Kelly Criterion for Prediction Markets: Position Sizing That Survives Bad Luck

Sizing is where most traders lose money they didn’t have to: the right thesis at the wrong size still ruins you. The Kelly criterion is the mathematically optimal answer — and the practically dangerous one, because it sizes to your estimated edge, and your estimate is wrong more often than you think. Here’s the formula, the worked example, and the fraction that actually survives contact with reality.

What is the Kelly criterion for prediction markets?

The bet size that maximizes long-run growth: for a binary contract at price p (your cost) with true probability q of paying $1, full Kelly fraction of bankroll is:

f* = (q − p) / (1 − p)     [for buying YES at p < q]

Buy that fraction of your bankroll’s worth of contracts — if q is honestly known. In practice q is your estimate with error bars, which is why practitioners run fractional Kelly — a quarter to half of f* — trading some growth for a massive reduction in variance and estimate-error risk. The leaderboard reality shows what full-Kelly-plus-overconfidence produces: spectacular run-ups followed by spectacular zero.

Worked example: the 60¢ market

With an estimated true probability of 70% against a 60¢ market price, full Kelly commits 25% of bankroll, half Kelly 12.5%, quarter Kelly 6.25% — and a mere five-point overestimate would have doubled the honest size.

You estimate a true probability of 70% (q = 0.70); the market prices YES at 60¢ (p = 0.60).

• Full Kelly: f* = (0.70 − 0.60) / (1 − 0.60) = 0.25 — 25% of bankroll into this position.
• Half Kelly: 12.5%. Quarter Kelly: 6.25%.

Notice the leverage of error: if your true q was 65% (not 70%), full Kelly drops to 12.5% — you’d have sized double your honest number on a five-point estimation error. That asymmetry is the entire argument for fractions: your edge estimate is the weakest input in the formula, and Kelly punishes overestimates brutally (a “sure thing” at full Kelly that loses half its probability is near-certain bankroll destruction over a series).

The three caps that keep you solvent

1. The per-position cap. Whatever Kelly says, most practitioners cap single positions at 5-20% of bankroll — event risk (resolution wording, disputes — see the resolution machinery) doesn’t care about your math. 2. The correlation cap. Positions on the same event (or correlated events) sum: five 10% positions on one election is a 50% bet wearing five tickers — the discovery filter tracks concentration before sizing does. 3. The drawdown cap. A rule set in advance: down X% from peak → size halves until you’re back. It converts tilt from a spiral into a pause — the behavioral defense the seven mistakes list centers on.

Where Kelly breaks down (be honest about these)

Perps leverage isn’t Kelly territory — funding, liquidation ladders, and session gaps make position sizing a buffer problem (see the leverage playbook and the forecaster); Kelly assumes clean binary outcomes, not margin calls. Multi-outcome events need the basket-aware version — neg-risk conversions change effective prices (see negative risk). Thin books cap implementable size below Kelly’s answer anyway — the depth check outranks the formula.

FAQ

What fraction of Kelly should I actually use? A quarter to half of full Kelly for binary markets — the growth cost is small, the variance and estimate-error protection is large. Full Kelly assumes your probability estimates are exact; they aren’t.

How do I calculate Kelly for a prediction market? f* = (q − p)/(1 − p) where q is your true-probability estimate and p the contract price. Multiply the fraction by your bankroll for position size — then apply the caps above.

Why did Kelly sizing blow up my account? Almost certainly: full Kelly on an overestimated edge, no correlation cap across correlated markets, or no drawdown rule — each converts estimation error into ruin. Fractional Kelly with caps is the survivable version.

Does Kelly work for perps and leveraged positions? Not directly — leverage adds funding, liquidation, and gap risk that break the clean-binary assumption. Size perp positions by buffer math (see the leverage playbook) and treat Kelly as the event-market tool.

How do I estimate q honestly? Start from a base rate, adjust only for information you can cite, and keep a calibration log: your stated probability vs outcomes over 50+ resolutions. Most traders discover they’re systematically overconfident — which is exactly what the fraction corrects for (see markets vs polls for estimation discipline).

Educational content, not financial advice. Not affiliated with Polymarket. Last updated October 2026.