The Kelly criterion is a position-sizing rule that prescribes the fraction of capital to risk on a bet (or trade) in order to maximize the long-run growth rate of wealth — formally, to maximize expected logarithmic wealth, which is the same as maximizing the expected geometric (compounded) growth rate. Derived by John L. Kelly Jr. at Bell Labs in 1956 by reinterpreting Claude Shannon's information theory, it answers a question that fixed-percentage rules cannot: given a measurable edge and payoff, what is the mathematically optimal bet size? Its core tension is that this optimum is optimal only when the edge and odds are known exactly — and in markets they never are. Bet the full Kelly fraction off an overestimated edge and a growth-maximizing rule becomes a wealth-destroying one. This section covers the criterion, its inputs, the universal practice of betting below full Kelly, and the survival math (risk of ruin) that bounds it.

The core formula

For a discrete bet where a loss costs the entire stake, the Kelly fraction is

f\* = p − q/b = (bp − q) / b

where p is the win probability, q = 1 − p, and b is the net odds (profit per unit wagered on a win) — equivalently, for trading, f\ = W − (1 − W)/R, with W the win rate and R the win/loss payoff ratio (Kelly 1956; Wikipedia, Kelly criterion). For a continuously-distributed return stream the analogue is f\ = (μ − r) / σ² — excess mean return over variance — which is algebraically identical to the optimal mean-variance leverage, tying Kelly directly to portfolio theory (Wikipedia; Thorp 2006). The output is a fraction of current capital, so Kelly is inherently a compounding, bet-a-percentage rule rather than a fixed-dollar one.

The core tension: optimal vs. survivable

Kelly's appeal is that it is provably growth-optimal over a long horizon — no other fixed fraction compounds faster, and it cannot be ruined in the literal infinite-bet sense because it always bets a fraction. But "growth-optimal" hides brutal short-run behaviour: full Kelly carries roughly a 50% chance of a 50% drawdown at some point, and the growth curve is dangerously asymmetric. Underbetting only slows compounding; overbetting degrades it fast, and beyond ~2× Kelly the long-run growth rate turns negative — guaranteed ruin despite a genuine positive edge (Thorp 2006; Wikipedia). Because real edges are estimated (and usually overestimated), and because the penalty for landing on the overbet side is so much worse than for the underbet side, the entire practical apparatus of Kelly is about deliberately sizing below the formula's output. That is the unifying theme of this branch's three sub-topics.

When it matters — and when it doesn't

Kelly is most useful where an edge is repeatable, roughly quantifiable, and bets are sequential — blackjack, sports betting, systematic trading strategies with a logged track record. It is the right mental model whenever capital compounds across many independent opportunities. It is least useful for one-off or unquantifiable decisions (no stable p and b to estimate), for utility functions that aren't logarithmic (an investor near a hard liability or short horizon should not maximize log-growth), and whenever inputs are too noisy to estimate — which describes most discretionary trading. In those cases a simpler edge-agnostic rule (fixed-fractional 0.5–2% risk; volatility targeting) is more robust, and Kelly serves only as a ceiling the simpler rule should never approach.

Adoption & standing

The criterion sat largely unnoticed until Ed Thorp learned of it from Shannon, applied it to blackjack, then ran Kelly-style sizing at the quant hedge fund Princeton-Newport Partners — the success that put it on Wall Street's radar (Poundstone, Fortune's Formula; Thorp 2006). Today it is widely taught and used in quantitative and systematic circles, gambling, and parts of the hedge-fund world, typically at a fractional setting. It is much less common — and arguably inappropriate — as a literal sizing rule for discretionary retail traders, who lack the stable, validated inputs it demands. The most famous live controversy is Paul Samuelson's repeated attack on the "Kelly fallacy": he argued that maximizing expected log wealth is not optimal for anyone whose utility isn't logarithmic, and that "it dominates in the long run" is misleading over finite horizons. The honest split: the math of Kelly and fractional Kelly is exact and settled; whether log-growth is the right objective, and whether any given trading edge can be estimated reliably enough to size off, are genuinely contested.

Strengths & limitations

Strengths. Provides a principled, non-arbitrary link from measured edge and variance to bet size; correctly assigns small bets to thin edges and refuses to bet a negative edge at all; unifies position sizing with portfolio theory (the continuous form is mean-variance leverage).

Limitations. It assumes known, stationary probabilities, one bet resolved before the next, no transaction costs, and an infinite horizon — all false in markets. It is non-robust to estimation error in the worst direction (overbetting), and it ignores correlation: several open positions in the same regime act like one oversized Kelly bet. The single most common misuse is plugging an overfit backtest's win rate and payoff into the formula and betting the raw f\* — combining an inflated edge estimate with the unforgiving overbet side of the growth curve, which is how a positive-edge system still blows up.

Map of this section

• Full vs Fractional Kelly — why almost no professional bets full Kelly; the parabolic growth result (half Kelly keeps ~¾ of the growth at roughly half the volatility), and the ¼–½ Kelly convention.
• Estimating Edge & Odds — how W and R are derived, their sampling error, the upward plug-in bias, and why the asymmetric penalty means you should shrink the inputs and the fraction.
• Risk of Ruin — the survival side of the coin: the probability that a losing streak depletes capital, why over-sizing (not low win rate) blows up accounts, and how this justifies the fixed-fractional 0.5–2% floor most traders should actually operate at.

Sources

• John L. Kelly Jr., "A New Interpretation of Information Rate," Bell System Technical Journal 35(4), 1956 — original derivation (log-growth objective, discrete formula).
• Wikipedia, Kelly criterion — discrete form f\ = p − q/b, continuous form f\ = (μ−r)/σ², log-wealth objective, >2× Kelly → negative growth, Samuelson critique. https://en.wikipedia.org/wiki/Kelly_criterion
• Edward O. Thorp, "The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market" (2006) — continuous f\* = m/s², drawdown behaviour, fractional-Kelly rationale.
• William Poundstone, Fortune's Formula (2005) — history: Shannon → Thorp → Princeton-Newport adoption.
• John Larry Kelly Jr. — Wikipedia (Bell Labs / Shannon information-theory lineage). https://en.wikipedia.org/wiki/John_Larry_Kelly_Jr.
• Disputed: Samuelson's "Kelly fallacy" critique of the log-utility objective remains unresolved at the level of whether growth-optimality is the right goal; this doc treats the Kelly/fractional-Kelly math as settled and its applicability to real, estimated edges as contested.