Bell Labs' Kelly criterion turned an edge into exponential growth without the ruin

In 1956 John L. Kelly Jr., a physicist at Bell Labs working on television signal transmission, reframed a gambler's problem in information theory's language. In his example, a bettor with a 'private wire' on horse races knows the true outcome probabilities while the odds on offer do not. Kelly asked not how much the bettor could win on one bet, but how fast capital could grow over a lifetime of bets, defining G as the exponential growth rate of wealth measured with base-two logarithms.

The answer exposed a trap. Betting the entire bankroll every time maximizes expected wealth, yet as Kelly put it, the bettor 'would probably be broke and, in fact, would be broke with probability one if he continued indefinitely.' Instead, the bettor should place a fraction of capital chosen so the expected logarithm of wealth is maximized. At exactly that fraction the growth rate equals Shannon's rate of information transmission, and with probability one that bettor eventually pulls ahead of anyone allocating differently.

Kelly himself noted the theory fit 'the totality of inside information available to the investor.' Edward Thorp took that instruction literally. After using MIT's IBM 704 to prove blackjack could be beaten and publishing a favorable strategy in the Proceedings of the National Academy of Sciences in 1961, Thorp used the Kelly criterion to decide how much of his bankroll to risk when the count was favorable — and then moved the same machinery from casino to market.

As co-founder of Princeton Newport Partners, Thorp ran the rule for nearly two decades: the fund compounded at 19.1% a year from 1969 to 1988, more than double the S&P 500 over the period, and never had a losing year. The edge-hunting and the bet-sizing were separate disciplines, and the second one is what kept the first one alive.

Maximize expected log capital, not expected profit

The logarithm is additive across repeated bets, so the law of large numbers applies to growth rates; expected wealth is not, which is why maximizing it permits ruin.

Overbetting converts a positive edge into a negative growth rate — the paper shows a bettor can have favorable odds and still, with probability one, go broke by betting too much of the bankroll.

The growth-rate bound ties bet size to the information advantage itself: the better the private wire, the larger the fraction worth risking, making sizing a measurement problem.

For a fund taking repeated trades for decades, the criterion answers the question that decides survival: how much can be staked so the drawdown never ends the game.

Having an edge and surviving to compound it are different problems. Maximizing one-bet expected profit still goes broke; maximizing expected log wealth maximizes the path you actually live.

The criterion spread from Bell Labs through gambling into quantitative finance: Thorp's blackjack strategy appeared in PNAS in 1961 and Beat the Dealer in 1966, and Princeton Newport's 19.1% compound from 1969 to 1988 without a losing year made log-optimal sizing the standard of the quant industry. The fund ended in 1988 over an unrelated employee's securities-regulation trouble, not over the strategy. Kelly's name is now attached to position-sizing debates in markets, sports betting, and even daily-leverage design.