Kelly Criterion for Trading: Optimal Position Sizing Explained
April 4, 2026 · By Ashim Nandi
The Kelly criterion is a formula for calculating the position size that maximizes long-term portfolio growth. It balances the tradeoff between betting too much (risking ruin) and betting too little (leaving returns on the table). In practice, no professional trader uses full Kelly. Most use quarter to half Kelly because the theoretical optimum produces drawdowns that are psychologically and financially unbearable.
What the Kelly Criterion Is
In 1956, John Kelly at Bell Labs discovered something profound while working on information theory. The formula for maximizing long-term growth in a repeated betting game is mathematically identical to maximizing information transmission through a noisy channel. It is not an analogy. It is the same equation.
For traders, the core insight is this: there exists a mathematically optimal fraction of your capital to risk on each bet, given your edge and the odds. Bet more than this fraction and you risk ruin. Bet less and you grow slower than necessary. Kelly finds the exact balance point.
The Formula
For simple bets with binary outcomes:
Kelly % = W - (L / R)
Where:
- W = Win probability
- L = Loss probability (1 - W)
- R = Win/loss ratio (average win / average loss)
For stock and portfolio applications, the continuous form is more appropriate:
Kelly % = (Expected Return - Risk-Free Rate) / Variance
Or equivalently:
f = (mu - r) / sigma^2*
Where mu is expected return, r is the risk-free rate, and sigma is standard deviation (volatility).
The key relationship: position size increases with edge but decreases quadratically with volatility. If volatility doubles, you cut position size by 75%, not 50%. Volatility punishes far more aggressively than most traders realize.
Worked Example: The 60% Coin
You have a 60% win rate and 2:1 reward-to-risk ratio.
Kelly % = 0.60 - (0.40 / 2.0) = 0.60 - 0.20 = 0.40 (40%)
Full Kelly says risk 40% of your capital per trade.
That sounds aggressive because it is. A string of four losses at 40% risk per trade would draw your account down by roughly 87%. Mathematically optimal for long-run growth. Practically unsurvivable for most traders.
The Microsoft Example
Your backtest shows Microsoft has a 15% expected annual return. The risk-free rate is 5%. Microsoft's annualized volatility is 25%.
Kelly % = (0.15 - 0.05) / (0.25)^2 = 0.10 / 0.0625 = 1.60 (160%)
Full Kelly wants you leveraged 1.6x. 160% of your capital in a single stock.
Can you handle a 50% drawdown? Because that is what full Kelly produces at this allocation. The math is correct. The human on the other side of the math is the problem.
Full Kelly vs. Fractional Kelly
This is why every professional uses fractional Kelly.
| Approach | Fraction | Drawdown Profile | Who Uses It |
|---|---|---|---|
| Full Kelly | 100% | Maximum growth, maximum drawdown (50%+) | Almost nobody |
| Half Kelly | 50% | ~75% of full Kelly growth, far lower drawdowns | George Soros |
| Quarter Kelly | 25% | ~50% of full Kelly growth, manageable drawdowns | Warren Buffett |
| Eighth Kelly | 12.5% | Conservative growth, minimal drawdowns | Risk-averse institutions |
The math behind fractional Kelly is elegant. Half Kelly achieves roughly 75% of the growth rate of full Kelly while cutting drawdowns dramatically. Quarter Kelly achieves about half the growth rate with drawdowns most traders can actually endure.
Warren Buffett has stated he would invest up to 40% of a portfolio in a single security, but only "under conditions coupling extremely high probability our facts and reasoning are correct with very low probability anything could drastically change underlying value." Then he added: "We're obviously only going to 40% in very rare situations." That 40% ceiling, applied to concentrated value positions, aligns with roughly quarter Kelly.
George Soros famously leveraged to 200% of fund value when shorting the British pound in 1992. That trade had five independent edges stacking up: economic fundamentals, political intelligence, technical setup, reflexive influence, and asymmetric payoff. The practical risk was under 1%. Half Kelly on a near-certain outcome with massive payoff asymmetry. That trade comes around once every 20 years.
Applying the Microsoft Example with Fractional Kelly
Starting from the full Kelly of 160%:
| Fraction | Allocation | Position in $100K Portfolio |
|---|---|---|
| Full Kelly | 160% | $160,000 (leveraged) |
| Half Kelly | 80% | $80,000 |
| Quarter Kelly | 40% | $40,000 |
| Eighth Kelly | 20% | $20,000 |
Quarter Kelly on the Microsoft example produces a $40,000 position in a $100,000 portfolio. Still aggressive for a single stock, but within the range of concentrated portfolio managers.
When NOT to Use Kelly
Kelly is powerful but fragile. It breaks in specific, predictable ways.
