Learn / Research
Are prediction markets accurate?
Short answer
Accuracy is measured over many events, not one. A market is well calibrated if contracts priced near 70 cents win about 70% of the time. Prices are often informative, but thin markets, longshot bias and fees mean they should be tested, not assumed.
One event cannot prove a market wrong
After every upset, someone says the prediction markets got it wrong. But a market that gave the winner a 30% chance said the upset would happen about three times in ten. If 30% events never happened, that would be the error. Accuracy is a property of many forecasts, not one.
Calibration: do 70% events happen 70% of the time?
Group every contract by its price at a fixed time before settlement, say 24 hours out, then check how often each group resolved Yes. A well-calibrated market lines up with the diagonal:
| Price bucket | Contracts | Resolved Yes | Reading |
|---|---|---|---|
| 0–10¢ | 400 | 3% | Longshots overpriced |
| 30–40¢ | 250 | 36% | Calibrated |
| 60–70¢ | 300 | 66% | Calibrated |
| 90–100¢ | 500 | 97% | Slightly underpriced |
Two choices change the answer, so state them: the snapshot time (prices hours before resolution are much better informed than prices weeks out) and the price used (last trade, midpoint or ask). Small buckets are noisy. A few hundred contracts per bucket is a reasonable minimum.
Brier score: one number for forecast quality
The Brier score is the average of (forecast − outcome)², with outcome 1 if the event happened and 0 if not. Lower is better. Three forecasts of 80% (happened), 60% (did not) and 30% (did not) score (0.04 + 0.36 + 0.09) ÷ 3 = 0.163. Always saying 50% scores 0.25, so any useful forecaster must beat that.
Comparing a model with a market
To test whether a model adds information, record the model’s probability and the market price at the same moment for every event, then score both. Four games:
| Game | Model | Market | Result | Model error² | Market error² |
|---|---|---|---|---|---|
| 1 | 65% | 58% | Yes | 0.123 | 0.176 |
| 2 | 40% | 47% | No | 0.160 | 0.221 |
| 3 | 72% | 70% | No | 0.518 | 0.490 |
| 4 | 55% | 61% | Yes | 0.203 | 0.152 |
| Mean | 0.251 | 0.260 |
The model scored slightly better, and four games prove nothing. You would need hundreds of events, scored on prices that existed before kickoff, to separate skill from luck. And a better Brier score still has to clear fees before it becomes a trading edge.
What the research says
Studies of the Iowa Electronic Markets, such as Berg, Nelson and Rietz (“Prediction market accuracy in the long run,” International Journal of Forecasting, 2008), found market prices were often closer to final election results than polls. Wolfers and Zitzewitz’s 2004 review in the Journal of Economic Perspectives describes markets as generally well calibrated while noting their limits. In betting markets more broadly, the favorite-longshot bias, where unlikely outcomes are overpriced, is one of the most consistently documented patterns.
When prediction markets are less reliable
- Thin markets. A few hundred dollars of depth can be moved by one motivated trader.
- Long horizons. Capital locked for months has a cost, which can distort prices on distant events.
- Longshots. Cheap contracts often trade above their true frequency, and fees are a large share of their price.
- Ambiguous rules. If traders disagree about what counts, the price mixes the event with the rule interpretation.
- Enthusiast crowds. Markets dominated by fans of one side can lean toward their preference.
Measure it yourself
Kelix Lines puts a simulation-based sports model beside available Kalshi, Polymarket and sportsbook prices with timestamps, so you can record both at the same moment and score them as games resolve. Read the Lines methodology for the model’s inputs and limitations, and our backtesting guide for turning a calibration gap into a tested strategy.
Frequently asked questions
If a 70% favorite loses, was the market wrong?
Not necessarily. A 70% forecast should lose about 3 times in 10. Accuracy shows up only across many forecasts, through calibration and scoring rules such as the Brier score.
What is a Brier score?
The average of (forecast − outcome)² across events, where outcome is 1 or 0. Lower is better. Always forecasting 50% scores 0.25, so a useful forecaster must beat that.
Are prediction markets better than polls or models?
Research has often found markets competitive with polls and expert forecasts, but results depend on liquidity and the event type. Compare them on the same events with the same scoring rule.
How does Kelix compare a model with a market?
Kelix Lines shows a model probability beside available market prices with timestamps, so you can measure disagreement and later check which was better calibrated.
Educational content, not financial, legal or tax advice. Prediction markets involve risk, including loss of capital. Examples are illustrative calculations, not live quotes or trading results. Kelix is independent of Kalshi, Polymarket, Robinhood and sportsbooks.