Kelly Criterion for Position Sizing: The Math Behind Optimal Bets

Born in a 1956 Bell Labs paper about information theory, the Kelly Criterion has become one of the most important formulas in quantitative finance. It answers a deceptively simple question: given an edge, how much should you bet? The answer has profound implications for position sizing, portfolio construction, and long-term wealth accumulation.

1. Origins: John L. Kelly Jr. and Bell Labs

In 1956, John Larry Kelly Jr., a physicist at Bell Telephone Laboratories in Murray Hill, New Jersey, published a paper titled “A New Interpretation of Information Rate” in the Bell System Technical Journal, Vol. 35, No. 4, pp. 917–926. Kelly was not thinking about gambling or investing. He was working on a problem in information theory: how to maximize the rate of information transmission over a noisy communication channel.

Kelly’s insight was to frame the problem in terms of a gambler who receives tips over a noisy wire. The gambler knows the tips are correct with some probability p and incorrect with probability q = 1 - p. The question: what fraction of his bankroll should the gambler wager on each race to maximize his long-term growth rate? Kelly showed that the answer was a specific formula that maximized the expected logarithm of wealth — equivalent to maximizing the geometric growth rate of capital over time.

The connection between information theory and gambling was not entirely accidental. Claude Shannon, the father of information theory and Kelly’s colleague at Bell Labs, had independently been thinking about gambling systems and would later apply similar ideas to the stock market alongside Edward Thorp. But it was Kelly who first formalized the relationship between information advantage, optimal bet size, and long-term growth.

2. The Formula: Simple Yet Powerful

In its simplest form — for a binary bet with even payoff odds — the Kelly Criterion states:

f* = 2p - 1

where f* is the optimal fraction of capital to wager and p is the probability of winning. If you have a fair coin (p = 0.5), Kelly says bet nothing. If you have a coin that lands heads 60% of the time at even odds, Kelly says bet 20% of your bankroll on each flip.

The general formula for unequal odds is:

f* = (bp - q) / b

where:

• f* = fraction of capital to bet
• b = net odds received on the bet (e.g., b = 1 for even money, b = 2 for 2-to-1 odds)
• p = probability of winning
• q = probability of losing = 1 - p

Example: A Biased Coin at 2-to-1 Odds

Suppose you have a coin that comes up heads 40% of the time, and someone offers you 2-to-1 odds on heads. Here, p = 0.40, q = 0.60, b = 2. Plugging in:

f* = (2 × 0.40 - 0.60) / 2
f* = (0.80 - 0.60) / 2
f* = 0.20 / 2
f* = 0.10

Kelly says bet 10% of your bankroll on each flip. Despite the coin being biased against you (only 40% heads), the 2-to-1 odds more than compensate, giving you a positive expected value per bet.

The Continuous (Investment) Form

For continuous investment returns — the case most relevant to stock trading — the Kelly fraction takes a different form:

f* = μ / σ²

where μ is the expected excess return (above the risk-free rate) and σ² is the variance of returns. This is sometimes written as f* = (expected return - risk-free rate) / variance. The intuition is clean: you should bet more when your edge is larger and less when returns are more volatile.

For a stock with an expected annual excess return of 12% and annual volatility of 30%, the full Kelly allocation would be:

f* = 0.12 / (0.30)² = 0.12 / 0.09 = 1.33

That is 133% of your capital — Kelly says you should use leverage. This illustrates both the power and the danger of full Kelly sizing.

3. What Kelly Maximizes (and Why It Matters)

The Kelly Criterion maximizes the expected logarithm of wealth, which is equivalent to maximizing the long-term geometric growth rate of capital. This is a crucial distinction from maximizing expected value.

Consider a bet where you double your money with 60% probability and lose everything with 40% probability. The expected value of betting your entire bankroll is positive: 0.60 × 2 + 0.40 × 0 = 1.20, a 20% expected gain. But if you actually bet everything repeatedly, you will eventually go broke with certainty, because a single loss wipes you out. The expected value is positive, but the geometric growth rate is negative infinity.

