What Is the Kelly Criterion and Why Does It Optimize Position Size?

The Kelly Criterion is a mathematical formula that calculates the optimal fraction of your bankroll to wager on each bet—given your edge (win rate and payoff ratio)—to maximize long-term logarithmic wealth. It answers a deceptively simple question: If I know my historical win rate and the odds, what percentage of my account should I bet on this trade to maximize wealth over infinite iterations?

Derived in 1956 by John Kelly, a Bell Labs researcher, the formula has captivated traders, gamblers, and professional investors for seven decades. Its appeal is intuitive: instead of guessing at a percentage (2%, 5%, whatever), you calculate the optimal percentage based on data. Apply Kelly consistently, and probability theory guarantees you'll accumulate wealth faster than any other constant-fraction strategy, provided your edge is real and persistent.

Yet Kelly has a dark side. At full Kelly, you accept volatility that few humans can endure psychologically. Account swings of 30-40% are normal. Many traders who've applied Kelly at full strength have been bankrupted not by bad math, but by losing psychological discipline mid-drawdown and making desperate decisions. This tension—between mathematical optimality and human reality—defines the Kelly Criterion's place in trading practice.

Quick definition: The Kelly Criterion is a formula that calculates the optimal percentage of bankroll to risk per bet to maximize logarithmic wealth: f* = (p × b - q) / b, where p is win probability, q is loss probability, and b is the win/loss odds ratio.

Key takeaways

• The Kelly formula directly ties position size to your demonstrated edge: better odds = larger position, worse odds = smaller position
• Full Kelly maximizes long-term wealth but creates 30-40% account swings that often trigger psychological capitulation
• Half-Kelly and quarter-Kelly reduce position sizes to more tolerable levels while preserving most of the mathematical advantage
• Kelly only works if your edge is real, persistent, and accurately measured; a slight overstimation of win rate can lead to ruin
• Professional traders and hedge funds use fractional Kelly (one-quarter to one-half) as a practical compromise between growth and survivability

The Classical Kelly Formula

The Kelly Criterion calculates the optimal bet size (as a fraction of bankroll) as:

f* = (p × b - q) / b

Where:

• f* is the optimal fraction of bankroll to risk
• p is the probability of winning (0 to 1)
• q is the probability of losing (q = 1 - p)
• b is the win/loss odds ratio (reward divided by risk)

Let's define these variables precisely with a trading example. You backtest a swing-trading system on ES (the S&P 500 e-mini futures contract):

• p (win probability): Your system wins 60% of trades. So p = 0.60.
• q (loss probability): It loses 40% of trades. So q = 0.40.
• b (odds ratio): Your average winner captures 2 points; your average loser gives up 1 point. So b = 2/1 = 2.

Plugging into the formula:

f* = (0.60 × 2 - 0.40) / 2
f* = (1.20 - 0.40) / 2
f* = 0.80 / 2
f* = 0.40

Kelly tells you to risk 40% of your bankroll per trade. If your account is $100,000, you'd risk $40,000 on each trade.

Before you recoil: this is mathematically optimal under one critical assumption—that your edge is exactly as measured, persistent indefinitely, and that you can psychologically tolerate a near-ruin experience without abandoning the system.

Building the Kelly Formula from First Principles

Understanding where Kelly comes from illuminates why it works—and when it fails. The Kelly formula maximizes the expected logarithm of wealth over infinite trials:

log(Wealth_Final) = log(Wealth_Initial) + (n × p × log(1 + f×b) + n × q × log(1 - f))

Where n is the number of bets. Kelly finds the value of f that maximizes this expression.

The logarithmic utility function (log of wealth, not linear wealth) embeds a critical assumption: your utility from doubling your wealth is the same whether you go from $100 to $200 or from $1 million to $2 million. This is called constant relative risk aversion—your preference for risk stays consistent regardless of wealth level. For a trader with $100,000, doubling to $200,000 feels great; for a billionaire, doubling feels neutral. Log utility reflects this.

Taking the derivative of the wealth equation with respect to f and setting it to zero yields the Kelly formula. The mathematics is elegant, but the practical implication is brutal: at full Kelly, you're optimizing for logarithmic wealth, not for minimizing bankruptcy risk.

A Numeric Walkthrough: Kelly in Action

Let's simulate a trader using full Kelly on a $50,000 account over 20 trades, with the same edge as before (60% win rate, 2:1 odds, 0.40 Kelly).

