The Kelly Criterion: The Mathematics of Optimal Betting

Imagine you have a coin that lands heads 60% of the time. A casino offers you even-money bets: bet a dollar, win a dollar. You have $1,000 in your pocket and unlimited time to play. How much should you bet on each flip?

Most people's instincts here are badly calibrated. Some say bet big — you have the edge, press it. Others say bet small — you don't want to risk ruin. A few say bet everything, since you'll almost certainly win in the long run. All of these intuitions miss a deep mathematical truth that wasn't discovered until 1956, when a physicist at Bell Laboratories named John Kelly Jr. solved the problem completely. His answer was exact: bet exactly 20% of your bankroll on every flip.

Not 10%, not 50%, not "it depends." Twenty percent. And if your bankroll grows or shrinks, your next bet grows or shrinks with it. This is the Kelly Criterion, and it turns out to be one of the most consequential formulas in the history of money.

The Concept

The Kelly Criterion is a formula for choosing bet sizes that maximizes your long-run wealth growth. It was first published in July 1956 in the Bell System Technical Journal under the title "A New Interpretation of Information Rate" — a deliberately cryptic name, because AT&T's executives had objected to the word "gambling" appearing in a Bell Labs publication. The paper's original working title had been "Information Theory and Gambling," which was far more accurate.

The formula itself is elegant:

f = (bp - q) / b

Where: - f is the fraction of your bankroll to bet - p is the probability of winning - q is the probability of losing (which equals 1 - p) - b is the net odds you receive (on an even-money bet, b = 1)

For an even-money bet, this simplifies beautifully to f = p - q, or equivalently f = 2p - 1. Plug in our 60% coin: f = 0.6 - 0.4 = 0.20. Bet 20% of whatever you currently have, every single time.

There's a deeper way to express the same idea: Kelly = Edge / Odds. Your "edge" is how much you expect to profit per dollar bet. Your "odds" is what you get back per dollar bet. Divide one by the other, and you have the optimal fraction.

The reason this formula is correct — and not just a useful heuristic — comes from what it maximizes. Kelly didn't set out to maximize expected profit. He set out to maximize the expected logarithm of wealth, which turns out to be equivalent to maximizing the long-run geometric growth rate of your bankroll. These are very different objectives. Expected profit would tell you to bet everything on every flip, since betting $1,000 gives you the highest expected dollar return. But betting everything means one loss — which will happen 40% of the time — wipes you out entirely. You never recover from zero.

The geometric growth rate accounts for this asymmetry. Losses hurt more than equivalent gains help. Lose 50%, you need a 100% gain to get back to where you started. Kelly's formula is the unique bet size that threads this needle perfectly: aggressive enough to compound wealth rapidly, conservative enough to avoid catastrophic losses.

Why It Matters

The Bell Labs Connection

John Larry Kelly Jr. was born on December 26, 1923, in Corsicana, Texas, and died tragically young — a stroke at 41, on March 18, 1965 — before the world had any idea what he'd set in motion. He was a World War II Navy pilot with a PhD in physics from the University of Texas at Austin, and he'd joined Bell Laboratories, the legendary research institution in New Jersey where Claude Shannon had recently invented information theory.

Shannon's 1948 paper "A Mathematical Theory of Communication" had established the mathematical foundations of how information can be reliably transmitted over noisy channels. His key insight: there's a fundamental limit, called channel capacity, on how much information can flow through any channel, regardless of how clever your encoding scheme is.

Kelly noticed something startling. A gambler receiving tips about bet outcomes through a noisy communication channel is mathematically identical to an engineer decoding signals through a noisy wire. The maximum achievable growth rate in Kelly's gambling formula is precisely equal to Shannon's channel capacity. The "value" of a betting edge, measured in bits of information, is the same quantity Shannon used to measure signal quality.

This is not an analogy or an approximation — it's a mathematical identity. Optimal betting is optimal information decoding. A completely fair gamble (p = 0.5) has zero channel capacity and zero Kelly growth rate; you cannot extract wealth from pure noise. Every bit of genuine edge translates, with perfect mathematical precision, into a corresponding rate of wealth growth.