1. When Your Edge Estimate Is Wrong
Kelly is only as good as your inputs. If you estimate a 60% win rate but your true win rate is 52%, full Kelly will oversize dramatically. Overestimating edge is the fastest path to ruin with Kelly. This is the primary reason professionals use fractional Kelly. The fraction acts as a buffer against estimation error.
The probability and statistics feedback loop exists precisely to calibrate these inputs. Without statistically validated win rates across sufficient sample sizes (200+ trades), Kelly inputs are guesses. And Kelly with guessed inputs is dangerous.
2. When Returns Are Not Independent
Kelly assumes each trade is independent of the last. In reality, market returns are correlated. Losses cluster. Winning streaks cluster. Market regimes create extended periods where your strategy's win rate differs significantly from the long-run average.
A strategy with a 60% overall win rate might have a 75% win rate in trending regimes and a 40% win rate in choppy regimes. Applying Kelly based on the overall 60% will oversize in choppy regimes and undersize in trending ones.
3. When Fat Tails Exist
Kelly assumes a known distribution of outcomes. Markets have fat tails. Extreme moves happen more often than models predict. A full Kelly position that looks optimal under normal distribution assumptions can produce catastrophic losses when a tail event strikes.
4. When You Cannot Tolerate the Drawdowns
Full Kelly produces theoretical maximum drawdowns of 50% or more. For most traders and most institutions, a 50% drawdown triggers career risk, margin calls, or psychological collapse long before the law of large numbers rescues the account. If you cannot sit through the drawdown, the optimal fraction is lower than Kelly suggests.
5. When Transaction Costs Are Significant
Kelly does not account for commissions, slippage, or market impact. In markets with high friction, the effective edge is smaller than the gross edge. Kelly should be calculated on net expected return, after costs.
A Simpler Alternative: Risk-Based Position Sizing
Most traders do not need Kelly at all. They need something simpler.
Position Size = (Account Size x Risk %) / Stop Distance
Example: $50,000 account. Buying Nvidia at $140 with a stop at $135. That is $5 risk per share.
1% of $50,000 = $500. Divide by $5 per share = 100 shares maximum.
No probability estimates required. No complex formulas. Just account size, risk percentage, and stop distance. This approach, combined with proper risk management and portfolio heat limits, keeps most traders alive long enough to build the track record needed to calculate Kelly inputs properly.
The practical path: start with risk-based sizing. Track 200+ trades. Calculate your actual win rate and payoff ratio. Then graduate to fractional Kelly with statistically validated inputs. ATOM supports both approaches, starting with simple risk-based sizing for newer traders and offering Kelly-based calculations once sufficient trade history exists to generate reliable inputs.
ATOM's Kelly Calculator
ATOM computes Kelly fractions using your actual trade data rather than assumed probabilities. This matters because Kelly is only useful when the inputs are accurate.
The process:
- Measure actual win rate and average win/loss ratio from your trade log (minimum 200 trades)
- Segment by market regime, so Kelly adjustments reflect current conditions
- Apply a fractional multiplier (default: quarter Kelly) as a buffer against estimation error
- Integrate with portfolio heat limits, so Kelly-suggested sizing never violates overall portfolio risk constraints
The result is position sizing that adapts to your measured edge rather than assumed edge, while remaining conservative enough to survive the estimation errors that make full Kelly impractical.
FAQ
Is full Kelly ever appropriate for retail traders?
Almost never. Full Kelly assumes perfect knowledge of your win rate and payoff ratio, independent returns, and the psychological ability to endure 50%+ drawdowns. None of these conditions hold for most retail traders. Quarter Kelly or simpler risk-based sizing (1-2% risk per trade) is more appropriate for anyone who has not validated their edge across hundreds of trades in multiple market regimes.
What happens if I bet more than Kelly suggests?
Overbetting relative to Kelly reduces long-term growth rather than increasing it. This is counterintuitive but mathematically proven. At 2x Kelly, your expected growth rate drops to zero. Beyond 2x Kelly, you are mathematically guaranteed to go broke given enough time. The growth curve is not linear. It peaks at Kelly and declines symmetrically on either side.
How do Buffett and Soros size their positions relative to Kelly?
Buffett operates at roughly quarter Kelly, concentrating in high-conviction positions but with a strict 40% ceiling per security and only under conditions of extreme certainty. Soros operates closer to half Kelly in normal conditions, with rare exceptions (like the 1992 pound trade) where stacked independent edges justified extreme concentration. Both are far below full Kelly in their standard operations.
Can I use Kelly for crypto or highly volatile assets?
You can, but the quadratic penalty on volatility makes it critical. If a stock has 25% annualized volatility and a crypto asset has 75% volatility, Kelly will suggest a position roughly 9x smaller in the crypto asset (all else equal), because volatility enters the denominator as a squared term. This is the formula correctly reflecting the reality that volatile assets need dramatically smaller positions.