Kelly Criterion for Prediction Markets: Sizing Positions When Odds Move
You spotted the edge at 11:43. A Fed meeting market priced at 44 cents, and your read of the minutes put fair value closer to 62. You confirmed the number twice. By 11:58, when you finally opened the order ticket, the market had drifted to 51 cents and you still had no idea how much of your bankroll to put behind it.
That fifteen-minute gap is not a research problem. It is a sizing problem. Most Polymarket traders lose money not because they pick wrong but because they size wrong, and the kelly criterion is the framework that fixes that specific failure.
Quick Answer
The kelly criterion calculates the optimal fraction of your bankroll to stake on a position where you hold a measurable edge. In binary prediction markets the formula simplifies to f* = p – q, where p is your estimated true probability and q equals 1 minus p. A 65% estimate against a 52 cent market produces a full Kelly stake of roughly 30% of bankroll. Most professional traders use half Kelly or quarter Kelly instead, trading some growth rate for a large reduction in variance.
Key Takeaways
- The kelly criterion maximises the geometric growth rate of your bankroll over a long series of positions, not the expected value of a single trade, and conflating the two is the single most common misuse of the formula.
- In binary markets the math collapses to f* = p – q. A 65% estimate against a market at 52 cents produces a full Kelly fraction of 0.30, meaning 30% of bankroll on that one position.
- Full Kelly is only correct when your probability estimate is exact. Since every estimate carries error, professional traders default to half Kelly, which keeps roughly 75% of the growth rate while cutting variance by close to 75%.
- Slippage eats real edge. A market that moves 5 to 10 cents between the moment you spot mispricing and the moment your order fills has already shrunk the position you thought you were taking.
- Kelly blowups are not bad luck. They come from overstating your edge and then sizing confidently against a number that was never right.
- Correlated positions do not stack safely. Two markets tied to the same underlying outcome, each sized at half Kelly, behave like one position run at full Kelly against the shared risk.
- DG3’s Kelly Sizing Tool pulls live Polymarket prices and your fair value input into one calculation, so the stake you see is the stake you should actually take.
What Is the Kelly Criterion?
Kelly Criterion: A formula that sets the fraction of your bankroll to risk on a position with positive expected value, built to maximise the long-run compounding rate of your capital rather than the payoff of any single bet.
J.L. Kelly Jr. derived it at Bell Labs in 1956 while working on signal transmission, not gambling. Ed Thorp picked it up for blackjack a few years later and then carried it into convertible bond arbitrage, where it quietly became one of the more consequential ideas in quantitative finance.
The general formula is f* = (bp – q) / b, where b is net odds received, p is your estimated win probability, and q is 1 – p. Binary prediction market contracts settle at $1 or $0, which strips the formula down to f* = p – q, or equivalently 2p – 1.
Say your estimate is 65%. f* = 2(0.65) – 1 = 0.30. Stake 30% of bankroll at full Kelly. The logic underneath is that Kelly maximises expected log wealth, and over enough repeated positions that beats every other fixed-fraction staking rule, including ones that win more often per trade.
How to Calculate Your Kelly Stake for Prediction Markets
The process only works if you do the steps in order. Reverse them and you are not using Kelly, you are rationalising a position you already wanted to take.
Step 1: Form your probability estimate before you look at the price. Write it down. If you check the market first, your estimate anchors to it, and the whole exercise becomes circular.
Step 2: Read the current Polymarket YES price as the market’s implied probability. A YES at 0.52 means the crowd is pricing this at 52%.
Step 3: Calculate your edge. Your estimate minus the market price. 0.65 – 0.52 = 0.13.
Step 4: Apply the binary formula. f* = p – q = 0.65 – 0.35 = 0.30.
Step 5: Pick your fraction. Half Kelly multiplies by 0.5, quarter Kelly by 0.25.
Step 6: Convert to dollars. A $1,000 bankroll at half Kelly: $1,000 x 0.30 x 0.5 = $150.
