Strategy
Kelly criterion: how big should each trade be?
The Kelly formula tells you the optimal bet size given your edge and the odds. Most prediction market traders bet way too much. Here is how to use Kelly without blowing up.
The Kelly criterion is a formula from information theory that tells you the optimal fraction of your bankroll to bet, given your edge and the odds offered. It was developed by John Kelly at Bell Labs in 1956, weaponized by professional gamblers (Ed Thorp, the Princeton-Newport Partners) and now appears in every serious quant trading desk's risk framework.
Most retail prediction market traders bet too much. Kelly is the antidote.
The formula
For a binary bet with probability p of winning and net odds b (i.e., you win b dollars for every 1 dollar risked):
f* = (bp - q) / b, where q = 1 - p.
f* is the fraction of your bankroll to bet. If it comes out negative, the bet has negative expected value — don't take it.
Worked example
Polymarket has a contract trading at 0.40. You estimate the true probability is 0.55. Buying Yes at 40 cents:
- p (your estimate of winning) = 0.55
- You risk 0.40 to win 0.60. Net odds b = 0.60 / 0.40 = 1.5
- q = 1 - 0.55 = 0.45
- f* = (1.5 × 0.55 - 0.45) / 1.5 = (0.825 - 0.45) / 1.5 = 0.25
Kelly says bet 25% of your bankroll. That number is going to feel large to most people. That's the catch.
Why almost everyone uses fractional Kelly
Full Kelly is mathematically optimal if your probability estimate is exact. It is also extremely volatile. A full-Kelly bettor will routinely see drawdowns of 50%+ even with positive edge.
Worse, your probability estimates are never exact. If you think p = 0.55 and the truth is 0.50, full Kelly says bet 25%, but the truth is you have no edge. You're now gambling, not investing.
The standard professional fix: use half Kelly or quarter Kelly. Multiply f* by 0.5 or 0.25. You sacrifice a little long-run growth in exchange for dramatically smaller drawdowns and more forgiving treatment of estimate errors.
What counts as your bankroll?
Your bankroll is the amount you've earmarked for prediction market trading and would be comfortable losing in full. Not your savings. Not your emergency fund. Not "the amount I deposited last week plus winnings". A clear, predetermined bankroll.
If you have $1,000 set aside for prediction markets, your quarter-Kelly maximum bet on the 0.55-probability example above is $62.50. Most beginners would put $300 on something they're confident about. That's why most beginners blow up.
When you have many simultaneous bets
Kelly assumes one bet at a time. In practice you have dozens of positions. The naive application — independently Kelly-size each one — leads to over-exposure when bets are correlated.
Two practical adjustments:
- Cap total exposure. Never have more than 30–50% of bankroll in active positions regardless of what individual Kelly numbers say.
- Treat correlated bets as one bet. Five bets all on different aspects of a single election are really one bet on the election outcome. Size them together.
When Kelly doesn't apply
Kelly assumes you have an edge — that your probability estimate is genuinely better than the market's. If you don't, Kelly will tell you to bet zero (or negative, meaning bet the other side). If you can't articulate why your estimate is better than the consensus, the answer is: you don't have edge here. Skip the trade.