Kelly Criterion: The Math of How Much to Bet
Having a genuine edge and knowing how much to risk on it are two entirely different skills, and more investors blow up from the second failure than the first. The Kelly criterion is the formula that tries to answer the sizing question rigorously, and its most important lesson is how brutally it punishes even a small overestimate of your own edge.
The core principle
The Kelly criterion is a formula, originally developed for signal transmission problems and quickly adopted by gamblers and later by investors, that calculates the fraction of your capital to wager on a bet with a known or estimated edge in order to maximize the long-run compound growth rate of your capital. For a simple win-or-lose bet, the formula is f = (bp − q) / b, where f is the fraction of capital to bet, b is the net odds received on a win (how many units you gain per unit risked), p is your estimated probability of winning, and q is your probability of losing, equal to 1 − p.
The formula's genuine insight is that betting more than Kelly recommends does not just increase risk, it actually reduces expected long-run growth, because oversized bets increase the odds of a catastrophic drawdown that takes an outsized number of future winning bets simply to recover from. Betting less than Kelly, by contrast, reduces growth more gently and predictably, trading some upside for meaningfully less volatility. This asymmetry, that overbetting is punished far more severely than underbetting, is the central practical lesson most serious users of the formula take away from it, more than the specific number it outputs.
The formula's biggest practical weakness is that it requires you to know your true edge, your true probability of winning, with precision. In gambling with fixed, known odds, that is sometimes possible. In investing, where your "edge" is really just a probabilistic belief about future prices, that input is almost always an estimate, and a noisy one, which means the output of the formula inherits all of that uncertainty. This is why disciplined users nearly always apply fractional Kelly, deliberately betting only a fraction, commonly half or a quarter, of what the raw formula recommends, explicitly building in a margin of safety against the near certainty that their edge estimate is somewhat wrong.
How the math works
Example 1: a simple even-money bet with a real edge. Suppose you believe, based on genuine research, that you have a 55% chance of winning a bet that pays even money, meaning b = 1. Then f = (1 x 0.55 − 0.45) / 1 = (0.55 − 0.45) / 1 = 0.10, so full Kelly says to risk 10% of your capital on this single bet. On a $50,000 account, that is a $5,000 position. Betting the full 10% repeatedly, assuming the 55% edge holds exactly, maximizes the long-run compound growth rate of the account, but with substantial swings along the way even when every input is correct.
Example 2: how a small overestimate of your edge changes the outcome. Now suppose your true probability of winning was actually only 51%, not the 55% you believed, a plausible and common-sized overconfidence error. Full Kelly at the true 51% probability gives f = (1 x 0.51 − 0.49) / 1 = 0.02, or just 2% of capital, meaning that acting on your inflated 55% belief led you to bet 5 times larger than your true edge justified. Using half Kelly instead of full Kelly on your original 55% estimate would have sized the bet at 0.10 / 2 = 5% of capital, a $2,500 position rather than $5,000, cutting the damage from the overconfidence error roughly in half even though the underlying probability error was never corrected.
How it shows up in real portfolios
A systematic trader who has backtested a strategy and found a statistically favorable win rate faces the exact problem Kelly is built for: given this apparent edge, how large should each position be? The honest answer usually starts with acknowledging that a backtested win rate is an estimate drawn from historical data, not a guaranteed future probability, which is precisely why professional quantitative trading desks that use Kelly-style sizing internally almost universally scale it down, often running at a quarter Kelly or less, treating the full formula's output as a theoretical ceiling rather than an operating target.
A concentrated stock investor convinced they have identified a mispriced company is implicitly making a Kelly-style sizing decision every time they choose what fraction of their portfolio to allocate to that single idea, whether or not they ever run the formula explicitly. An investor who is right about the company's underlying value but wrong about their own confidence level, sizing a position as if their edge were larger and more certain than it actually is, experiences exactly the kind of outsized drawdown the Kelly framework predicts, even when the original investment thesis eventually proves correct.
A professional poker player or sports bettor, the population Kelly was most directly built for, provides the clearest real-world illustration: players who consistently bet full Kelly or beyond, even with a genuinely documented edge, experience far more volatile bankroll swings and a meaningfully higher rate of complete bankroll ruin during a normal losing streak than players using a fractional Kelly approach with the identical underlying edge, which is why fractional sizing has become close to standard practice among serious professionals in that world.
An investor evaluating a factor tilt, such as overweighting value stocks based on a documented long-run historical premium, faces a version of the same sizing question even though no one is calling it Kelly betting. The historical premium is itself an estimate drawn from a finite sample of market history, subject to the same overconfidence risk as any other edge estimate, which is a reasonable argument for sizing factor tilts modestly relative to a market-cap-weighted core, rather than concentrating heavily around a single historical pattern that may not repeat with the same magnitude going forward.
A retail options trader who has had a string of profitable trades on a particular setup often begins increasing position size in a way that, whether consciously or not, mirrors betting closer to full Kelly on an edge that was never rigorously quantified in the first place. A short run of favorable outcomes is weak evidence of a durable edge, and sizing decisions based on recent results rather than a genuinely tested probability estimate are one of the more common ways traders convert a real but modest edge into an eventual, sometimes account-ending, loss.
Actionable breakdown
- Before applying Kelly-style sizing at all:
- Confirm you have a real, evidence-based edge.
- State your win probability as a specific number.
- Write down why that number is honest, not hopeful.
- When sizing the position:
- Calculate full Kelly first as a reference ceiling.
- Default to half or quarter Kelly in practice.
- Recalculate whenever your edge estimate changes.
- Ongoing discipline:
- Track realized results against your assumed edge.
- Reduce sizing further if results run below expectations.
Common pitfalls
- Overestimating your edge: a modest error in your probability estimate produces a wildly oversized recommended bet, since the formula amplifies small input errors.
- Betting full Kelly on real capital: tolerating far more volatility and drawdown risk than most people can withstand psychologically without abandoning the strategy.
- Applying the formula with no real edge: the output is meaningless, or actively destructive, when the underlying probability estimate is just a guess.
- Treating one calculation as permanent: failing to resize as new information changes your actual edge, leaving an outdated bet size in place.
Related concepts
For the broader danger of oversized positions even without a formula involved, see leverage and margin. For the discipline of matching risk to what you can actually withstand, see risk tolerance. For a real-world venue where sizing errors are punished quickly and visibly, see day trading. For the guide on sizing risk across a full portfolio, see the risk guide.
The bottom line
Use the Kelly criterion as a ceiling to bet less than, not a target to bet exactly, because the cost of overestimating your edge is always larger than the cost of underestimating it.