Kelly sizing resolves this paradox. By betting the optimal fraction (in this case f* = (1 × 0.60 - 0.40) / 1 = 0.20, or 20%), you guarantee the highest possible long-run compound growth rate while never risking total ruin. This is because the log utility function heavily penalizes large losses relative to equivalent gains — losing 50% of your capital requires a 100% gain to recover, not just a 50% gain.

Key Properties of Kelly Sizing

• Maximizes geometric growth rate: No other fixed-fraction strategy will produce faster long-term wealth accumulation.
• Never risks total ruin: Since you only bet a fraction of your capital, you can theoretically never reach zero (though in practice, discrete bet sizes create a small ruin probability).
• Dominance in the long run: Given enough time, a Kelly bettor will almost surely have more wealth than any other bettor using a different fixed-fraction strategy.
• Bet zero when there is no edge: If p ≤ q/b (the bet has negative or zero expected value), the Kelly fraction is zero or negative, meaning you should not bet at all (or bet on the other side).

4. Edward Thorp: From Blackjack to Wall Street

The person most responsible for bringing the Kelly Criterion from information theory to practical use is Edward O. Thorp, a mathematics professor at MIT and later UC Irvine. Thorp first learned of Kelly’s work through Claude Shannon at MIT in the late 1950s.

In 1962, Thorp published Beat the Dealer, which presented a card-counting system for blackjack. The book demonstrated that by keeping track of which cards had been played, a player could identify situations where the deck favored the player over the house. Thorp used the Kelly Criterion to determine optimal bet sizes: bet more when the count indicated a larger edge, bet less (or the minimum) when the house had the advantage. The system worked. Casinos changed their rules in response.

Thorp then turned his attention to financial markets. In 1967, he co-authored Beat the Market with Sheen Kassouf, which described a hedging strategy using warrants and their underlying stocks — essentially a precursor to modern options arbitrage. Thorp later founded Princeton Newport Partners, one of the first quantitative hedge funds, which reportedly generated annualized returns of approximately 19.8% (before fees) over its lifetime from 1969 to 1988, with only three losing months. Thorp used Kelly-based position sizing throughout.

5. The Danger of Full Kelly: Why Practitioners Use Fractional Kelly

In theory, full Kelly sizing is optimal. In practice, it is extremely aggressive and almost no professional trader or fund manager uses it without modification. The reasons are both mathematical and practical.

Volatility of the Kelly Path

Full Kelly sizing produces the highest long-term growth rate, but the path to that growth is brutally volatile. The standard deviation of the log-wealth process under full Kelly is approximately equal to the expected growth rate itself. This means that drawdowns of 50% or more are common and expected. A full Kelly bettor should expect to see their bankroll halved roughly once every few years — even with a genuine, persistent edge.

For most investors, drawdowns of this magnitude are psychologically and practically intolerable. Margin calls, redemption requests, or simple loss of nerve will force a liquidation long before the mathematical long run arrives.

Estimation Error: The Silent Killer

The Kelly formula requires accurate estimates of the probability of winning (p) and the payoff odds (b). In gambling games like blackjack, these can be estimated with precision. In financial markets, they cannot. Estimation error in either parameter directly translates to suboptimal sizing, and the Kelly function is asymmetric: overbetting is far more destructive than underbetting.

If the true Kelly fraction is 20% and you bet 10% (half Kelly), you give up some growth but your bankroll still grows steadily. If you bet 40% (double Kelly), your expected geometric growth rate actually becomes negative — you will go broke in the long run despite having a positive edge. This asymmetry makes conservatism a rational response to uncertainty.

