Kelly Criterion

How much of your capital should you risk on a bet you believe is in your favour? The intuitive answer — as much as possible — is mathematically guaranteed to destroy you. The correct answer was derived in 1956 by John Kelly Jr., a physicist at Bell Labs, and it changed the foundations of how rational actors should think about sizing any commitment under uncertainty.

Kelly's formula is deceptively simple. For a binary bet with probability p of winning and probability q = 1 − p of losing, where a win pays b-to-1, the optimal fraction of capital to wager is: f* = (bp − q) / b. The formula can be restated even more compactly: f* = edge / odds, where edge is the expected profit per dollar risked and odds is the payoff on a winning bet. If you have a 60% chance of winning an even-money bet, the Kelly fraction is (1 × 0.6 − 0.4) / 1 = 20% of your capital. Not 100%. Not 50%. Twenty percent.

The number shocks people who encounter it for the first time. A bet where you win 60% of the time feels overwhelmingly favourable — the kind of edge most investors, gamblers, and founders would mortgage their house for. Yet the formula says to risk only a fifth of your bankroll. The reason is the asymmetry between multiplicative gains and multiplicative losses. A 50% loss requires a 100% gain to recover. A 75% loss requires a 300% gain. Betting too large on any single outcome — even one with a genuine, quantifiable edge — exposes you to sequences of losses that compound geometrically toward zero. The Kelly criterion is the precise mathematical boundary between growth and ruin.

What Kelly actually maximised was not expected wealth but the expected logarithm of wealth — the geometric growth rate. This distinction is the entire intellectual content of the model. The arithmetic expected value of a bet tells you what happens on average across a population of bettors at a single moment. The geometric growth rate tells you what happens to a single bettor across many sequential bets over time. In any system where outcomes are multiplicative — where each result changes the base from which the next result is calculated — these two quantities diverge. The arithmetic average overstates the individual's prospects. The geometric growth rate captures the actual compounding trajectory. Kelly's contribution was proving that the bet size which maximises the geometric growth rate is unique, calculable, and always smaller than the bet size which maximises expected value.

The formula emerged from information theory, not gambling. Kelly was working on the problem of signal transmission over noisy channels — building on Claude Shannon's foundational work at Bell Labs. Kelly noticed that a gambler receiving partially reliable information about the outcome of a horse race faces a mathematically identical problem to a communications engineer receiving a partially reliable signal over a noisy line. The optimal strategy in both cases is to allocate resources in proportion to the reliability of the information: bet more when the signal is strong, less when it's weak, and nothing when the signal is pure noise. The connection between information theory and optimal wagering was not a metaphor. It was a mathematical identity.

Ed Thorp — a mathematics professor at MIT who had independently developed card-counting systems for blackjack — recognised the formula's significance immediately. Thorp applied Kelly sizing first at the blackjack table, where his card-counting system provided a quantifiable edge that fluctuated hand by hand, and then in financial markets through his hedge fund Princeton Newport Partners. The results were extraordinary: nineteen consecutive years of positive returns with annualised performance exceeding 20%, achieved not through superior prediction of market direction but through superior sizing of exposure to identified mispricings. The same mispricings were visible to other sophisticated market participants. The difference was that Thorp sized his positions to maximise geometric growth while his competitors sized to maximise expected return — and the competitors, one by one, were eliminated by adverse sequences that their sizing could not survive.

The Kelly criterion's most counterintuitive property is what happens when you exceed it. Betting exactly the Kelly fraction maximises long-run geometric growth. Betting less than Kelly — say, half-Kelly — reduces the growth rate but also reduces the volatility of the path, producing a smoother equity curve with smaller drawdowns. This is why most professional practitioners use a fractional Kelly approach. But betting more than Kelly does something far worse than just reducing growth: it actually decreases the geometric growth rate below what under-betting would produce, while simultaneously increasing volatility and ruin probability. At twice the Kelly fraction, the expected geometric growth rate drops to zero — you are statistically guaranteed to neither grow nor shrink over the long run, while experiencing enormous swings. Above twice Kelly, the geometric growth rate turns negative — you are mathematically certain to go broke. The overbettor destroys wealth faster than the random bettor, despite having a genuine edge, because the sizing converts the edge into a liability.

