Tech & Science
Tech & Science accuracy
Every settled Tech & Science market gets scored against the price it was charging an hour before resolution. The numbers below are scoped to this category — for the full catalog view, head back to the index.
Events scored
17,977
Settled markets with a usable pre-settlement snapshot.
Favorite hit rate
90.1%
How often the pre-settlement favorite was the actual winner.
Brier score
0.0208
Mean squared probability error. Lower is more accurate. 0.25 is a coin flip.
91.7% better than random guess
Calibration
Markets that priced an outcome inside each 10pp band — how often did it actually happen? In-line means the observed rate landed inside the bucket (well calibrated).
By platform
How each platform’s pre-settlement pricing held up against the actual results, counted only on markets that at least two platforms priced. Scored against a whole book this would measure which categories a venue lists rather than how well it prices them — crypto resolves close to a coin flip and dominates some books entirely.
| Platform | Events scored | Favorite hit rate | Brier score |
|---|---|---|---|
| Kalshi | 979 | 87.5% | 0.0073 |
| Polymarket | 835 | 91.9% | 0.0900 |
| Gemini | 625 | 94.1% | 0.0156 |
| Polymarket US | 207 | 66.7% | 0.0733 |
| Novig | 34 | 79.4% | 0.1126 |
| Azuro | 5 | 100.0% | 0.0162 |
| Limitless | 2 | 100.0% | 0.1752 |
How we score this
- Corpus: Settled markets with both a settlement timestamp and a winning outcome. Voided and canceled markets are excluded.
- Snapshot: For each market we read the price an hour before settlement. The one-hour buffer keeps the settlement spike — where the eventual winner pre-prints to near 1.0 in the final minutes — out of the scoring. This instant is relative to settlement, not to kickoff: for a sports fixture it usually lands after the game has been played, while the venue is still waiting to resolve. Read these figures as a record of how markets priced outcomes, not as a forecast you could have placed a bet on.
- Favorite: The outcome with the highest cross-platform consensus price at snapshot time. Ties broken by stable canonical id. The "favorite hit rate" is the share of markets where the favorite was the actual winner.
- Brier score: Per market, the mean squared distance between the snapshot probability and the realized result (1 for the winner, 0 for losers). Overall Brier is the mean across markets. 0.0 is perfect; 0.25 is a coin flip on a binary market. Each market’s prices are normalized to sum to 1 before scoring: venues quote different things — an executable ask on some, a book midpoint on others — so scoring the raw quote would grade a venue on its pricing convention rather than its judgement. Markets where an outcome went unquoted carry no Brier, because a partial book cannot be normalized.
- Per-platform: Each platform is scored against its own pricing, normalized the same way, and only on markets where it had an opinion on the eventual winning outcome — so neither coverage gaps nor a venue’s quoting convention penalize accuracy. The comparison counts only markets at least two platforms priced: measured across a whole book it would rank which categories a venue lists rather than how well it prices them, since some question types resolve close to a coin flip and make up most of some books. Platforms with few contested markets are published but marked provisional rather than hidden.
- Calibration: Each canonical outcome’s normalized probability is binned into one of ten 10pp-wide buckets. For each bucket we publish the observed hit rate — the share of outcomes in that bucket that actually won — measured against the average probability the bucket predicted rather than the bucket’s midpoint, since predictions do not sit at the middle of their bin. A well-calibrated market lands on what it predicted; persistent skew up or down points at systematic over- or under-confidence.