Fractional Kelly in Practice

The standard remedy is fractional Kelly: multiply the full Kelly fraction by a constant between 0 and 1. Common choices in the industry:

The mathematical relationship is: at fraction c of full Kelly, the growth rate is c(2 - c) times the full Kelly growth rate. Half Kelly achieves 0.5 × (2 - 0.5) = 75% of the growth rate with substantially lower variance. Quarter Kelly achieves 0.25 × (2 - 0.25) = 43.75% of the growth rate but with dramatically reduced drawdowns.

6. Kelly vs. Fixed Fractional Sizing

Many traders use fixed fractional sizing: risk a constant percentage of capital per trade regardless of the trade’s expected edge. For example, risk 1% of capital per trade, always. This approach is simple, intuitive, and widely recommended in trading education.

Kelly sizing is fundamentally different. It varies the bet size based on the magnitude of the edge. A trade with a large expected edge and low variance gets a larger allocation. A trade with a marginal edge and high variance gets a tiny allocation. This is both its strength and its weakness.

Advantages of Kelly over Fixed Fractional

• Optimal growth: Kelly maximizes long-term compound returns. Fixed fractional sizing, unless it happens to coincide with the Kelly fraction, grows wealth more slowly.
• Adaptive: Position sizes adjust to the quality of each opportunity. High-conviction signals get more capital.
• Mathematically grounded: The fraction has a clear derivation and interpretation, not just a rule of thumb.

Advantages of Fixed Fractional over Kelly

• No estimation required: Fixed fractional does not require you to estimate p and b for each trade. In many real-world contexts, these estimates are unreliable.
• Simplicity: One parameter (the risk fraction) versus per-trade edge estimation.
• Robustness: When your edge estimates are wrong — and they usually are — fixed fractional does not compound the error through sizing.

In practice, the best approach is often a hybrid: use Kelly-like logic to adjust sizes based on conviction (so you bet more on strong signals and less on marginal ones), but cap the maximum allocation using a fixed fractional limit (so no single trade can cause catastrophic damage even if your edge estimate is completely wrong).

7. Practical Worked Examples

Example 1: A Simple Trading Edge

You have a trading strategy that wins 55% of the time with an average win of $1,200 and an average loss of $1,000. What is the Kelly bet size?

First, compute the odds: b = average win / average loss = 1200 / 1000 = 1.2. Then p = 0.55, q = 0.45.

f* = (bp - q) / b
f* = (1.2 × 0.55 - 0.45) / 1.2
f* = (0.66 - 0.45) / 1.2
f* = 0.21 / 1.2
f* = 0.175

Full Kelly says risk 17.5% of capital per trade. At quarter Kelly (0.25 × 0.175 = 0.044), that becomes approximately 4.4% of capital per trade.

Example 2: A High Win-Rate Scalping Strategy

A scalping strategy wins 80% of the time, but the average win is $200 and the average loss is $600. Here b = 200/600 = 0.333, p = 0.80, q = 0.20.

f* = (0.333 × 0.80 - 0.20) / 0.333
f* = (0.267 - 0.20) / 0.333
f* = 0.067 / 0.333
f* = 0.20

Full Kelly: 20%. Quarter Kelly: 5%. Despite the high win rate, the poor risk-reward ratio keeps the Kelly fraction modest.

Example 3: When Kelly Says Don’t Bet

A strategy wins 40% of the time with average win of $1,000 and average loss of $800. Here b = 1000/800 = 1.25, p = 0.40, q = 0.60.

f* = (1.25 × 0.40 - 0.60) / 1.25
f* = (0.50 - 0.60) / 1.25
f* = -0.10 / 1.25
f* = -0.08

The Kelly fraction is negative. This means the bet has negative expected value. Kelly says do not take this trade (or, if you can, bet the other side). No amount of position sizing can turn a negative-edge bet into a winning strategy.

8. Multi-Asset Kelly: Portfolio Applications

The original Kelly formula applies to a single bet in isolation. For a portfolio of multiple simultaneous positions, the math extends to the multi-asset Kelly framework. The optimal allocation vector is:

f* = Σ² μ

where Σ is the covariance matrix of asset returns and μ is the vector of expected excess returns. This is closely related to mean-variance optimization (Markowitz, 1952), but with the specific objective of maximizing log utility rather than minimizing variance for a given return.

