get position size. If your stop is $2 away from entry on a stock, that's 6,250 shares; if it's $0.50 away, it's 25,000 shares.
Check that the resulting position fits margin, liquidity, and any exchange or broker constraints before entering.

Pro Tip: Always compute position size from the stop distance, not from a dollar amount you feel comfortable with: the stop defines your actual risk per unit, and skipping that step is how traders end up over-leveraged without realizing it.

Kelly math can also produce a fraction greater than 100%, which happens when the edge is large relative to the payoff variance, for example a very high win rate combined with a payoff ratio close to 1. A result above 1.0 technically calls for leverage greater than your account equity. In practice this is a signal that your win rate or payoff estimate is likely inflated by a small sample or survivorship-biased backtest, not a genuine invitation to lever up.

Fractional Kelly and the growth versus drawdown trade-off

Full Kelly is growth-optimal, but the ride is rougher than most traders can stomach, which is why fractional Kelly, scaling the computed fraction down by a constant multiplier c (commonly 0.5), is the default in practice rather than the exception.

The relationship between the fraction used and the growth achieved follows a known pattern: g(c f*) / g(f*) = c(2 − c), where g is the growth rate and c is the fraction of full Kelly applied.

Half Kelly sizing reduces the probability of a 50% drawdown from roughly 50% at full Kelly to about 12.5%, a steep cut in tail risk for a 25% sacrifice in long-run growth rate.

That asymmetry, a large reduction in catastrophic drawdown probability for a comparatively modest reduction in growth, is the entire argument for fractional Kelly in live trading. Choosing the right fraction comes down to a few operational criteria:

Drawdown tolerance: set the fraction so your worst plausible drawdown stays inside what you can psychologically and financially survive without abandoning the strategy.
Estimation confidence: the less certain you are about your win rate and payoff inputs, the smaller the fraction, since errors compound multiplicatively at higher sizing.
Capital constraints: smaller accounts facing fixed costs (commissions, minimum lot sizes) often can't follow fractional Kelly precisely and need rounding rules built in.
Correlation across positions: running several Kelly-sized trades at once on correlated instruments stacks risk that per-trade Kelly never accounted for.

Most practitioners settle between a quarter and a half of full Kelly, treating the computed f* as a ceiling they deliberately stay well under.

Multivariate Kelly: portfolios, covariance, and constraints

Sizing a single asset is the easy case. Real portfolios hold multiple positions simultaneously, and that changes the math because Kelly for a portfolio depends on the full covariance structure, not just each position's individual edge.

The portfolio-level formulation replaces the scalar f* with a vector f* = Σ⁻¹ (μ − r), where μ − r is the vector of expected excess returns for each asset and Σ is the covariance matrix of returns. The inverse covariance term is what makes this different from just computing Kelly separately for each position: it automatically shrinks allocations to correlated assets and can even push a position negative (a short) if it offsets risk elsewhere in the book.

This matters because computing Kelly independently for each trade and simply adding the fractions overstates how much you should actually be risking. Two setups that each look like a clean quarter-Kelly opportunity on their own can combine into far more portfolio-level risk than either alone if they move together, since a shared drawdown hits both at once. The covariance term is the only part of the formula that accounts for that overlap.

In practice, full matrix inversion only gets you partway there, because unconstrained Kelly portfolios routinely call for leverage, short positions, or concentration levels that no real account can or should run. The same Frontiers research on practical implementation recommends adding explicit constraints, no leverage, no shorting, maximum per-position weight, before solving. Three approaches handle this in practice:

Constrained quadratic optimization (solvers like quadprog) finds the Kelly-like allocation that maximizes expected log growth subject to no-short and no-leverage constraints, trading a closed-form answer for a numerically solved one.
Proportional fractional scaling takes the unconstrained Kelly vector and scales every position down by the same multiplier until the largest position and total exposure both fit within your caps, a cruder but much simpler fix.
Monte Carlo groping algorithms, described in academic work on multivariate Kelly, simulate many allocation candidates and search for the one that maximizes simulated growth under constraints, useful when the asset universe is too large or the return distribution too non-normal for a clean matrix solution.