This is the insight that separates Kelly from every other risk framework: the relationship between bet size and growth rate is not monotonic. There is a peak — the Kelly fraction — and both sides of the peak slope downward, but the right side (overbetting) slopes into ruin while the left side (underbetting) merely slopes into slower growth. The asymmetry means that the penalty for betting too large is categorically worse than the penalty for betting too small. A bettor who uses half the Kelly fraction captures 75% of the maximum growth rate. A bettor who uses double the Kelly fraction captures 0%. The errors are not symmetric, which is why every practitioner who has survived long enough to write about their experience advocates sizing conservatively relative to the theoretical optimum.

The framework extends beyond gambling and investing into any domain where you allocate a scarce, depletable resource to opportunities with uncertain outcomes. A founder deciding how much runway to burn on a product bet. An executive deciding how much of the R&D budget to concentrate on a single initiative. A professional deciding how much career capital to invest in a risky pivot. In each case, the Kelly logic applies: there is an optimal allocation that maximises the long-run growth rate, and that allocation is always smaller than intuition suggests, because intuition is calibrated to expected value rather than geometric growth.

The history of financial blowups is overwhelmingly a history of Kelly violations. Long-Term Capital Management leveraged Nobel Prize–winning models at 25:1 and was destroyed in 1998 by the same convergence trades that Thorp had executed profitably for decades at fraction-of-Kelly sizing. Bear Stearns and Lehman Brothers held mortgage-backed securities with genuine positive expected returns but sized their exposure so far beyond any Kelly-compatible fraction that a correlated adverse move — the 2008 subprime crisis — converted portfolio-level mark-to-market losses into institutional death. In each case, the operators had correctly identified an edge. In each case, the sizing of their commitment to that edge was the variable that determined whether the edge built wealth or destroyed the institution. The edge was right. The fraction was wrong. And in multiplicative systems, a wrong fraction overwhelms a right edge with mathematical certainty.

The Kelly criterion does not tell you what to bet on. It tells you how much — and in that distinction lies the difference between a strategy that compounds across decades and a strategy that looks brilliant until the sequence of outcomes that exposes its sizing arrives. Every sequence arrives eventually. The Kelly fraction is what determines whether you are still at the table when it does.

The Kelly criterion hides wherever someone has a genuine edge but must decide how aggressively to exploit it. The signal is a decision-maker who has deliberately sized their commitment below the maximum their edge would seem to justify — who has reserved capital, time, or reputation as a buffer against the variance that even favourable odds produce over sequential trials.

The opposite signal — the Kelly violation — is equally diagnostic: a decision-maker who has committed a fraction of their resources that exceeds what their edge justifies, converting a positive-expectation opportunity into a ruin-probability accelerant. The violation is most visible after the fact, when an adverse sequence reveals that the sizing was incompatible with survival.

Common misapplication: Treating the Kelly fraction as a target rather than a ceiling.

The formula outputs the maximum bet size compatible with long-run geometric growth. It is not a recommendation to bet that amount. In practice, the inputs — probability of winning, magnitude of payoff — are estimated with substantial uncertainty. Errors in these estimates propagate directly into the Kelly fraction, and estimation error in the direction of overconfidence produces overbetting, which is the catastrophic side of the Kelly curve. This is why Thorp, the most successful practitioner of Kelly sizing in history, consistently used half-Kelly or less. The theoretical optimum assumes perfect knowledge of probabilities and payoffs. No real-world decision-maker has perfect knowledge. The gap between theoretical optimum and practical implementation should always be filled with conservatism.

A second misapplication is applying Kelly to a single, unrepeatable bet. The formula derives its power from sequential compounding — the geometric growth rate across many repeated bets. For a truly one-shot decision with no future bets to compound into, expected-value reasoning may be more appropriate. However, most decisions that feel like one-shot bets are actually sequential: the founder's company is one venture in a potential series, the investor's stock pick is one position in a career of positions. The question is not "is this the only bet I will ever make?" but "will I need capital, reputation, or health after this bet resolves?" If yes, Kelly applies.

The practitioners who have applied the Kelly criterion most successfully share a structural insight that precedes the mathematics: the recognition that how much you bet matters more than what you bet on. In each case below, the operator had a genuine edge — an analytical, informational, or structural advantage — and the Kelly framework determined how aggressively to exploit that edge without converting it into a liability through overbetting.

