The Kelly Criterion tells you the exact fraction of your capital to bet on each trade to maximize long-run growth. Bet more than Kelly and you risk ruin. Bet less and you leave money on the table. It's the mathematical backbone of position sizing — and the reason most traders are either overbetting catastrophically or underbetting unnecessarily.
The Kelly Criterion is a formula that calculates the optimal fraction of your capital to risk on a bet with a known edge. It was developed by John Kelly at Bell Labs in 1956 and has since become the foundation of position sizing theory in both gambling and trading.
The basic formula for a simple win/lose bet:
f* = (bp − q) / b
Where:
Example: You have a trading strategy with a 55% win rate (p = 0.55) and a 1.5:1 reward-to-risk ratio (b = 1.5). Average win = $1,500, average loss = $1,000.
f* = (1.5 × 0.55 − 0.45) / 1.5 = (0.825 − 0.45) / 1.5 = 0.25 or 25%
Kelly says to risk 25% of your capital per trade to maximize long-run geometric growth. This is where the formula connects directly to Vince's optimal f framework — both are solving for the same peak on the growth curve.
Kelly is the peak of the Terminal Wealth Relative (TWR) curve. Below Kelly, your capital grows — but slower than optimal. Above Kelly, your capital growth *decreases* and eventually turns negative.
Vince's canonical example — a 50/50 coin flip winning $2 or losing $1:
| Fraction Bet | TWR After 40 Bets |
|---|---|
| 10% | 4.66× |
| 25% (Kelly) | 10.55× |
| 40% | 4.66× |
| 50% | 1.00× (breakeven) |
| 60% | Ruin |
The asymmetry is the critical insight: overbetting by the same amount as underbetting produces worse results. Betting 40% (15% over Kelly) gives you the same result as betting 10% (15% under Kelly). And anything past 50% leads to ruin.
This is why professional traders use fractional Kelly — typically half-Kelly or quarter-Kelly — to stay safely on the left side of the curve.
Full Kelly maximizes *long-run geometric growth*. But the path to that optimal growth is terrifying:
Most practitioners use half-Kelly (f*/2):
For the example above: 25% / 2 = 12.5% risk per trade
Half-Kelly sacrifices approximately 25% of the optimal growth rate but reduces drawdowns by roughly 50%. The Handbook of Portfolio Mathematics demonstrates this through the geometric mean framework: halving the fraction barely hurts the geometric mean (because the growth curve is relatively flat near the peak) while dramatically improving the equity curve's smoothness.
Quarter-Kelly (f*/4 = 6.25%) is even more conservative and is common among professional fund managers who prioritize capital preservation.
Real trading isn't a simple win/lose bet. To apply Kelly:
1. Estimate your win rate from backtested data (minimum 200+ trades for statistical reliability).
2. Estimate your reward-to-risk ratio (average winning trade / average losing trade).
3. Calculate Kelly fraction using f* = (bp − q) / b.
4. Apply a reduction — use half-Kelly or quarter-Kelly.
5. Translate to position sizing: If half-Kelly says 12.5%, and your stop-loss is 3% from entry, your position size = (12.5% × Account) / (3% × Position Notional).
Kelly and the risk of ruin formula converge on the same truth from different angles. Kelly tells you the optimal bet. The risk of ruin formula tells you the probability of going broke at any given bet size. At full Kelly, the risk of ruin is technically zero over infinite bets (the geometric growth is positive). Below Kelly, it's also zero. Above Kelly, it becomes positive — and past 2× Kelly, ruin is virtually certain.
It provides a mathematical ceiling. Whatever sizing method you use, Kelly is the upper bound. If your method recommends risking more than Kelly, you're in the overbetting zone. This single check prevents the most catastrophic sizing errors.
It quantifies the cost of caution. Half-Kelly gives up ~25% of growth to halve drawdowns. Quarter-Kelly gives up ~44% of growth to quarter drawdowns. You can make an informed choice about where on this spectrum you're comfortable.
It forces you to know your edge. You can't calculate Kelly without knowing your win rate and reward-to-risk ratio. If you don't know these numbers, you can't size properly — and Kelly makes that gap obvious.
Using Kelly with unreliable statistics. If your backtest has 30 trades, your win rate estimate is noisy. Kelly applied to noisy estimates produces dangerous sizing. Require at least 200 trades (ideally 500+) before trusting Kelly calculations. Use confidence intervals, not point estimates.
Ignoring correlation between trades. Kelly assumes independent bets. If you take 5 correlated crypto positions, each at half-Kelly, your effective portfolio bet may be well above Kelly. Adjust per-position Kelly downward when positions are correlated.
Forgetting that Kelly is about growth, not returns. Kelly maximizes geometric growth rate — the compound growth of your account over many bets. It does not maximize expected return per trade or minimize risk. It finds the balance point. If your goal is capital preservation rather than growth, use a smaller fraction.
They're mathematically related. Kelly provides a closed-form solution for simple bets (known odds). Optimal f (Vince) generalizes to any distribution of trade outcomes — you numerically search for the fraction that maximizes TWR given your actual historical trade results. For simple cases, they converge to the same answer. For complex, multi-outcome trading distributions, optimal f is more general.
Yes. As you accumulate more trades, your estimates of win rate and reward-to-risk improve. Recalculate Kelly quarterly or after every 100 trades. If your edge is shrinking (win rate declining or reward-to-risk compressing), Kelly will automatically reduce your sizing.
Yes, with adjustments. Use half-Kelly or less (crypto's fat tails and regime changes make full Kelly too aggressive). Account for correlation between crypto positions (most move together during sell-offs). And use rolling windows for your statistics rather than all-time averages — the market you're trading today may have different characteristics than 2 years ago.
---
*This article is part of The Codex — PARAGON's structured learning library.*