Accuracy, explained
How accurate are AI sports predictions? 10,867 graded matches
Accuracy claims are cheap. This is the full graded record of one public prediction model — every league, every sport — plus the test that matters more than the hit rate: whether the probabilities mean what they say.
Numbers recalculated 19 Oct 2026 from graded fixtures
Short answer
Hit rate by sport
Football is graded on three outcomes (home, draw, away), so a draw counts as a miss unless the model picked it. Baseball, basketball and ice hockey have no draw, which is why their numbers are not directly comparable with football.
| Sport | Graded matches | Correct | Hit rate |
|---|---|---|---|
| Baseball | 5,399 | 2,961 | 55% |
| Football | 2,639 | 1,252 | 47% |
| Ice hockey | 1,518 | 823 | 54% |
| Basketball | 1,311 | 877 | 67% |
| League | Graded | Hit rate | Last 30 days | Strong picks |
|---|---|---|---|---|
| MLB | 3,643 | 55% | 59% (156) | 56% (1,783) |
| NHL | 1,518 | 54% | 60% (126) | 58% (665) |
| NBA | 1,311 | 67% | — | 78% (464) |
| KBO League | 921 | 54% | 63% (75) | 53% (239) |
| NPB | 835 | 56% | 57% (80) | 59% (431) |
| MLS | 521 | 46% | 48% (75) | 38% (8) |
| Premier League | 444 | 46% | 50% (28) | 100% (7) |
| LaLiga | 408 | 50% | 54% (28) | 100% (15) |
| Serie A | 392 | 45% | 52% (27) | 78% (9) |
| Ligue 1 | 320 | 48% | 69% (26) | 100% (3) |
| Bundesliga | 312 | 49% | 58% (26) | 57% (14) |
| Champions League | 136 | 54% | 72% (18) | 100% (3) |
| K League 1 | 106 | 42% | 39% (18) | 0% (1) |
Strong picks are matches where the top probability clears a league-specific threshold. Figures in brackets are sample sizes.
A 50/50 match stays 50/50
A useful prediction is not one that is always confident — it is one whose probabilities can be trusted. If a model says 60%, the favourite should win about six times in ten. We group every graded match by how confident the model was in its pick and check how often that pick actually won.
| Model confidence | Average stated | Actually won | Matches |
|---|---|---|---|
| 30%–40% | 37% | 34% | 588 |
| 40%–50% | 45% | 46% | 1,432 |
| 50%–60% | 54% | 55% | 4,511 |
| 60%–70% | 65% | 56% | 2,286 |
| 70%–80% | 74% | 60% | 1,348 |
| 80%+ | 86% | 72% | 685 |
Confidence = probability of the outcome the model picked. Buckets with fewer than 20 matches are hidden.
The closer the "actually won" column sits to the "average stated" column, the better calibrated the model is. In 40%–50%, 50%–60% the two columns are within three points of each other — in those bands the stated probability really does match how often the pick wins. Outside those bands the model has been overconfident: in 30%–40%, 60%–70%, 70%–80%, 80%+ its picks won less often than stated, by as much as 14 points (80%+). Favourites still win most of the time, just not as often as the raw number suggests.
On the 3,774 matches where bookmaker odds were available, the model's Brier score was 0.527 against 0.512 for the market (lower is better). The market is still the sharper forecaster on these matches — which is exactly why the model blends market odds into its own estimate.
Why 90% accuracy claims don't add up
Promises of 85% or 90% on single games should be checked against a full record. In this record, even the picks the model was most confident about (80%+) won 72% of the time. A record you can check — with misses left in — is worth more than any headline number.
How the record is kept
- Every prediction is timestamped before kick-off; anything created later is excluded.
- Results are graded automatically once final. Postponed or abandoned games are not counted.
- Nothing is deleted or re-scored. The live numbers are on the accuracy page.
The method behind the probabilities is described on the methodology page, and the full benchmark, including calibration curves, is published by Scorebase.