Shannon, who worked alongside Kelly at Bell Labs and is widely regarded as one of the greatest intellects of the 20th century, recognized the formula's importance and encouraged Kelly to publish it.

Ed Thorp Takes It to the Casino

Kelly died not knowing his formula would change Wall Street. That transformation was engineered by Edward O. Thorp, a mathematics professor who learned of Kelly's work in November 1960 from Shannon himself.

Thorp had been analyzing blackjack using an IBM 704 computer, looking for card-counting strategies that could give a player a genuine edge over the house. When Shannon handed him Kelly's paper, Thorp immediately recognized it as the missing piece: now he knew not just when to bet more (when the deck was favorable) but exactly how much more to bet.

With investor Manny Kimmel, Thorp tested the strategy in Nevada casinos, winning $11,000 in a single weekend in Reno and Lake Tahoe. His 1962 book Beat the Dealer, which explicitly named "the Kelly gambling system," became a bestseller with over 700,000 copies sold. For the first time, a mathematically provable strategy for beating a casino was in the public's hands.

Thorp didn't stop there. In 1969 he founded a hedge fund, Princeton/Newport Partners, applying Kelly-optimal position sizing to convertible bond arbitrage and derivatives. The results over nineteen years were almost incomprehensible: 19.1% gross annualized returns (versus the S&P 500's 10.2% over the same period), not a single losing year in nineteen years of operation, and only three down months in the entire run — the worst under 1%. The fund grew from $1.4 million to $273 million.

This is arguably the most compelling real-world validation of a single mathematical formula in financial history.

The Details

The Biased Coin, Worked Through

Let's return to our original example and trace through what the Kelly approach actually looks like in practice.

You have a 60/40 coin, even-money payoffs, $1,000 starting bankroll. Kelly says bet 20%.

Flip 1: Bet $200. Win → bankroll is now $1,200, next bet is $240. Lose → bankroll is $800, next bet is $160.

The key is that the bet tracks the bankroll. You never bet a fixed dollar amount; you always bet the same fraction. This automatic adjustment is what gives the Kelly strategy its remarkable properties. After a win, you bet more — compounding aggressively. After a loss, you bet less — preserving capital. The strategy is inherently self-protective.

The long-run geometric growth rate for this bet is approximately 2% per flip. That sounds modest, but with compounding, starting from $1,000 and flipping 500 times, you'd expect a bankroll around $20,000. Starting with $10,000 and flipping 1,000 times? Around $4 million. The magic of compound growth, harvested at the mathematically maximum rate.

Now consider what happens if you deviate. Bet 50% instead of 20% — it feels more aggressive, and your expected dollar profit per flip is higher. But the geometric growth rate actually decreases, because the variance destroys compounding. Bet 100% of your bankroll every time, and the very first losing flip wipes you out completely; no recovery possible. The Kelly fraction is not arbitrary — it's the exact peak of a growth curve that rises steeply from zero and then falls back to zero at 100%.

The Over-Betting Problem

There's an asymmetry in the Kelly curve that has profound practical implications: betting more than Kelly is worse than betting less than Kelly by the same amount. If the true Kelly fraction is 20%, betting 30% is worse than betting 10% — even though both are equally far from the optimum.

Why? Because losses are multiplicative. When you over-bet and lose, the damage to your compounding base is disproportionately large. When you under-bet and lose, you've merely sacrificed some growth. One path can lead to ruin; the other just leads to slower wealth accumulation.

This asymmetry has a critical implication: you must know your edge precisely. If your edge is smaller than you think — if p is 55% when you thought it was 65% — your Kelly fraction is also smaller than you think. Betting what you believe is Kelly will actually be over-betting, with all the associated risks. Estimation error is not symmetric; overconfidence is more dangerous than underconfidence.

Fractional Kelly: The Practitioner's Approach

For this reason, essentially every serious practitioner who uses Kelly-based thinking uses a fraction of the Kelly bet, not the full amount.

