How Accurate Are Prediction Markets? Kalshi and Polymarket Scorecard

By Arthur Angelo, PredictionScout Research · Updated Oct 11, 2026, 9:47 AM ET

Key findings

• The favorite usually wins, and more often the further you are from coin-flip prices. One day before the outcome became clear, the higher-priced side won 70.0% of 55,376 questions. A week out it won 74.9% of 19,500; a month out, 80.6% of 7,553 (only longer-running markets reach that far back, so these are different mixes of questions).
• Mid-priced favorites won less often than their prices implied. Favorites priced 50–60¢ a day out won 52% (7,880 of 15,081); 70–80¢ favorites won 65% (4,815 of 7,354); 80–90¢ favorites 76% (4,683 of 6,180). Only 95¢-plus favorites matched their price (97.4% of 12,432). The gap was larger on Kalshi than on Polymarket (see the contract table below).
• Kalshi vs Polymarket. One day out, Kalshi’s favorite won 68.7% of 27,221 questions and Polymarket’s 71.4% of 28,155. The two exchanges list different kinds of markets, so this is not a head-to-head on the same events; the category table below is a fairer comparison.
• By topic: favorites did best in world & other (89.0% of 707 a day out) and worst in sports (61.3% of 30,226). How lopsided the typical price was explains much of the gap.
• Long shots. Across every outcome contract, those priced under 10¢ a day out paid off 3.2% of the time (average price 2.8¢; 22,514 contracts). Contracts at 90¢ or more paid off 94.5% of the time (average price 96.1¢).

Does a 70¢ favorite win 70% of the time?

How often the favorite won

The favorite is the outcome with the highest price at that moment. For yes/no markets that means whichever side was above 50¢.

When the favorite was priced at…

Both exchanges combined. If prices are honest probabilities, each row’s win rate should land inside its price range.

By topic

Favorite win rate by category, both exchanges combined, plus each exchange separately one day out. Brier score (lower is better, 0 is perfect) is across every outcome contract one day before the decision.

Every contract, by price

This counts each outcome contract on its own (a five-candidate race is five contracts), using the YES price one day before the decision. It answers the classic calibration question: of all contracts trading at, say, 70–80¢, how many paid out?

Brier score across all contracts one day out: Kalshi 0.1846, Polymarket 0.1459 (34,916 and 44,868 contract prices).

Download the data

Question-level CSVCalibration tables CSV

The question file has one row per graded question: exchange, category, the favorite and its price 1 day, 1 week and 1 month before the decision, whether it won, the price of the eventual winner, volume and a link to the market. Free to reuse under CC BY 4.0.

Methodology

• Which markets. Every market that settled YES or NO and closed between Jan 1, 2025 and Oct 9, 2026 (UTC), with at least 50,000 contracts traded on Kalshi or at least $50,000 of volume on Polymarket. Volume floors keep out thin markets where a single small trade can set the price. Voided, cancelled and 50/50 resolutions are excluded. On Kalshi we used every series with at least 1 million lifetime contracts, skipped 15-minute and hourly series (they are over before a 1-day horizon exists) and skipped multi-leg parlay contracts.
• Sports are sampled. Sports make up most qualifying markets by count (62,378 Kalshi sports events and 34,708 Polymarket sports events). We included every non-sports market and a fixed random sample of 8,833 Kalshi and 10,736 Polymarket sports events (all markets in each sampled event). The “both exchanges” figures describe this sample. If sports are weighted back to their real share, the 1-day favorite win rate is 64.9% instead of 70.0%.
• Decision time. Many markets stay open after the answer is public (an election market can stay open for weeks while results are certified). So we measure back from the moment the outcome became clear: the first hour after which the eventual winner’s price stayed at 99¢ or higher until close. If that never happened, we use the market’s close time. Horizons are exactly 24 hours, 7 days and 30 days before that moment, and a market only counts at a horizon if it was already trading then.
• Price. Kalshi: the midpoint of the hourly closing bid and ask when the spread is 10¢ or less, otherwise the last trade (the same rule as our Fed study). Polymarket: the hourly price history from its public CLOB API. Prices are read as implied probabilities.
• Questions vs contracts. When an event has several mutually exclusive outcomes (who wins a race, which bracket a number lands in) and exactly one settled YES, we treat it as one question and the favorite is the highest-priced outcome; we require the priced outcomes to add up to at least 80¢ so a missing outcome cannot pose as the favorite. Every other market is a yes/no question on its own. The contract-level table counts every outcome contract separately.
• Categories come from Kalshi’s series category and Polymarket’s market tags, mapped to eight shared groups.
• Check on the method. Measured from each market’s scheduled close instead of the decision time, the share of contracts whose favored side won one day out would be 75.0% instead of 74.4%. 20% of markets had a clear outcome at least 24 hours before they closed, which is why we do not use close time.
• Limits. Most questions are not close calls, which flatters any overall “the favorite won” number; the price-bucket tables are the fairer test. Kalshi volume is counted in contracts and Polymarket volume in dollars, so the floors are similar but not identical. Polymarket’s Gamma API reports no volume for most markets that ended between about March 5 and March 26, 2026, so those markets fall below our floor and are missing; the March 2026 Polymarket count is low for that reason, not because fewer markets settled. Outcomes are taken from each exchange’s settlement; we did not re-check each result against outside sources (our Fed study did, and all 29 matched FRED).
• Updates. The market list is refreshed weekly and new settled markets are added; past markets do not change. Last run: Oct 11, 2026, 9:47 AM ET.

Data sources: prices and volumes come from the public APIs of Kalshi and Polymarket. All trademarks belong to their owners. PredictionScout is independent and not affiliated with any exchange. Nothing here is investment, betting, or legal advice.

Related research: Biggest prediction-market upsets · Fed decision accuracy · Prediction market statistics 2026 · All research