Trade sequence (assume 60-40 split, actual order matters): Win, Win, Loss, Win, Loss, Win, Win, Win, Loss, Loss, Win, Win, Loss, Win, Win, Loss, Win, Win, Loss, Win.

After Trade 1 (Win): $50,000 + ($50,000 × 0.40 × 2) = $90,000 After Trade 2 (Win): $90,000 + ($90,000 × 0.40 × 2) = $162,000 After Trade 3 (Loss): $162,000 - ($162,000 × 0.40) = $97,200 After Trade 4 (Win): $97,200 + ($97,200 × 0.40 × 2) = $175,000

(Continuing through all 20 trades with the exact win sequence above...)

After Trade 20: $267,340

The account tripled in 20 trades. This is the power of Kelly: when your edge is real, position sizing compounds wealth explosively.

Now contrast with a trader using only 10% per trade (overly conservative):

Same trades, 10% allocation, 2:1 odds: After Trade 20: $96,500 (minimal growth)

And a trader using 60% Kelly (half-Kelly would be 20%, but let's say over-leveraged at 60%):

After Trade 20: $18,500 (bankruptcy!)

This last scenario illustrates why Kelly calibration is so sensitive. At 60% (50% overestimation of the optimal 40%), the account hits zero before the 20-trade sequence completes. A trader hitting zero bankroll stops trading—you can't recover from that.

The Win Rate and Odds Ratio Dependency

Kelly's position size is exquisitely sensitive to your inputs. If you overestimate your win rate or odds ratio, you'll risk too much. If you underestimate, you'll risk too little.

Scenario 1: True 60% win rate, 2:1 odds.

f* = (0.60 × 2 - 0.40) / 2 = 0.40 (40% Kelly)

Scenario 2: You think you have 60%, but it's actually 55%.

f* (calculated) = (0.60 × 2 - 0.40) / 2 = 0.40
f* (actual) = (0.55 × 2 - 0.45) / 2 = 0.325 (32.5%)

You're risking 40% when you should risk 32.5%. Over 100 trades, this 7.5-percentage-point overestimation can lead to ruin.

Scenario 3: True edge is 55%, 1.5:1 odds.

f* = (0.55 × 1.5 - 0.45) / 1.5 = 0.233 (23.3% Kelly)

A slightly worse win rate and lower payoff ratio halve your optimal position size. The formula is unforgiving: it scales to your actual edge, not your hopes.

Why Full Kelly Is Dangerous in Practice

Kelly assumes infinite trials (or at least many, many iterations), but traders operate with finite capital and finite careers. A sequence of just 10 trades using full Kelly, where half win and half lose (exactly the expectation), might produce one of several outcomes:

Lucky sequence (e.g., 5 wins, then 5 losses): Bankroll barely dented; you recover quickly.

Unlucky sequence (e.g., 5 losses immediately): Bankroll collapses, and you may quit before the wins arrive.

The law of large numbers guarantees that over 1,000 trades, your win rate converges to the true 60%. But in the first 100 trades, you might experience a 50% win rate (49 wins, 51 losses)—statistically plausible. If you're using full Kelly on a 60% assumed edge, and you hit 50% for the first 100 trades, your account can shrink 30-50%. Many traders psychologically capitulate during this drawdown, violating the system and locking in losses.

Furthermore, Kelly assumes no slippage, commissions, or errors. Real trading includes all three. A 1% drag from commissions and slippage reduces your effective edge, making your optimal Kelly fraction smaller. If you're using full Kelly without accounting for drag, you're oversizing.

Half-Kelly and Quarter-Kelly: The Professional Standard

To balance Kelly's mathematical optimality with human psychology, most professionals use fractional Kelly:

Practical Position Size = (Kelly Percentage) / Divisor

Half-Kelly divides by 2:

Full Kelly: 40% → Half-Kelly: 20% Full Kelly: 25% → Half-Kelly: 12.5%

Quarter-Kelly divides by 4:

Full Kelly: 40% → Quarter-Kelly: 10% Full Kelly: 25% → Quarter-Kelly: 6.25%

Why these fractions? Research and empirical observation show:

• Half-Kelly maintains roughly 75% of Kelly's long-term wealth accumulation while cutting maximum drawdown to 15-20% instead of 40%.
• Quarter-Kelly maintains roughly 50% of Kelly's edge while reducing drawdown to 8-12%, matching what many traders can psychologically endure.

