A Physicist's Insight About Optimal Betting
John Kelly was a physicist at Bell Labs in the 1950s working on information theory for long-distance telephone lines. His insight about optimal bet sizing came from thinking about noisy communication channels, not gambling. The core idea: if you know the probability of an outcome and the odds being offered, there is a mathematically optimal fraction of your bankroll to wager that maximizes long-term capital growth.
The formula is deceptively simple. Kelly fraction equals (bp minus q) divided by b, where b is the net odds received on the wager, p is the probability of winning, and q is the probability of losing (1 minus p). If a prediction market contract pays 2-to-1 and you estimate a 60% probability of the event occurring, Kelly says to bet 40% of your bankroll.
Why That Number Should Make You Uncomfortable
If 40% made you flinch, your instincts are sound. Full Kelly sizing is extremely aggressive. A string of losses, which will inevitably happen even with a genuine edge, can draw down your portfolio by 50% or more before the law of large numbers kicks in.
Ed Thorp, the mathematician who first applied Kelly's ideas to blackjack and later to financial markets, became one of the strongest advocates for fractional Kelly. In his 2007 paper for the World Scientific Handbook, Thorp demonstrated that betting half Kelly gives you roughly three-quarters of the optimal long-run growth rate with half the volatility. The tradeoff is overwhelmingly favorable. Bill Gross, who ran PIMCO (one of the world's largest bond funds), credited Thorp's Kelly Criterion work as a foundational part of his risk management approach.
Most professional bettors and quantitative traders use between one-quarter and one-half Kelly. The reason is not just comfort. Thorp himself pointed out that people have a background tendency to overestimate their probability of winning. Fractional Kelly acts as built-in insurance against your own overconfidence.
Why Prediction Markets Make Kelly Practical
Kelly sizing is more practical in prediction markets than in most financial applications because the payoff structure is clean and binary. A contract pays $1 or $0. If you can buy at 45 cents and you estimate the true probability at 55%, the expected value math is straightforward.
The critical input, and where most people go wrong, is the probability estimate itself. Kelly assumes you know the true probability. In reality, your estimate carries its own uncertainty. If you think an event has a 55% chance of occurring but your confidence interval spans 45-65%, your effective edge might be much smaller than you think, or nonexistent.
A Practical Workflow
Here is an approach that works. First, estimate the probability of the outcome independently, before looking at the market price. Write it down. Then compare your estimate to the market price. If they are within 5 percentage points, you probably do not have a meaningful edge. If they differ by more, ask yourself specifically why. What do you know that the market does not?
If you can articulate a genuine informational or analytical edge, calculate full Kelly, then take one-quarter to one-half of that amount. As your portfolio grows or shrinks, recalculate sizes as a percentage of current bankroll, not initial capital. This dynamic sizing is what keeps Kelly working over time rather than blowing up after a losing streak.
When Kelly Breaks Down
Kelly assumes outcomes are independent (they often are not, especially in correlated markets). It assumes you can make fractional bets (usually possible in markets). And it assumes your probability estimates are accurate, which is the biggest and most common failure mode.
Overconfidence in probability estimates is the single most dangerous input error. If you think you have a 70% edge on something that is actually 50-50, full Kelly will systematically destroy your capital. This is why fractional Kelly is not conservative temperament disguised as math. It is genuine insurance against model error.
Explore these tools on Blockcircle: Prediction Markets Mispricing Engine