How accurate are prediction markets?
Every prediction market quotes a probability. At Vectorial Data we record every resolved Yes/No Polymarket market with at least $100,000 traded, with its price one day and one week before its scheduled end, and check how often it came true. Free to cite and download.
In Vectorial Data's open dataset of 15,587 resolved Polymarket markets (Nov 1, 2025 – Nov 20, 2026), events priced at 70–80% one day before their scheduled end happened 74% of the time (n=518).
These markets are live on Vectorial Data Terminal
Trade prediction markets on politics, sports, crypto and the economy from your own wallet, with a plain-language summary of every market and the odds history next to it.
Calibration plot
Each bubble = one 10-point band of forecasts. Bubble size = number of markets in that band. Dashed diagonal is perfect calibration. Bubbles below the line = overconfident; above = underconfident.
Patterns
Of 8,434 markets priced at 80% or more (or 20% or less) one day before their scheduled end, 7,929 resolved that way.
Sports (0.199) was the hardest to call; Tech & AI (0.019) the easiest.
Of 15,587 markets, the side priced at 50% or more one day before the scheduled end won 12,312 times.
Why some analyses report much better numbers
3,801 markets in the same period closed more than a day before their scheduled end, usually because the event had already happened. Measured 24 hours before their actual close, as many analyses do, their Brier score is 0.054, and 72% were already within 5 points of the result. Counting them would take our score from 0.133 to 0.117 (measured Nov 8, 2026). We measure against the scheduled end, a time fixed in advance, and leave them out.
Near-certain markets pull the average down: a long shot priced at 1–2% almost never happens and adds almost nothing to the error. In one widely shared analysis of 7,661 markets, three in four sat in the 0–5% band; in ours, 40% are below 10%. That's why the calibration bands above matter more than one average: they show accuracy where the outcome was actually in doubt.
Brier by category
| Category | Markets | Brier (1 day) | Live markets |
|---|---|---|---|
| ⚽Sports | 8,919 | 0.199 | Live markets → |
| 🏛️Politics | 2,572 | 0.061 | Live markets → |
| 🎬Culture | 1,855 | 0.051 | Live markets → |
| ₿Crypto | 1,016 | 0.027 | Live markets → |
| 🤖Tech & AI | 617 | 0.019 | Live markets → |
| 📈Economy | 492 | 0.031 | Live markets → |
| 🌍Other | 64 | 0.064 | Live markets → |
| 🌡️Climate | 52 | 0.063 | Live markets → |
Recent resolutions
Use the odds, not just read them
On Vectorial Data Terminal every live market shows its odds history, a plain-language summary and the resolution rules. Your funds stay in your own wallet.
Download and cite
- Every market of 2026, one row each: CSV · JSON
- Every market of 2025, one row each: CSV · JSON
- Summary (this page as data): /api/calibration
- License: CC BY 4.0. Free to use with attribution.
- Cite as: Vectorial Data Terminal, “Polymarket calibration dataset”, terminal.vectorialdata.com/metrics.
Methodology
- Universe: Polymarket markets whose outcomes are Yes and No (not team-vs-team or over/under), resolved YES or NO, with at least $100,000 traded and open at least 7 days before their scheduled end.
- Reference time: the market's scheduled end date, a time known in advance.
- Excluded: markets that closed more than a day before their scheduled end. They close because the result is already known, so their last-day price already reflects it. This also leaves out early YES outcomes of “by this date” questions.
- Forecast: the YES price from Polymarket's hourly price history closest to 24 hours (and 7 days) before the reference time, within 6 hours.
- Brier score: the mean of (forecast − outcome)², where outcome is 1 for YES and 0 for NO. 0 is perfect; always saying 50% scores 0.25.
- Calibration: markets grouped by forecast in 10-point bands, comparing the average forecast with how often the event happened.
- Category: Polymarket's own section tags (Politics, Sports, Crypto…), or the question text when a market has none.
- Updated daily. Each market enters once, after it resolves. Downloads are split by year of the reference time.