Whichever method you use, re-optimizing on a fixed schedule rather than continuously is standard practice, since covariance and expected returns both drift and constant re-solving amplifies estimation noise into constant, costly rebalancing.

A practical workflow for sizing and validating Kelly-based positions

Turning the formulas above into a live sizing rule takes more than plugging numbers into an equation once. Investopedia's overview of the Kelly criterion frames this as a workflow: estimate your inputs from a trustworthy sample, compute the fraction, scale it down, translate it into real position sizes, cap your exposure, and then validate the whole thing before risking live capital.

Estimate W and R from a representative, out-of-sample trade record. Use a sample large enough to smooth out noise, and separate the data you used to build the strategy from the data you use to estimate its win rate and payoff; using the same trades for both inflates your edge estimate.
Compute full Kelly using the binary formula for discrete trade setups or the μ − r over σ² approximation for continuously held positions.
Choose a fractional multiplier, typically landing between a quarter and a half of full Kelly, based on your drawdown tolerance and confidence in the input estimates.
Translate the fraction into a position size by multiplying equity by the fractional Kelly value to get dollar risk, then dividing by your stop-loss distance to get shares or contracts.
Cap single-position weight and aggregate correlated exposure, even when the raw Kelly math suggests more, since no formula protects you from a cluster of correlated bets going wrong together.
Backtest and stress-test the rule across different market regimes, not just the period the edge was discovered in.
Schedule re-optimization and ongoing monitoring rather than recalculating Kelly reactively after a losing streak, which tends to lock in exactly the wrong adjustment at the wrong time.

Each step has a specific failure mode if skipped, and the validation stage deserves the most scrutiny because it's where overconfident sizing gets caught before it costs real money. A few checks matter most:

Sensitivity to mean estimation: small changes in your assumed win rate or average payoff can swing full Kelly substantially, since the formula is a direct, linear function of those inputs. Run the calculation across a plausible range of W and R rather than a single point estimate.
Rebalancing frequency and window width: the Frontiers study on practical Kelly implementation found that shorter rolling estimation windows with more frequent rebalancing can raise reported compound growth in backtests, but at a real cost in transaction fees and drawdown that a naive backtest can understate.
Heavy-tailed stress tests: because real return distributions have fatter tails than a clean Gaussian model assumes, the same research recommends testing sizing rules against heavy-tailed return assumptions, which tends to push the safe sizing fraction lower and raise measured drawdown risk.
Path-dependent drawdown metrics: average maximum drawdown and conditional expected drawdown (CED) should be standard outputs of any Kelly backtest, not an afterthought, since the growth-optimal property says nothing about how bad the path there gets.

Before sizing live capital, test the sizing rule on historical data and stress scenarios across multiple regimes rather than trusting a single backtest window.

This is precisely the stage where a dedicated backtesting platform earns its place in the workflow. Running a fractional Kelly rule against a strategy's actual historical trade record, rather than a theoretical win rate pulled from memory, surfaces whether the inputs you fed into the formula hold up out of sample. Reports on win rate, profit factor, and maximum drawdown let you see directly whether a quarter-Kelly or half-Kelly sizing rule would have produced a survivable equity curve on the data you actually have, instead of the one you assumed.

Pro Tip: Run your sizing rule against both the original version of a strategy and any rule changes you've made, side by side, so you can see whether the improvement actually reduced drawdown or just moved it around.

Common pitfalls and robust adjustments to Kelly sizing

Kelly's biggest practical danger is that it's far more sensitive to errors in your mean return estimate than to errors in variance, so a modest overestimate of your edge produces an oversized position long before anyone notices the variance assumption was also off. Bayesian or shrinkage approaches to estimating expected returns pull noisy point estimates toward a more conservative prior, producing sizing policies that adjust with new evidence rather than overreacting to a short recent streak.