Run the Fed example through the full process. A rate-hold market prices YES at 0.44. Your read of the minutes puts true probability at 0.62. Edge is 0.18. Full Kelly is 0.62 – 0.38 = 0.24, or 24% of bankroll. On a $2,000 account at half Kelly, that is $2,000 x 0.24 x 0.5 = $240. If the market later corrects to 0.62, you can take the profit without waiting on the actual Fed decision.
Full Kelly vs Half Kelly vs Fractional Kelly
Fractional Kelly: Staking a fixed proportion of the full Kelly amount, most often 0.5x or 0.25x, to price in the fact that your probability estimate is never perfectly accurate.
| Kelly Fraction | Stake (on 24% Full Kelly) | Variance | Growth Rate vs Full Kelly | Practical Use |
|---|---|---|---|---|
| Full Kelly (1x) | 24% of bankroll | Highest | 100% | Near-certain, mechanically derived edge only |
| Half Kelly (0.5x) | 12% of bankroll | ~75% lower | ~75% | Standard professional default |
| Quarter Kelly (0.25x) | 6% of bankroll | Much lower | ~44% | Uncertain edge, new market categories |
Half Kelly is the professional default for a structural reason. The Kelly growth curve is concave, so moving from full to half Kelly costs you a modest amount of growth but buys a disproportionate cut in variance and drawdown. At full Kelly, drawdowns above 50% are not an edge case, they are mathematically expected over a long enough run, and most traders quit before the compounding ever shows up.
Thorp himself, after running Kelly through blackjack and then convertible arbitrage, settled on half Kelly for real capital. His reasoning was blunt: estimation error is permanent, and the concave shape of the Kelly curve means over-betting costs more than under-betting.
A rough decision framework:
- High confidence, mechanically derived edge: 0.75x to 1.0x Kelly
- Standard research-based edge: 0.5x Kelly
- Uncertain edge or new market category: 0.25x Kelly
- Gut feel or very thin conviction: flat small stake, or skip the position entirely
Bankroll Management and Kelly Across Multiple Positions
Managing exposure across several open Polymarket positions matters as much as getting the single-position math right.
Correlated positions inflate real exposure. YES on “Candidate A wins” and YES on “Party A controls the Senate” move together most of the time. Size each at half Kelly independently and your actual exposure behaves like one full Kelly bet on the shared outcome. Treat correlated clusters as a single sizing unit, not five separate decisions.
Cap total Kelly exposure. Many systematic traders keep total open Kelly across all positions under 100 to 150% of bankroll, which stops five simultaneous half-Kelly positions from quietly becoming 60%+ of capital at risk on one afternoon.
Ruin risk compounds. A single 20% Kelly stake carries low catastrophic risk on its own. Ten correlated 20% positions do not. Kelly discipline is a portfolio problem before it is a per-trade formula.
Slippage is a tax on the fraction you calculated. Markets can move 5 to 10 cents between order submission and fill on Polymarket. If you sized at 52 cents and filled at 57, your effective edge shrank before the position even opened. On high-conviction, liquid-market positions, sizing slightly under Kelly leaves room for that gap.
Common Mistakes
Mistake 1: Using the market price as your probability estimate. Kelly needs your independent number. Anchoring to the market and then asking “is this good value” is not Kelly, it is rationalisation dressed up as math.
Mistake 2: Applying full Kelly without pricing in estimation error. The formula assumes your input is exact. It never is. A 15% overestimate of your true edge is enough to turn full Kelly from optimal into destructive.
Mistake 3: Re-running Kelly on the new price mid-position. You bought YES at 0.45, the market is now at 0.62, and you re-apply the formula to 0.62 as if the earlier position never happened. That double-counts exposure. Kelly re-sizing only applies to the marginal stake you are adding, not the whole position again.
Mistake 4: Treating every open position as independent. Five correlated sports markets, each sized at a “cautious” quarter Kelly, can add up to something close to full Kelly exposure on one underlying outcome.
Mistake 5: Applying Kelly where there is no real edge. The formula sizes an edge you already have. It does not manufacture one. Running Kelly math on a coin-flip market just produces a confident-looking stake on a coin flip.