The pattern across these cases is consistent: the Kelly-aware operator achieves superior long-run results not by finding better opportunities than competitors but by sizing exposure to those opportunities so that adverse sequences are survivable. The edge identifies the opportunity. The Kelly fraction determines the sizing. The sizing determines whether the edge compounds into wealth or volatility extinguishes it.

The Kelly criterion sits at the intersection of probability theory, risk management, and capital allocation. Its core logic — that the optimal sizing of any commitment is determined by the ratio of edge to variance, and that overbetting is categorically more dangerous than underbetting — creates natural connections to models that address survival, compounding, and the structural architecture of risk. Kelly is rarely applied in isolation; its most powerful implementations emerge when combined with adjacent frameworks that either strengthen its survival logic, create productive friction with its conservative implications, or extend its sizing discipline into domains beyond portfolio mathematics.

The Kelly criterion appears wherever a decision-maker with a genuine edge must determine how aggressively to exploit it. The diagnostic question is always: has the sizing been determined by the magnitude of the edge relative to the variance of outcomes, or has it been determined by conviction, enthusiasm, or institutional pressure? Kelly-optimal sizing often looks timid to observers who evaluate boldness by commitment size rather than by the ratio of commitment to edge.

The Kelly criterion sits at the intersection of information theory, probability, and financial practice. Kelly's original paper provides the mathematical foundation. Thorp provides four decades of real-world evidence. Poundstone provides the narrative history connecting Shannon, Kelly, and Thorp into a coherent intellectual lineage. Together, they equip the reader to understand not just the formula but the deeper principle it encodes: in any system with multiplicative dynamics and the possibility of ruin, how much you bet determines your fate more reliably than what you bet on.

The application to startup founders is underexplored and important. A founder deciding how much personal savings to invest, how much runway to burn on a single product bet, or how aggressively to scale before achieving product-market fit is facing a Kelly problem. The edge is the founder's assessment of the probability that the bet will work. The sizing is the fraction of irreplaceable resources — money, time, reputation — committed to the bet. Most founders dramatically overbet because the culture of entrepreneurship celebrates all-in commitment and treats caution as lack of conviction. The Kelly framework reframes caution as mathematical sophistication: the founders who survive to build multiple companies are those who sized their commitments so that any individual failure was painful but not fatal.

The most elegant practitioners apply Kelly not as a formula but as a disposition. Buffett doesn't calculate a Kelly fraction before each investment. He has internalised the curve's shape: concentrate when the edge is obvious, diversify when it's not, and never commit enough to any single position that its failure would impair his ability to act on the next opportunity. The disposition matters more than the decimal because the inputs are never precise enough to justify decimal-level precision in the output. What matters is being on the correct side of the peak — the left side, where errors cost growth, rather than the right side, where errors cost everything.

The venture capital industry provides a fascinating test case. A typical VC fund has a meaningful ensemble-average return — the top-quartile funds return 3x or more — but the per-position edge is tiny and uncertain. The Kelly fraction for any individual startup investment, given a 10–15% probability of returning meaningful capital and enormous variance in payoff magnitude, is small: 2–5% of fund capital. This is precisely what successful VC fund construction looks like — forty positions at $5 million each in a $200 million fund. The fund structure is Kelly-optimal at the portfolio level even when individual GPs don't think of it that way. The funds that blow up are the ones that concentrate 20–30% into a single "high conviction" deal, overbetting an edge they cannot reliably quantify.

My operational rule: if you cannot write down, in one sentence, why your edge exists and how large it is, your Kelly fraction is close to zero and your position size should reflect that. The inability to articulate the edge is not a sign that you need more analysis. It is a sign that the edge may not exist — that your conviction is a feeling rather than a quantified advantage. The Kelly criterion turns feelings into numbers, and when the number is small, the correct response is a small bet, not a larger feeling.

The Kelly criterion has survived seventy years without modification because the problem it solves — how much to risk when you have an edge — is permanent. Markets change. Instruments change. The fundamental mathematics of sequential, multiplicative betting under uncertainty does not. Every generation rediscovers, through the expensive empirical method of blowup and bankruptcy, the same curve Kelly drew in 1956: there is a peak, the left side is forgiving, and the right side is fatal.