Polymarket's prices have been well calibrated in Nimb's own check of 1,353 resolved 2026 markets: outcomes priced around 60% a week before the end happened 62% of the time, those around 22% happened 34% of the time, and of 603 outcomes priced above 90% or below 10% a week out, 19 (3%) went the other way. The prices were far better than a guess from the base rate, but not perfect: in this sample, cheap outcomes between 5% and 30% happened somewhat more often than their prices implied.
How accurate is Polymarket?
Calibration asks a simple question: when the market says 70%, does it happen about 70% of the time? Nimb grouped outcomes by their price before the end and counted how often each group came true.
| Price 7 days before close | Outcomes | Average price | Happened |
|---|---|---|---|
| 0–5% | 499 | 0.8% | 1.2% |
| 5–15% | 124 | 9.3% | 16.1% |
| 15–30% | 125 | 21.8% | 33.6% |
| 30–50% | 132 | 40.0% | 49.2% |
| 50–70% | 156 | 60.2% | 62.2% |
| 70–85% | 57 | 76.1% | 80.7% |
| 85–95% | 31 | 89.4% | 93.5% |
| 95–100% | 20 | 97.9% | 100% |
Does accuracy improve closer to the end?
| Price taken | Outcomes | Brier score | Base-rate guess | Favourite won (10–90% only) |
|---|---|---|---|---|
| 1 day before close | 1,342 | 0.095 | 0.204 | 67.8% |
| 7 days before close | 1,144 | 0.118 | 0.203 | 65.4% |
| 30 days before close | 782 | 0.105 | 0.173 | 69.3% |
The Brier score is the average squared gap between price and result (0 is perfect; lower is better). "Base-rate guess" is the score of always saying the share of outcomes that came true, about 28%. Prices beat it by roughly half at every horizon. The 30-day row is a different, smaller set of markets (those open for at least a month), so it is not directly comparable.
How the check was done
- Sample. The most traded closed Polymarket markets with an end date between 1 January and 7 November 2026, binary Yes/No, at least $100k traded, resolved through UMA with a final price above 99%: 1,353 markets, closed between 30 October 2025 and 8 November 2026, fetched from the Gamma API on 8 November 2026.
- Prices. The last daily price from Polymarket's CLOB price history at or before 1, 7 and 30 days before each market closed.
- Caveats. 928 of the markets are outcomes of multi-outcome events, so they are not independent. Many markets close early when the event happens, which helps the 1-day row. The sample favours heavily traded, prominent markets; small markets may be less accurate. One year is one period. The script is in Nimb's repository (
pipeline/research/calibration.py) so the check can be repeated.
What does this mean for reading a price?
- Treat a Polymarket price as a reasonable estimate of the chance, not a verdict. A 20% outcome happens often enough to matter.
- In this sample low prices were, if anything, too low, the opposite of the long-shot bias found in many older studies. The favourite–long-shot bias.
- Calibration over many markets says nothing certain about one market: check its money, depth and rules. How to read a market · Accuracy research.
Frequently asked questions
What is Polymarket's accuracy track record?
In Nimb's check of 1,353 resolved 2026 markets, prices a week before the end were well calibrated, with a Brier score of 0.118 against 0.203 for a base-rate guess.
Were prediction markets right about the election?
One election is one outcome. Calibration over many markets is the better test; see prediction markets vs polls.
What is prediction market calibration?
How closely prices match how often things happen: a well-calibrated market's 70% outcomes happen about 70% of the time.
Last reviewed 8 November 2026 · Data from Polymarket · Not financial advice · Nimb is not affiliated with Polymarket.