The reduction in position size is more than offset by the increase in system survival: if you quit at a 40% drawdown, you accumulate zero wealth. If you survive a 15% drawdown psychologically, half-Kelly wins.

Applying Kelly to Equities and Futures

Equities example: You trade a breakout system on large-cap stocks. Backtesting on 100 trades:

• Wins: 55 (55% win rate, p = 0.55)
• Average winner: +1.5% (e.g., $150 gain on a $10,000 entry)
• Average loser: -1% (e.g., -$100 on a $10,000 entry)
• Odds ratio: b = 1.5 / 1 = 1.5

f* = (0.55 × 1.5 - 0.45) / 1.5
f* = (0.825 - 0.45) / 1.5
f* = 0.375 / 1.5
f* = 0.25 (25% Kelly)

Full Kelly says risk 25% of account per trade. For a $100,000 account, that's $25,000 risk per trade. But:

1. Verify position capital: $25,000 risk on a 1% stop means $2.5 million in stock value (25 times your account). This is clearly overlevered.
2. Use fractional Kelly: Apply half-Kelly (12.5%) instead.
3. Size to positions: 12.5% of $100,000 = $12,500 risk per trade.

Futures example: You trade ES (S&P 500 e-mini). Each 1-point move = $50 per contract. Backtesting:

• Win rate: 52% (barely profitable, p = 0.52)
• Average winner: 5 points = $250 per contract
• Average loser: 4 points = $200 per contract
• Odds ratio: b = 250 / 200 = 1.25

f* = (0.52 × 1.25 - 0.48) / 1.25
f* = (0.65 - 0.48) / 1.25
f* = 0.17 / 1.25
f* = 0.136 (13.6% Kelly)

On a $100,000 account, 13.6% Kelly = $13,600 risk. With a 4-point stop, that's 68 contracts ($13,600 / $200 per point per contract). This is massive leverage; a 3-tick move against you and you're wiped out. Practically, you'd apply quarter-Kelly: 3.4% = $3,400 risk = 17 contracts.

The Role of Data Quality and Sample Size

Kelly is only as reliable as your backtest data. Common errors:

Insufficient sample size: A backtest with 20 trades is too small. Kelly formula assumes at least 50-100 trades to stabilize the win rate estimate.

Curve-fitted overfitting: You optimize your system parameters to maximize Sharpe ratio on historical data, then apply Kelly to those optimized parameters. On live trading, the parameters don't generalize, and your actual edge is lower. Result: you're oversized and take catastrophic losses.

Ignoring market regime changes: Your 55% win rate in a trending market drops to 48% in a choppy market. Kelly sized for 55% will ruin you in chop if you don't detect the regime change.

Not accounting for slippage and commissions: Your backtest shows a 1% average win; after slippage and commissions, it's 0.8%. This 0.2% reduction in edge shrinks your Kelly fraction.

Professional traders reduce Kelly by 20-30% just to buffer for these unknowns. If the formula says 25%, they use 20%. If it says 40%, they use 28%. This "Kelly haircut" is humility in practice.

Kelly Criterion vs. Risk-of-Ruin

Risk-of-Ruin (RoR) is an alternative framework that calculates the probability of losing your entire bankroll before reaching a profit target. RoR and Kelly serve different purposes:

• Kelly optimizes long-term wealth growth assuming you never quit.
• RoR calculates the probability of catastrophic loss before success.

A trader using full Kelly on a 55% win rate system with 1:1 odds accepts a ~1% annual risk of ruin, provided they rebalance perfectly and don't deviate. But over a 20-year career, a 1% annual RoR compounds to a ~20% lifetime risk of ruin—non-trivial. Most traders demand RoR < 0.1% annually, which typically requires fractional Kelly.

Calibrating Kelly for Real Markets

Here's a practical process for applying Kelly to your system:

1. Backtest 50+ trades on out-of-sample data (data not used to build the system).
2. Record win rate, average winner, average loser.
3. Calculate full Kelly, then apply a 20% haircut (multiply by 0.80).
4. Use half or quarter of the result (divide by 2 or 4 more, depending on psychological tolerance).
5. Forward-test on paper trading for 20 trades.
6. If live trading, start with 50% of the calculated Kelly and scale up after 50 live trades of confirming the backtest results.

