Kelly Criterion for Prediction Markets: Position Sizing That Survives (2026)

Most prediction market traders don’t blow up because they’re wrong about events — they blow up because they sized positions as if being right were the only variable. The Kelly Criterion is the mathematical answer to position sizing: it maximizes long-run bankroll growth for a given edge. Full Kelly is aggressive; fractional Kelly is how professionals actually use it. Here is the complete framework with prediction-market-specific adjustments.

The Kelly formula for binary contracts

For a binary contract priced at p_market (your cost in cents) where you estimate the true probability as p_true:

f = (p_true - p_market) / (1 - p_market)      for buying YES

where f is the fraction of your bankroll to commit.

Worked example: Polymarket prices a market at 42¢. Your research says the true probability is 55¢:

f = (0.55 - 0.42) / (1 - 0.42) = 0.224

Full Kelly says commit 22.4% of your bankroll. That is a huge position — which is exactly why almost nobody runs full Kelly.

Why full Kelly is dangerous in practice

Kelly assumes you know p_true. You don’t — you estimated it. If your estimate is biased (and every estimate is), full Kelly systematically overbets, and the penalty for overbetting is asymmetric: overbetting beyond 2× Kelly guarantees negative growth no matter how good your edge is.

Worked example, continued: at half Kelly you commit 11.2% instead of 22.4%. Your long-run growth rate drops only modestly, but your variance drops by roughly half, and your tolerance for estimate error quadruples.

Prediction-market-specific adjustments

Three adjustments make Kelly usable on Polymarket and Kalshi:

1. Fees and spread reduce your realized edge. If the effective round-trip cost is 2 cents, subtract it from your edge before computing f. An edge that looks like 13 cents is really 11.

2. Resolution risk is not a normal loss. Kelly assumes defined odds. A market that resolves against the written criteria — UMA disputes, clarified rules, voided outcomes — can take your stake at 100% when your model said 70%. Multiply your p_true by a resolution-confidence factor for markets with ambiguous criteria.

3. Correlated positions are one position. Ten midterms markets that all hinge on the same polling shift are not ten independent Kelly bets — they are one bet on polling accuracy, sized ten times too large if you size each independently. Group correlated positions and size the group.

What this looks like in code

def kelly_fraction(p_true, p_market, fee_cents=0.02, resolution_conf=0.95):
    edge = p_true - p_market - fee_cents
    adjusted_true = p_market + edge * resolution_conf
    if adjusted_true <= p_market:
        return 0.0
    f = (adjusted_true - p_market) / (1 - p_market)
    return max(0.0, min(f * 0.5, 0.10))  # half Kelly, hard cap 10%

# Example: Polymarket at 42c, your estimate 55c
print(kelly_fraction(0.55, 0.42))  # ~0.097 → 9.7% of bankroll

The hard cap matters: it limits damage from any single bad estimate, bad resolution, or platform incident — see the June 2026 incident for why platform risk belongs in your sizing too.

FAQ

What is the Kelly Criterion in simple terms? A formula that tells you the bankroll fraction to bet when you have a measurable edge: edge divided by net odds. Bet more and you risk ruin; bet less and you leave growth on the table.

Why do professionals use fractional Kelly? Because full Kelly assumes a perfect probability estimate. Half Kelly gives up little growth but massively reduces variance and protects against estimate error.

Does Kelly work for prediction markets specifically? Yes — binary contracts are the natural Kelly case. The adjustments that matter are fees/spread, resolution ambiguity, and correlated markets.

How do I size correlated positions? Treat markets that share a common driver as one group, estimate the group edge once, and apply Kelly to the group total rather than each market independently.

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