The practical implications are significant. Correlated positions effectively increase the total bet size, so the Kelly framework naturally penalizes concentrated portfolios with correlated holdings. If two stocks have a correlation of 0.9, Kelly treats them almost as a single bet and reduces the combined allocation accordingly.

9. Critiques and Limitations

The Kelly Criterion is not universally accepted. Several prominent investors and academics have raised substantive objections.

Paul Samuelson’s Objection

The Nobel laureate economist Paul Samuelson was a vocal critic of the Kelly Criterion. In his 1971 paper and subsequent writings, Samuelson argued that maximizing the geometric growth rate was not the only rational objective. Different utility functions lead to different optimal strategies, and there is no a priori reason to prefer log utility over other risk preferences. Samuelson famously expressed this objection in a 1979 paper written entirely in one-syllable words to ensure it could not be misunderstood.

Estimation Error (Again)

The practical importance of estimation error cannot be overstated. In financial markets, the expected return and variance of any strategy are estimated with substantial uncertainty. Small errors in these estimates produce large errors in the Kelly fraction. Since overbetting is catastrophic and underbetting is merely suboptimal, rational traders should use fractional Kelly as a matter of course.

Non-Ergodic Returns

The Kelly Criterion assumes you can bet continuously and that your capital is infinitely divisible. In reality, minimum position sizes, transaction costs, and margin requirements create discrete constraints. For small accounts, the Kelly fraction for a given trade might round to zero.

10. How Alpha Suite Uses Kelly

Alpha Suite uses the Kelly Criterion as a core component of its position sizing engine, but with several practical modifications designed for real-world trading conditions.

The system’s default configuration uses KELLY_FRACTION=0.25 — quarter Kelly. This conservative choice reflects the inherent uncertainty in estimating edge from insider trading signals. Even when multiple insiders are buying a stock in a cluster pattern, the actual probability of a profitable trade and the expected magnitude of returns remain estimates, not known quantities.

The sizing pipeline works as follows: each signal receives a conviction score based on insider breadth, cluster intensity, dollar conviction, and time-weighted value. This conviction score, combined with technical overlays (volume-confirmed breakout, momentum, relative strength) and a first-passage probability model, produces an expected edge estimate. The Kelly formula converts this edge into a position size, which is then scaled by the quarter-Kelly fraction and further constrained by a maximum risk-per-trade limit (default 1% of account equity via MAX_RISK_PER_TRADE=0.010).

The result is a sizing system that allocates more capital to high-conviction signals — a cluster of five insiders buying with strong technicals gets a larger position than a single insider buy with ambiguous momentum — while never exposing the portfolio to the catastrophic drawdowns that full Kelly sizing would produce.

11. Summary: When to Use Kelly (and When Not To)

The Kelly Criterion is one of the most elegant results in applied mathematics. It provides a principled, theoretically optimal answer to the position sizing problem. But its practical value depends entirely on the quality of your edge estimates.

Use Kelly (fractional) when: you have a quantifiable, repeatable edge; you can estimate win probability and payoff ratio with reasonable accuracy; you are optimizing for long-term compound growth; and you are disciplined enough to stick with the system through inevitable drawdowns.

Use fixed fractional or simpler methods when: your edge is poorly defined or hard to quantify; you are trading discretionarily rather than systematically; your psychological tolerance for drawdowns is low; or you are primarily concerned with short-term risk management rather than long-term growth.

In most cases, the right answer is a pragmatic blend: Kelly-informed sizing with conservative fractions (quarter to half Kelly), hard position limits, and portfolio-level risk constraints. This captures the core insight of Kelly — bet proportionally to your edge — without requiring the unrealistic precision that full Kelly demands.