Bell Labs and the Optimal Bet
Kelly's paper was ostensibly about information theory — specifically, how a gambler with a noisy private information channel should bet to maximise their long-run wealth growth rate. The connection to Claude Shannon's information theory was explicit: Kelly showed that the optimal growth rate equals the mutual information between the private signal and the outcome. But the practical result — the betting formula — transcended its theoretical context and became foundational in gambling theory, portfolio management, and quantitative finance. Edward Thorp, the mathematician who proved that blackjack could be beaten through card counting, recognised Kelly's result as the missing piece for turning a theoretical edge into maximum practical profit. Thorp applied Kelly sizing to blackjack, then to warrant arbitrage, and built one of the first quantitative hedge funds using Kelly-derived position sizing. Warren Buffett's partner Charlie Munger has spoken publicly about Kelly's influence on Berkshire Hathaway's concentration strategy.
Why Logarithmic Utility?
Kelly's formula emerges from maximising the expected logarithm of wealth — not the expected value of wealth itself. This distinction is crucial. Maximising expected wealth leads to maximising bet size on any positive-EV opportunity, which quickly leads to ruin because a single loss can destroy the bankroll. The St. Petersburg paradox (a game with infinite expected value but which no rational person would pay much to play) illustrates why raw expected value is an inadequate guide to decisions under uncertainty. The logarithmic utility function captures a property called "constant relative risk aversion" — the disutility of losing X% of your wealth is independent of how much wealth you have. This is a reasonable model for anyone who thinks proportionally about money, and it leads naturally to Kelly sizing as the growth-maximising strategy that avoids eventual ruin. Betting too little is suboptimal — you grow more slowly than possible. But betting too much is catastrophic — Kelly proves that any bet above 2× the Kelly fraction leads to negative expected log growth, meaning certain ruin in the long run. The Kelly fraction is the asymptote: more is worse, less is merely suboptimal.
Fractional Kelly in Practice
Professional gamblers and fund managers almost universally use "fractional Kelly" — betting a fixed fraction (typically 25–50%) of the full Kelly bet. The motivation is parameter uncertainty: the Kelly formula requires knowing your true edge precisely, which you typically don't. If you overestimate your edge, full Kelly can be wildly too aggressive.
Applying Kelly to Sports Betting
Sports betting offers the rare possibility of genuine positive expected value for skilled analysts — unlike casino games where the probabilities are fixed. A bettor who can estimate the true probability of an outcome more accurately than the bookmaker's odds imply has a genuine edge. Applying Kelly requires three inputs: the decimal odds offered, your estimated true probability, and the resulting implied edge.
Kelly in Financial Trading
Kelly sizing was introduced to financial portfolio theory by Ed Thorp and independently developed by others. In the continuous-time setting, Kelly bet sizing for a stock with expected excess return μ and variance σ² gives the optimal leverage as f* = μ/σ² — the same formula as the Sharpe-ratio-maximising portfolio, scaled to full Kelly. Commodity Trading Advisors (CTAs) and quantitative hedge funds use Kelly-derived position sizing as a core risk management framework, though typically at fractional Kelly to manage drawdown risk.
When Kelly Fails
Kelly's formula is derived for i.i.d. (independent, identically distributed) bets with known probabilities. Real-world betting and trading violate these assumptions in several ways. Outcomes are often correlated — a sports bettor who has multiple games in the same league on the same day has correlated bets, and the multivariate Kelly formula (requiring covariance estimates) is much harder to apply. Edge estimates are uncertain and change over time as bookmakers adjust lines. And Kelly ignores psychological factors — the very large drawdowns that occur on the path to long-run optimal growth are too severe for most people to sustain behaviourally, even when mathematically justified. EngForge's Monte Carlo simulator lets you model thousands of simulated betting sequences, comparing full Kelly, half Kelly, and fixed-fraction strategies — visualising how variance and growth rate trade off across different bet sizing rules.