Kelly Criterion & Advanced Position Sizing

25 min read · Intermediate · Last updated August 2026

Most traders obsess over entries — which indicator to use, when to buy, when to sell. Yet the single most important determinant of long-term survival and profitability is how much you risk on each trade. Two traders with identical signal accuracy will produce wildly different equity curves depending on their position sizing. One compounds steadily; the other blows up.

The Kelly Criterion, developed by John L. Kelly Jr. at Bell Labs in 1956, provides the mathematically optimal answer to this question. It tells you exactly what fraction of your capital to risk on each bet to maximise long-term geometric growth. This guide covers the theory, the practical modifications that make it usable for trading, and the portfolio-level extensions that protect against correlated drawdowns.

1. Why Position Sizing Matters More Than Entries

Consider a system with a 60% win rate and 1:1 risk-reward. That’s a genuine edge. But if you risk 50% of your account on each trade, a run of just three consecutive losers — which has a 6.4% probability on any given sequence of three trades — reduces your capital to 12.5% of its starting value. Recovery from that drawdown requires a 700% return. The edge is real; the sizing destroyed it.

The same system with 2% risk per trade survives the same losing streak with a 5.9% drawdown — recoverable in under two weeks of normal trading. The signal didn’t change. The strategy didn’t change. Only the position size changed, and it made the difference between bankruptcy and compounding.

Academic research consistently shows that position sizing explains more variance in long-term trading outcomes than signal accuracy. Ralph Vince’s work on optimal-f, and later the Kelly Criterion’s adoption by hedge funds, formalised what experienced traders already knew: the size of the bet matters more than the quality of the bet.

2. The Kelly Criterion — Mathematical Derivation

The Kelly Criterion maximises the expected logarithm of wealth, which is equivalent to maximising the long-term geometric growth rate of your capital. For a simple binary outcome (win or lose), the formula is:

f* = (p · b − q) / b

Where f* is the optimal fraction of capital to wager, p is the probability of winning, q = 1 − p is the probability of losing, and b is the ratio of the net profit on a win to the net loss on a loss (the odds or reward-to-risk ratio).

For a system with 55% win rate and 1.5:1 reward-to-risk: f* = (0.55 × 1.5 − 0.45) / 1.5 = 0.25. Full Kelly says to risk 25% of your capital on each trade. That sounds aggressive — and it is. Full Kelly produces the highest long-term growth rate, but with drawdowns that would make most traders physically ill. The maximum drawdown under full Kelly sizing can exceed 80% before recovering.

The key insight is that Kelly optimises for terminal wealth, not for comfort. It assumes infinite time horizon, perfect knowledge of probabilities, and no psychological constraints. None of these hold in practice.

3. Why Half-Kelly Is the Practical Standard

Professional traders and quantitative funds almost universally use half-Kelly or less. Ed Thorp, who pioneered the use of Kelly in both blackjack and the stock market, advocated for half-Kelly or even quarter-Kelly in his fund Princeton Newport Partners.

The mathematics support this conservatism. Half-Kelly achieves 75% of full Kelly’s geometric growth rate while cutting the expected maximum drawdown roughly in half. The growth-to-drawdown ratio actually improves. You give up 25% of the upside to avoid 50% of the downside — a trade any rational investor should accept.

There’s a deeper reason: Kelly assumes you know your true win rate and payoff ratio with certainty. In trading, you don’t. Your estimated edge contains estimation error, and over-betting relative to your true edge is far more destructive than under-betting. If your real win rate is 53% but you think it’s 60%, full Kelly will over-size systematically. Half-Kelly builds in a margin of safety against estimation error.

The AI Trading Copilot Approach

The Risk Manager agent in the AI Trading Copilot enforces half-Kelly as the default sizing methodology. It takes the probability score from the Trader agent’s consensus (which serves as the win probability estimate) and the risk-reward ratio from the trade plan, then applies the half-Kelly formula to determine position size as a percentage of available capital.

4. Drawdown Mathematics and Circuit Breakers

Understanding drawdown mathematics is essential for setting appropriate risk limits. The probability of experiencing a drawdown of magnitude D with a win rate p and risk fraction f follows a geometric distribution. For a 55% win rate risking 2% per trade, the probability of a 20% drawdown is approximately 0.8%, which means it will happen roughly once every 125 trading sequences of meaningful length.

