AI Kelly Position Sizing: Bet the Right Amount

Position sizing is often more important than entry timing. A good strategy with poor sizing can blow up. A mediocre strategy with excellent sizing can survive and compound. The Kelly criterion is one of the most famous frameworks for optimal bet sizing.

The Kelly formula was originally developed for information theory and gambling. In trading, it tells you how much of your capital to allocate to a bet with known probabilities and payoffs. The catch is that those probabilities are never truly known in markets.

The Kelly Formula

For a simple bet with two outcomes:

f* = (p * b - q) / b

Where:

• f* = fraction of capital to bet
• p = probability of winning
• q = probability of losing = 1 - p
• b = average win / average loss

Example: If you win 60% of trades with a 2:1 payoff ratio:

p = 0.60
b = 2.0
f = (p * b - (1 - p)) / b
print(f)  # 0.40

Full Kelly says bet 40% of capital. That is almost always too aggressive.

Fractional Kelly

Most traders use fractional Kelly to reduce volatility:

• Half Kelly: 20% in the example above
• Quarter Kelly: 10%
• Many systematic traders use even less

Fractional Kelly sacrifices some theoretical growth for much smoother equity curves. A common rule is to start with quarter Kelly or less until you have high confidence in your edge estimates.

Estimating Inputs With AI

The hard part of Kelly is estimating p and b. Historical averages are often stale. AI can estimate them dynamically:

• Predict win probability from current market features
• Estimate expected payoff from volatility and recent price action
• Adjust for regime changes

For example, a model might predict that under current conditions, a breakout setup has a 55% win rate with a 1.8:1 payoff ratio.

Practical Implementation

def kelly_size(capital, win_prob, payoff_ratio, fraction=0.25):
    kelly = (win_prob * payoff_ratio - (1 - win_prob)) / payoff_ratio
    kelly = max(0, kelly)  # never bet negative
    return capital * kelly * fraction
 
position = kelly_size(100000, 0.55, 1.8, fraction=0.25)

This gives a conservative position size based on the current trade's estimated edge.

Kelly vs Fixed Fractional

Fixed fractional risk such as 1% per trade is simpler and often safer. Kelly is most useful when:

• You have many similar trades
• Edge varies meaningfully across setups
• You can estimate probabilities accurately

If your probability estimates are wrong, Kelly sizing can make losses worse.

A comparison table helps clarify the tradeoff:

Common Mistakes

• Using full Kelly
• Ignoring correlation between simultaneous trades
• Overestimating win rate from backtests
• Applying Kelly to strategies with skewed payoff distributions
• Forgetting that Kelly assumes you can reinvest continuously

When to Use Kelly

Use Kelly-based sizing when:

• You have a large sample of similar trades
• Your edge estimates come from out-of-sample data
• You want to scale positions up when edge is high
• You can handle larger position sizes emotionally

When Not to Use Kelly

Avoid Kelly when:

• You are a beginner
• Your strategy has fat-tailed or skewed returns
• You cannot estimate probabilities reliably
• You hold highly correlated positions

For a simpler alternative, see AI Stop Loss Dynamic ATR.

Combining Kelly With Risk Management

Kelly sizing should not exist in isolation. Pair it with:

• Maximum position limits per trade
• Portfolio heat limits for correlated trades
• Stop losses to define the actual loss amount
• Regular recalculation as market conditions change

For example, you might cap any single position at 5% of capital even if Kelly suggests 8%. This keeps a single mistake from dominating your outcome. The Kelly output is a starting point for discussion, not a final order size.

Kelly in a Multi-Strategy Portfolio

When running multiple strategies, Kelly can help allocate capital across them. Treat each strategy as a bet with its own return distribution. Strategies with higher risk-adjusted returns get more capital. Strategies in drawdown get less.

This is sometimes called growth-optimal portfolio allocation. It requires long track records and stable correlations between strategies.

Bottom Line

The Kelly criterion is a powerful tool for thinking about position size, but it is not a magic formula. Use fractional Kelly, validate your edge estimates out-of-sample, and never let one trade threaten your account.

Related reading: AI Risk Parity Portfolio | AI Stop Loss Dynamic ATR | AI Risk Management Framework