How accurate are prediction markets?
How often Polymarket prices came true, by topic and price range — measured on resolved markets, with the cases we left out and why.
Last updated Oct 10, 2026
The measurement
Across 17,347 resolved markets, the price seven days before resolution scored a Brier of 0.202, against 0.245 for always forecasting the base rate — a skill of 0.18.
All topics, seven days before resolution
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.4% | 2.1% | 1,304 | reliable |
| 5–15% | 9.7% | 13.7% | 1,181 | happened more often |
| 15–35% | 25.6% | 30.3% | 3,432 | happened more often |
| 35–65% | 49.0% | 48.7% | 8,870 | reliable |
| 65–85% | 73.1% | 68.5% | 1,758 | happened less often |
| 85–95% | 90.1% | 88.0% | 515 | reliable |
| 95–100% | 98.1% | 97.6% | 287 | reliable |
By topic
| Topic | Cases | Brier | Base-rate Brier | Skill |
|---|---|---|---|---|
| Sports | 14,358 | 0.221 | 0.249 | 0.11 |
| Culture | 942 | 0.098 | 0.144 | 0.32 |
| Politics | 1,013 | 0.083 | 0.159 | 0.48 |
| Crypto | 938 | 0.114 | 0.217 | 0.48 |
| World | 527 | 0.107 | 0.214 | 0.50 |
| Geopolitics | 398 | 0.119 | 0.223 | 0.46 |
| Tech | 377 | 0.096 | 0.223 | 0.57 |
| Economy | 250 | 0.108 | 0.219 | 0.51 |
Sports: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.4% | 1.7% | 291 | reliable |
| 5–15% | 9.8% | 13.8% | 669 | happened more often |
| 15–35% | 26.1% | 31.5% | 2,878 | happened more often |
| 35–65% | 49.1% | 48.8% | 8,420 | reliable |
| 65–85% | 72.7% | 67.2% | 1,558 | happened less often |
| 85–95% | 89.8% | 85.4% | 364 | happened less often |
| 95–100% | 98.4% | 96.1% | 178 | reliable |
Culture: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.2% | 1.5% | 400 | reliable |
| 5–15% | 9.6% | 16.8% | 220 | happened more often |
| 15–35% | 21.2% | 20.1% | 154 | reliable |
| 35–65% | 45.4% | 35.4% | 113 | happened less often |
| 65–85% | 75.9% | 80.8% | 26 | too few cases |
| 85–95% | 89.9% | 100.0% | 16 | too few cases |
| 95–100% | 97.9% | 100.0% | 13 | too few cases |
Politics: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.2% | 1.8% | 436 | reliable |
| 5–15% | 9.5% | 10.4% | 221 | reliable |
| 15–35% | 21.3% | 20.3% | 153 | reliable |
| 35–65% | 47.5% | 44.6% | 92 | too few cases |
| 65–85% | 76.6% | 81.6% | 49 | too few cases |
| 85–95% | 90.3% | 88.2% | 34 | too few cases |
| 95–100% | 97.5% | 100.0% | 28 | too few cases |
Crypto: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.9% | 3.7% | 242 | reliable |
| 5–15% | 9.3% | 5.3% | 133 | could be chance |
| 15–35% | 23.0% | 20.5% | 220 | reliable |
| 35–65% | 49.7% | 50.8% | 183 | reliable |
| 65–85% | 75.5% | 78.5% | 65 | too few cases |
| 85–95% | 91.4% | 98.0% | 51 | too few cases |
| 95–100% | 97.4% | 100.0% | 44 | too few cases |
World: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.2% | 2.4% | 207 | reliable |
| 5–15% | 9.1% | 14.5% | 83 | too few cases |
| 15–35% | 24.3% | 33.8% | 74 | too few cases |
| 35–65% | 48.9% | 54.5% | 66 | too few cases |
| 65–85% | 76.8% | 81.0% | 42 | too few cases |
| 85–95% | 90.1% | 91.4% | 35 | too few cases |
| 95–100% | 97.3% | 100.0% | 20 | too few cases |
Geopolitics: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.4% | 2.6% | 152 | reliable |
| 5–15% | 9.0% | 20.3% | 69 | too few cases |
| 15–35% | 24.7% | 37.3% | 51 | too few cases |
| 35–65% | 48.0% | 55.8% | 52 | too few cases |
| 65–85% | 76.2% | 90.3% | 31 | too few cases |
| 85–95% | 90.1% | 87.1% | 31 | too few cases |
| 95–100% | 96.9% | 100.0% | 12 | too few cases |
Tech: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.0% | 2.2% | 137 | reliable |
| 5–15% | 9.5% | 10.0% | 50 | too few cases |
| 15–35% | 24.1% | 29.6% | 54 | too few cases |
| 35–65% | 48.8% | 45.8% | 48 | too few cases |
| 65–85% | 76.4% | 82.8% | 29 | too few cases |
| 85–95% | 90.6% | 91.2% | 34 | too few cases |
| 95–100% | 97.8% | 100.0% | 25 | too few cases |
Economy: by price range
| Price range | Average price | Happened | Cases | Reading |
|---|---|---|---|---|
| 0–5% | 1.0% | 1.0% | 103 | reliable |
| 5–15% | 9.1% | 26.9% | 26 | too few cases |
| 15–35% | 25.9% | 24.4% | 41 | too few cases |
| 35–65% | 49.2% | 65.6% | 32 | too few cases |
| 65–85% | 73.8% | 77.3% | 22 | too few cases |
| 85–95% | 92.2% | 90.0% | 10 | too few cases |
| 95–100% | 97.9% | 100.0% | 16 | too few cases |
Rows behind the figures
- Counted
- 17,347
- Resolution date still in the future
- 197
- No confirmed result
- 1,006
- Short up-or-down series
- 11,265
- Too little traded
- 87,739
- No price seven days before
- 17,203
- Launch price of exactly 50%
- 3,256
Computed Oct 19, 2026. Download every cell as CSV
What is measured
For every resolved Polymarket market in our corpus we keep the price seven days and one day before resolution, and the result. Markets are grouped by topic and by price range, and for each group the average price is set against how often the event actually happened.
If prices were perfectly calibrated, markets priced at 25 percent would resolve YES about a quarter of the time. Where the observed rate differs from the average price, that difference is the finding — in either direction.
Which markets are left out
A market only counts when its price was a judgment. The rows below are excluded, and the counts above show how many each rule removed:
- Markets with less than $10,000 traded over their lifetime — a price hardly anyone traded against says little.
- Prices of exactly 50 percent. For most markets that is the launch price, recorded before any trading, not an assessment. A real market that happened to sit at exactly 50 percent is lost with them; that is the smaller error.
- Short up-or-down series that resolve within hours or a day.
- Markets without a confirmed result, and markets whose recorded resolution date still lies in the future.
How to read a difference
A group is reported as running high or low only when it holds at least 100 cases, the difference is at least three percentage points, and the 95 percent interval around the observed rate excludes the average price. A difference of under three points is shown as reliable; a larger one whose interval still includes the price is shown as one that could be chance; a group under 100 cases as too small to say — none of them as an edge.
The Brier score measures the prices as forecasts: the average squared distance between price and result. It is shown next to the score of simply forecasting the base rate every time. Skill above zero means the prices carried information beyond that base rate.
The limit of this number
This is a historical frequency for groups of markets. It is not a probability for any single market and not advice. It describes the markets our corpus holds, beginning in June 2026; a change in who trades or what gets listed can change it.
Sports games make up the largest share of the corpus, so the figures across all topics lean towards them. The per-topic tables are there so that a politics market is not read against a basketball game.