Common Mistakes

Applying Kelly without sufficient backtest data. You have 15 trades, see a 60% win rate, and calculate Kelly as 30%. Unlock! With a larger sample (100 trades), your win rate might drop to 52%, and Kelly falls to 15%. The formula is sensitive to sample size.

Using Kelly on curve-fitted, overoptimized systems. You've spent months tweaking parameters to achieve the best possible backtest. The system is likely overfit. Any real edge is smaller than your backtest suggests. Apply an aggressive Kelly haircut or avoid Kelly altogether.

Ignoring concurrent positions. You're using 25% Kelly on three concurrent trades. Your aggregate risk is 75%—a single adverse move across all three positions wipes you out. Kelly assumes sequential, independent bets; multiple concurrent positions violate this.

Not recomputing Kelly after live trading. Your backtest said 20% Kelly. After 50 live trades, your win rate is 47%, not 55%. You should recalculate Kelly on the combined data (backtest + live), but many traders don't, staying overexposed.

Confusing Kelly percentage with account risk percentage. Kelly tells you the optimal fraction to bet; it's not the same as "risk 2% of account per trade." They can align in some cases, but they're fundamentally different frameworks.

FAQ

Can I use Kelly Criterion if my system's edge varies by trade type?

Yes. Segment your trades by type; calculate Kelly for each segment independently. For example, if you have breakout trades (55% win rate) and reversal trades (48% win rate), compute separate Kelly fractions. Size each trade using its segment's Kelly.

What if my Kelly percentage is greater than 100%?

Mathematically, this means your edge is so strong that you should bet the entire bankroll repeatedly. In practice, it signals either an error in your calculations or a curve-fitted, overoptimized system. Even the best real-world trading edges rarely produce Kelly > 50%. If you see > 100%, recheck your data.

Should I use Kelly on individual trades or a portfolio of trades?

Kelly is typically applied to individual wagers in betting contexts. In trading, you can apply it per-trade or to a portfolio of concurrent positions. Per-trade is simpler; portfolio-level Kelly requires more complex correlation analysis. Start with per-trade Kelly.

How often should I recompute Kelly?

Quarterly or semi-annually. If your system's win rate drifts after 30-50 new trades, recalculate. If it stays consistent, you're fine. Never recalculate after a single lucky or unlucky streak.

Is Kelly better than fixed fractional sizing?

Kelly is more efficient (better long-term wealth) if your edge is stable and measurable. Fixed fractional (e.g., 2%) is simpler and forgiving if your edge estimate is uncertain. Many traders use fixed fractional until they have 100+ trades and can reliably estimate Kelly.

What if I'm trading options or other derivatives with non-linear payoffs?

Kelly still applies, but you must carefully define "odds ratio" as the average payoff per unit risk, accounting for any non-linearity. For spread strategies (e.g., vertical spreads with capped losses), use the capped loss as your "risk" in the Kelly formula.

Related concepts

• Where the Kelly Criterion Comes From
• Half-Kelly: The Practitioner's Choice
• Quarter-Kelly for Conservative Traders
• Fixed Fractional Position Sizing
• What Risk Means

Summary

The Kelly Criterion is the gold standard for position sizing in betting and trading: it calculates the optimal fraction of your bankroll to risk per trade given your win rate and payoff odds. Applied correctly, Kelly maximizes long-term logarithmic wealth faster than any other constant-fraction strategy. But at full Kelly, the ride is volatile—30-40% drawdowns are normal—and many traders abandon the system before the law of large numbers delivers the promised gains.

In practice, professional traders use fractional Kelly: half-Kelly (50% of calculated) or quarter-Kelly (25% of calculated). These fractions preserve 75-90% of Kelly's mathematical advantage while reducing drawdowns to psychologically tolerable levels (15-20% for half-Kelly, 8-12% for quarter-Kelly). The practical recipe is to backtest at least 50-100 trades, calculate full Kelly, apply a 20% haircut for real-world friction (slippage, commissions, regime change), then use half or quarter of the result.

Kelly is not a magic formula; it's a guide calibrated by data. Misestimate your win rate by 5 percentage points, and Kelly will ruin you. But in a trader with a proven, documented edge and the discipline to stick to the plan through drawdowns, Kelly is the closest thing to a free lunch in finance.

Next

→ Where the Kelly Criterion Comes From