More critically, the time to recover from a drawdown scales non-linearly. A 10% drawdown requires an 11.1% gain to recover. A 20% drawdown requires 25%. A 50% drawdown requires 100%. At 75%, you need a 300% return. This asymmetry is why preventing large drawdowns is more important than maximising returns.

The AI Trading Copilot implements a 15% drawdown circuit breaker. If the paper trading portfolio drops 15% from its high-water mark, the system halts all new signals until the trader manually reviews and resets. This is not a suggestion — it’s a hard stop enforced by the Risk Gate, the final stage of the signal pipeline.

Daily and Weekly Loss Limits

Beyond the portfolio-level circuit breaker, professional risk management includes daily loss limits (typically 2–3% of equity) and weekly loss limits (typically 5–6% of equity). These prevent the compounding effect of tilt — the tendency to increase risk after losses in an attempt to recover quickly. The Kelly framework provides the per-trade sizing; the loss limits provide the temporal constraints.

5. Correlation-Adjusted Position Sizing

Standard Kelly treats each bet as independent. In portfolio trading, positions are rarely independent. If you hold long positions in AAPL, MSFT, and GOOGL, a broad tech selloff hits all three simultaneously. The effective risk is much higher than the sum of individual position risks suggests.

Correlation-adjusted sizing accounts for this by reducing individual position sizes when the portfolio already holds correlated exposure. The adjustment uses the average pairwise correlation between existing positions and the proposed new position. If the correlation exceeds a threshold (commonly 0.6), the new position size is scaled down proportionally.

The formula modifies the Kelly fraction: f_adjusted = f_kelly / (1 + avg_correlation × n_correlated), where n_correlated is the number of existing positions with correlation above the threshold. This ensures that portfolio-level risk remains bounded even when individual positions appear small.

The Correlation agent in the AI Trading Copilot computes real-time pairwise correlations across all positions and feeds this data to the Risk Manager. When a new signal triggers for an asset that is highly correlated with existing exposure, the position size is automatically reduced — or the signal is vetoed entirely if adding it would breach the portfolio correlation limit.

6. Portfolio-Level Risk Controls

Individual trade sizing is necessary but not sufficient. Portfolio-level controls address systemic risks that per-trade rules miss:

• Maximum total exposure: No more than 20–30% of capital deployed at any time across all positions. This leaves 70%+ in cash as a buffer against correlated adverse moves.
• Sector concentration limits: No more than 10% exposure to a single sector. Prevents a sector-specific event from causing outsized damage.
• Single-name limits: No individual position exceeding 5% of portfolio equity, regardless of what Kelly suggests. This caps idiosyncratic risk.
• Directional balance: Net long or short exposure should not exceed 15% unless the Regime agent explicitly confirms a trending regime. In unclear regimes, the portfolio stays close to neutral.
• Volatility scaling: When realised volatility exceeds its 90th percentile, all position sizes are halved automatically. High-volatility environments amplify all risks.

These constraints operate as a layered defense system. A single trade might pass Kelly sizing, but if adding it would breach a portfolio-level limit, the Risk Gate vetoes it. The philosophy is that no single constraint is perfect, but multiple overlapping constraints create robust protection.

7. Implementing Kelly in a Live Trading System

Moving from theory to practice requires several adaptations. First, your win rate and payoff ratio estimates must come from out-of-sample data, not in-sample backtest results. Using optimised backtest statistics to size real trades is a recipe for over-betting.

Second, update your estimates using a rolling window rather than all-time statistics. Market conditions change, and your edge today may differ from your edge last year. A 60-day rolling win rate provides a responsive estimate without over-reacting to short-term variance.

Third, implement Kelly fraction capping. Even if the formula suggests risking 8% per trade, cap the maximum at 3–4% to protect against estimation error and tail events. The AI Trading Copilot caps half-Kelly sizing at 3% per position, with an override requiring explicit manual confirmation for any size above 2%.

Finally, maintain a position sizing journal. Record the Kelly fraction suggested, the actual size taken, and the outcome. Over time, this reveals whether your sizing is improving your risk-adjusted returns or whether your estimates need recalibration.

Position sizing is where mathematics meets discipline. The Kelly Criterion gives you the optimal framework; half-Kelly gives you the practical implementation; and portfolio-level controls ensure that no single mistake can be fatal. Master these, and the quality of your entries becomes a secondary concern.