A few concrete fixes address this directly:

Use fractional Kelly as the default, not full Kelly, specifically because it dampens the impact of mean-estimation error.
Cap single-position and correlated-bucket exposure independent of what the formula outputs, since no cap exists inside the Kelly equation itself.
Shrink win-rate and payoff estimates toward a conservative prior when the sample of trades is small or recent.
Backtest across varied rebalancing cadence and out-of-sample windows, including realistic transaction costs and slippage, and review conditional expected drawdown (CED) and average maximum drawdown as explicit outputs, not side notes.

When Kelly is the right benchmark and when it isn't

Kelly earns its place as the mathematical ceiling for sizing decisions, the number that tells you the absolute most you could rationally risk for maximum long-run growth under your stated assumptions. Treating it as a live autopilot is where traders get into trouble, because the formula has no opinion about your psychology, your liquidity constraints, or how confident you actually are in the win rate you fed it.

The more useful habit is to compute full Kelly, then deliberately throw most of it away: use a quarter or half of the result, validated against real out-of-sample trade data, with hard caps on correlated exposure that the formula itself will never give you. Kelly is a discipline for thinking about edge and risk together, not a command to be obeyed at face value. Treat the output as one well-reasoned input among several, and you get most of the benefit without the drawdown that full Kelly would have handed you on a bad stretch.

— WAJDI

Validate your Kelly sizing rules with Backtestify

Computing a Kelly fraction is only half the job. Knowing whether that fraction would have survived real market conditions is the part most traders skip, and it's the part a dedicated backtesting platform can address. Backtestify

With a backtesting platform, you can run a fractional Kelly sizing rule against a strategy's actual historical trade record instead of a theoretical win rate, and see the resulting win rate, profit factor, and maximum drawdown side by side with the original version of the rule.

Out-of-sample backtesting on real historical data, so your Kelly inputs come from verified performance rather than assumption.
Win rate, profit factor, and max drawdown reports generated automatically for every strategy you test.
Side-by-side comparisons between an original strategy and a sizing-adjusted version, so you can see exactly what a fractional Kelly rule changes.
A published strategy library of tested rules, including ones sourced from popular trading content creators, if you want a representative starting point rather than building one from scratch.

If you already have a strategy with a known win rate and payoff ratio, pull up our strategy library and test a half-Kelly sizing rule against its actual trade history before committing capital. Pro access runs $29 per month or $190 per year at Backtestify, and a Free tier is available if you want to start with a single strategy test.

FAQ

What is the Kelly Criterion in trading?

The Kelly Criterion is a formula that calculates the fraction of capital to risk on a bet or trade in order to maximize long-run compound growth. In trading, it's typically computed from a strategy's historical win rate and average win-to-loss ratio, then used as an upper bound rather than an exact position size.

How do you calculate the Kelly Criterion for trading?

For a binary win/loss setup, Kelly fraction equals your win rate minus the quotient of your loss rate divided by your payoff ratio: f* = W − [(1 − W) / R]. For continuously held positions, the common approximation is f* = (μ − r) / σ², using expected return, the risk-free rate, and return variance.

What is fractional Kelly and why do traders use it?

Fractional Kelly means scaling the full Kelly fraction down by a constant, commonly to half, before sizing a position. Research summarized by Thorp and the Berkeley review shows half Kelly cuts the probability of a 50% drawdown dramatically while only reducing long-run growth by about 25%, which is why it's the more common live sizing choice.

How does Kelly Criterion compare to fixed-fraction or risk-parity sizing?

Kelly sizes positions based on edge and payoff variance for each specific strategy, while fixed-fraction sizing risks the same percentage on every trade regardless of estimated edge, and risk parity allocates based on volatility contribution across a portfolio rather than expected return. Kelly can produce a mathematically higher long-run growth rate when inputs are accurate, but it's more sensitive to estimation error than either alternative.

Can Kelly Criterion sizing lead to ruin if inputs are wrong?

Yes. Because Kelly fractions respond linearly to your assumed win rate and payoff ratio, overestimating your edge from a small or biased sample can push full Kelly sizing well past what the real, unknown edge supports. Using a fractional Kelly multiplier and shrinking uncertain estimates toward a conservative prior both reduce this risk.