How DG3 Helps
The hard part of using the kelly criterion on Polymarket is rarely the formula. It is getting the probability estimate right before the formula ever runs. That means stripping the vig out of the market price, adjusting for thin-liquidity distortion, and comparing your own read against a clean baseline.
DG3’s Fair Value Engine devigs live Polymarket prices in real time, giving you a reference number that is not inflated by platform margin. When your own estimate diverges from that fair value, the size of the gap is your starting input for Kelly.
The Kelly Sizing Tool takes that fair value estimate plus the live market price, runs the binary formula, and shows full, half, and quarter Kelly dollar stakes against your bankroll setting in one view, before you confirm anything. It also totals your current open Kelly exposure across positions, so correlated risk does not stack past your intended cap without you noticing.
Frequently Asked Questions
Q: What is the kelly criterion in trading? A: It is a formula that sets the optimal fraction of capital to risk on a position with positive expected value, aimed at maximising long-run geometric growth. J.L. Kelly Jr. developed it in 1956 and Ed Thorp carried it into real markets, and it guards against both under-betting, which wastes compounding, and over-betting, which risks ruin.
Q: How do you calculate Kelly stake for prediction markets? A: Use f* = p – q for binary markets, where p is your estimated win probability and q is 1 – p. Subtract loss probability from win probability, multiply by your chosen fraction, then multiply by bankroll to get a dollar figure.
Q: Should I use full Kelly or half Kelly on Polymarket? A: Half Kelly is the professional default because full Kelly is only correct when your probability estimate is exact, which it never fully is. Half Kelly keeps roughly 75% of maximum growth while cutting variance by close to the same amount.
Q: What is fractional Kelly and why do traders use it? A: Fractional Kelly multiplies the full Kelly stake by a fixed number, commonly 0.5 or 0.25, to account for estimation error. It sacrifices some theoretical growth in exchange for a real reduction in catastrophic downside.
Q: What is the risk of ruin at full Kelly? A: Drawdowns of 50% or more are mathematically expected over a long sequence at full Kelly, even with genuine edge. Half Kelly cuts expected maximum drawdown to roughly a quarter of that level, and quarter Kelly pushes practical ruin risk close to zero.
Q: How do I use the kelly criterion on Polymarket specifically? A: Read the YES price as implied probability, form your own fair value estimate independently or through DG3’s Fair Value Engine, calculate edge as your estimate minus the devigged price, apply f* = p – q, choose a fraction, and convert to a dollar stake against your bankroll.
Q: What fraction of Kelly should I use when my edge estimate is uncertain? A: Quarter Kelly or lower. The formula is only as good as its inputs, and when your probability is closer to a judgment call than a calculation, quarter Kelly builds that uncertainty into the size before it costs you real capital.
Q: How does DG3 calculate Kelly position size? A: The Kelly Sizing Tool takes your fair value probability and the live Polymarket price, runs the binary formula, and shows full, half, and quarter Kelly dollar stakes against your bankroll, alongside your total open Kelly exposure across all current positions.
Q: How do I know if I’m over-betting on Polymarket? A: Track entries and outcomes across at least 100 positions. If your realised win rate consistently sits below your average entry price, you are either picking wrong, over-betting into weak edges, or both, and calibration tracking over time separates the two problems.
Q: Can I use the kelly criterion across multiple Polymarket positions? A: Yes, but correlated positions need to be sized as a single unit rather than independently, since the combined exposure compounds. Cap total open Kelly across your full portfolio rather than treating each market in isolation.
Final Thoughts
The kelly criterion does not predict anything. It sizes a position after the prediction work is already done, and it only adds value when that underlying estimate is sound.
The estimation is where the real work lives, building a probability read that is genuinely independent of the market price and sharper than the crowd’s. Kelly is what you apply once that groundwork is in place, not a substitute for it.
For prediction markets specifically, the binary formula keeps the arithmetic simple. What separates traders who survive long enough to see the compounding work is the fraction they choose, the discipline to treat correlated positions as one exposure, and the humility to size smaller when the edge is anything less than certain.
Explore Fair Value Prediction Markets to see how DG3 builds the probability estimate that feeds directly into your Kelly stake.