The mathematics here is surprisingly favorable. Research by MacLean, Ziemba, and Blazenko (1992) quantified the trade-off: half-Kelly captures approximately 75% of the long-run growth rate of full Kelly, while cutting volatility roughly in half. That's a remarkably efficient trade. You give up a quarter of your expected growth in exchange for dramatically smoother outcomes and much more protection against the inevitable errors in your edge estimates.

The probability numbers are stark. A full Kelly bettor has approximately a 1-in-3 chance of seeing their bankroll drop to half before it doubles. A half-Kelly bettor? About 1-in-9. If you genuinely value peace of mind (and most people do, in ways they underestimate), half-Kelly or quarter-Kelly is the rational choice.

Bill Gross, who later became the "Bond King" and co-founded PIMCO managing nearly $1 trillion in assets, experienced this firsthand. After reading Thorp's Beat the Dealer in 1966, Gross went to Las Vegas with $200 and, sizing his bets proportionally to his current count-estimated edge, grew that stake to $10,000 over a summer. He carried the framework into fixed-income investing, reportedly saying that PIMCO was "being run like a professional blackjack table from the standpoint of risk management."

The Stock Market Version

For continuous investments like stocks, the Kelly formula takes a different form:

f = (μ - r) / σ²

Where μ is the expected return of the asset, r is the risk-free rate, and σ² is the variance of returns. This fraction tells you how much of your capital to allocate to a risky asset to maximize long-run growth.

Plug in typical numbers for a stock: expected excess return of 8%, volatility of 20%. The Kelly fraction is 0.08 / 0.04 = 2.0 — meaning Kelly suggests 200% allocation, i.e., levering up 2-to-1. This is one reason full Kelly is rarely applied literally to equity portfolios; the leverage implied by the formula often exceeds what's practical or psychologically sustainable. But the direction of the insight is sound: assets with higher Sharpe ratios deserve larger allocations.

The Deep Philosophical Point

Kelly's criterion connects to a question that's deeper than money: what does it mean to play an optimal long-run strategy in a world of uncertainty?

The Kelly criterion embodies a specific answer: maximize the expected logarithm of wealth, which is equivalent to maximizing the median of the long-run distribution. Over many trials, a Kelly bettor will beat any fixed-fraction bettor with probability approaching 1. This is a theorem, not a conjecture.

But "long run" is doing heavy lifting here. The long run might be very long — hundreds or thousands of bets. Over a finite horizon, or in a world where your life circumstances could change (job loss, medical crisis, family obligation), maximizing geometric growth is not necessarily the right objective. Someone with strong risk aversion or a short time horizon might rationally choose a much smaller fraction than Kelly suggests. The formula answers the question it's designed to answer; whether that's your question depends on your actual situation.

Takeaways

The Kelly Criterion gives the exact fraction of your bankroll to bet on any favorable wager to maximize long-run wealth growth. For an even-money bet with a 60% win probability, that's 20% — not more, not less.

Kelly was discovered at Bell Labs in 1956 as a consequence of information theory. The formula is mathematically equivalent to Shannon's channel capacity: the maximum rate of wealth growth equals the maximum rate of information transmission through a noisy channel.

*Betting more than Kelly is worse than betting less. The curve peaks at the Kelly fraction and falls asymmetrically; overconfidence about your edge is more dangerous than underconfidence.

Most serious practitioners use half-Kelly or less. Half-Kelly captures ~75% of maximum growth with ~50% of the volatility — an efficient trade-off, especially when edge estimates are uncertain.

The formula works anywhere you have a genuine, quantifiable edge: blackjack, sports betting, equity investing, venture capital. The math is the same; the difficulty lies in accurately estimating p and b.

Resources: - Kelly's original 1956 paper: Bell System Technical Journal, vol. 35, p. 917 - Ed Thorp's account of applying Kelly to blackjack and markets: The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market (2006) - William Poundstone's Fortune's Formula* (2005) — the best narrative history of Kelly, Shannon, Thorp, and the formula's journey from Bell Labs to Wall Street