The Kelly Criterion

Imagine you have a biased coin that lands heads 60 % of the time. You start with $1,000 and can bet any fraction of your current bankroll on each flip, winning your stake if heads and losing it if tails. How much should you bet to get as rich as possible?

Bet nothing and you stay at $1,000 forever. Bet everything and a single tail wipes you out. Somewhere in between is a sweet spot — and in 1956 John L. Kelly Jr., an engineer at Bell Laboratories, found it exactly.

Kelly's insight came from information theory: each bet is like receiving a noisy signal, and the optimal strategy is to bet at the rate that maximizes the long-run growth rate of your wealth — the same way Shannon's entropy maximizes the rate of information transmission. The formula is elegant: bet the fraction $f^* = p - q/b$, where $p$ is the probability of winning, $q = 1 - p$ is the probability of losing, and $b$ is the net odds received (winning $b$ per $1$ risked). For our 60/40 coin with even payoffs ($b = 1$) that's $f^* = 0.60 - 0.40 = 0.20$ — bet exactly 20 % each time.

The result is proven optimal for maximizing the expected logarithm of wealth, which is equivalent to maximizing geometric growth. Bet less and you grow more slowly than necessary. Bet more — even a tiny bit more — and your long-run wealth shrinks toward zero with probability 1. The Kelly criterion is not a heuristic; it is a mathematical theorem.

The demo below runs 200 coin flips with a 60 % heads coin and tracks three bettors simultaneously: one who follows Kelly exactly (20 %), one who bets only 10 % (underbetting), and one who bets 40 % (overbetting).

<div class="controls">
  <label>{{lbl_prob}} <span id="p-val">0.60</span>
    <input type="range" id="prob" min="0.51" max="0.75" step="0.01" value="0.60">
  </label>
  <label>{{lbl_flips}} <span id="n-val">200</span>
    <input type="range" id="flips" min="50" max="500" step="50" value="200">
  </label>
  <button id="run" type="button">{{btn_run}}</button>
</div>
<canvas id="chart" width="560" height="280"></canvas>
<div class="legend">
  <span class="dot kelly"></span> {{leg_kelly}}
  <span class="dot under"></span> {{leg_under}}
  <span class="dot over"></span> {{leg_over}}
</div>
<div id="result" class="result"></div>

* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; margin: 0; color: #222; }
.controls { display: flex; flex-wrap: wrap; gap: .6rem 1.2rem; align-items: center; margin-bottom: .7rem; font-size: .88rem; }
label { display: flex; flex-direction: column; gap: .15rem; }
label span { font-weight: 700; font-size: .92rem; }
input[type=range] { width: 120px; }
button { font: 700 14px system-ui; padding: .4rem 1rem; background: #1d3557; color: #fff; border: none; border-radius: 8px; cursor: pointer; }
canvas { width: 100%; height: auto; border: 1px solid #dde3ea; border-radius: 8px; background: #f8fafc; display: block; }
.legend { display: flex; gap: 1.2rem; font-size: .82rem; margin-top: .4rem; }
.dot { display: inline-block; width: 12px; height: 12px; border-radius: 50%; margin-right: .3rem; }
.dot.kelly { background: #2a9d8f; }
.dot.under  { background: #e9c46a; }
.dot.over   { background: #e63946; }
.result { font-size: .88rem; font-weight: 600; margin-top: .5rem; min-height: 1.2em; color: #444; }

// Code not found

Hit Run to play out a fresh sequence. Notice how the overbetter's curve collapses — even when it briefly overtakes Kelly, it eventually ruins itself. Hit Run many times to build intuition: Kelly wins in the long run almost every time, while the overbetter oscillates wildly and trends downward.

Why does this specific formula win? The answer lives in logarithms and the law of large numbers.

After n bets, your wealth is the product of n multipliers. If you win W bets and lose L, your bankroll is:

$$\text{Bankroll} = (1 + f \cdot b)^W \cdot (1 - f)^L$$

Taking logarithms turns this product into a sum, which the law of large numbers makes predictable. The per-bet growth rate (the expected log-return) is:

$$G(f) = p \cdot \log(1 + f \cdot b) + q \cdot \log(1 - f)$$

Setting $dG/df = 0$ and solving gives $f^* = p - q/b$ — exactly the Kelly fraction. Because the growth rate is a strictly concave function of $f$, this is a global maximum: any other bet size yields a lower growth rate.

Key theorems:

• Kelly (1956): the log-optimal strategy maximizes the expected growth rate of the bankroll per bet.
• Breiman (1961): among all non-anticipating strategies, Kelly almost surely achieves the maximum long-run wealth — no other strategy can beat it asymptotically.
• Ruin: if $f > f^*$, the growth rate turns negative, and the bankroll converges to zero with probability 1, regardless of how favorable the bets are.

This connects to information theory: Kelly showed that the optimal growth rate equals the channel capacity — the mutual information between the bet outcome and prior knowledge. Gambling and communication are formally the same problem.

The Kelly fraction generalizes to multiple simultaneous bets and to continuous distributions. In the continuous case it becomes a portfolio optimization that maximizes the expected log-return, solved via the same first-order condition.

The Kelly criterion is not a curiosity — it is actively used wherever people must make repeated, favorable bets with uncertain outcomes:

• Sports and casino betting: professional bettors and card counters (most famously the MIT Blackjack Team) use Kelly sizing to maximize the growth of their bankrolls while surviving bad runs.
• Algorithmic trading: quantitative funds use Kelly (or fractional Kelly, typically half-Kelly) to size positions. Full Kelly can be too volatile for institutional mandates; half-Kelly gives roughly 3/4 of the growth with much smaller drawdowns.
• Venture capital: Kelly reasoning appears in portfolio construction — how much of the fund to put into each deal, given win-rate and return-multiple estimates.
• Information theory and communications: Kelly's original 1956 paper framed the problem as channel coding. The optimal bet fraction equals the channel's information rate, linking gambling and Shannon entropy.
• Evolutionary biology: growth-rate maximization under uncertainty appears in models of natural selection, where organisms that "bet" resources optimally outcompete those that don't.

See also linear programming for portfolio optimization and Bayesian inference for how to estimate p and b from data when the true probabilities are unknown.

The Kelly criterion answers one of the oldest questions in probability: how much should I bet? The answer is $f^* = p - q/b$ — your edge divided by your odds. Bet this fraction and you are provably maximizing your long-run wealth. Bet more and you are guaranteed, over enough time, to go broke regardless of how favorable the game is.

The formula's elegance comes from a deep connection: Kelly discovered that optimal betting is the same problem as optimal information transmission. Your edge over the casino is exactly your mutual information about the outcome, and the Kelly fraction is the rate at which you can reliably "transmit" that information into wealth.

In a world full of uncertain, repeated decisions — markets, games, investments — the Kelly criterion is the uniquely correct answer to how aggressively to act on your edge.