The most common question we get at PredictWire is “do prediction markets actually work?” Until today our answer relied on the same handful of academic papers everyone else cites. Today we release the data ourselves: the PredictWire Calibration Archive, an open and reproducible measurement of Polymarket’s real-world forecasting accuracy across 1,914 resolved binary markets representing $30.48 billion in lifetime trading volume.
The archive is built from 194,111 daily price snapshots pulled directly from Polymarket’s public CLOB endpoint, joined to each market’s actual resolved outcome. It is not a curated highlight reel. It is every market that closed with a clean Yes/No resolution and at least $100,000 in volume. The methodology is on the page; the underlying CSV is downloadable; every number can be recomputed from the raw data.
The Headline Finding
When prediction markets are confident, they are right. Across 1,341 contracts that closed at 80% Yes-or-higher or 20%-or-lower, the market’s directional pick was correct 97.1% of the time. That is the single number we keep coming back to. Confident calls resolve in the market’s favor almost 24 times out of 25.
The Brier score across all 1,914 closing prices is 0.0865, against a naive 50/50 baseline of 0.25 and a base-rate baseline of 0.190. The implied skill score versus base rate is +0.544 — substantial. Volume-weighted, the Brier improves to 0.0665, which tells us high-volume markets are even better-calibrated than the unweighted average.
The Reliability Picture
A handful of decile buckets from the closing-price reliability table:
| Predicted | N | Mean Pred. | Actual Yes |
|---|---|---|---|
| 0-10% | 1,086 | 0.9% | 1.2% |
| 30-40% | 86 | 34.9% | 39.5% |
| 50-60% | 129 | 54.9% | 47.3% |
| 70-80% | 56 | 74.4% | 75.0% |
| 90-100% | 146 | 98.2% | 99.3% |
That is what calibration looks like in practice. Markets pricing 90-100% Yes resolved Yes 99.3% of the time. Markets pricing 0-10% Yes resolved No 98.8% of the time. The full ten-bucket table is on the /accuracy/ page.
What Markets Are Best At — and Worst At
Categorized, the dataset shows a clean rank-order from most-calibrated to least:
| Category | N | Brier | Directional |
|---|---|---|---|
| economy | 68 | 0.0064 | 100.0% |
| sports | 226 | 0.0339 | 95.1% |
| politics | 438 | 0.0419 | 93.2% |
| crypto | 182 | 0.0568 | 90.7% |
| geopolitics | 141 | 0.1205 | 82.3% |
| other | 856 | 0.1305 | 77.8% |
Economy markets are the best-calibrated class, with a Brier of 0.0064 and 100.0% directional accuracy across 68 contracts (mostly Fed decisions, where consensus converges to near-certainty). Other is the worst, at Brier 0.1305 and 77.8% directional. That is the catch-all category — niche sports lines, novelty contracts, and one-off events where the wisdom-of-crowds advantage is thinnest. Geopolitics — wars, peace deals, regime change — is the next-worst at Brier 0.12. Hard problems are hard, and the markets correctly know they’re hard.
The Counter-Intuitive Finding
The dataset reveals one finding that contradicts conventional wisdom: prediction markets are not most accurate at the moment of resolution. They are most accurate 90 to 180 days before resolution. The Brier score in the 0-1 day window is 0.10; the Brier in the 90-180 day window is 0.07. This isn’t a contradiction so much as a survivor-bias story — easy contracts effectively settle long before their nominal end date and stop contributing to the late-window snapshots, leaving only the residual hard markets. Read the late-stage Brier as “the markets that were genuinely contested to the wire.”
Why We Built This
The prediction market industry has always had a credibility problem. Every reviewer site claims their picks are “highly accurate”; nobody publishes the data. We are the first to publish ours. The full dataset, the full methodology, the full caveats — all on the /accuracy/ page. The CSV is downloadable. The metrics are reproducible. Send the worst confident calls to your friends; we did the work to find them.
PredictWire will refresh the archive quarterly. The next update — adding Kalshi resolved contracts and PredictIt’s archive — is scheduled for July 2026. We will publish a transparent diff of every metric between runs.
Read the Full Archive
- → Read the full PredictWire Calibration Archive — full reliability tables, horizon analysis, category breakdown, methodology, and CSV download.
- PredictWire Methodology — how we rate platforms.
- Best Prediction Markets — current platform rankings.
Trade These Markets
Disclosure: PredictWire earns a commission on qualifying accounts opened through the links below. Our rankings and reviews are not influenced by these relationships. Full disclosure.
- Trade on Polymarket → — every market in this archive is hosted there.
- Trade US-regulated alternatives on Kalshi →
About this article: Written and reviewed by The PredictWire Research Team under our Editorial Standards. Platform rankings follow our public Methodology. Prediction market contracts carry risk of total loss. Nothing here is financial advice. Corrections: corrections@predictwire.io.