Are Football Predictions Accurate? What 10,000+ Matches Tell Us
How to measure the accuracy of football predictions (calibration, log loss, closing line), our own model's tested record, and the warning signs of a record you cannot verify.
Updated 2026-11-13 · PunditRival
Short answer
A good football model is well calibrated but does not beat the sharpest betting market. Tested walk-forward on 10,930 matches in 9 European leagues (2022-08-05 to 2026-01-14), PunditRival's Dixon-Coles model put probabilities within 1.3 percentage points of the observed frequency in every probability band. Its 1X2 log loss was 0.9866, against 0.9628 for Pinnacle's closing prices (lower is better) and 1.0986 for a blind one-third guess on every outcome. So the market was more accurate than the model.
| Predicted band | Outcomes | Avg predicted | Happened |
|---|---|---|---|
| 0.0–0.1 | 1,113 | 7.0% | 7.5% |
| 0.1–0.2 | 4,470 | 15.9% | 15.4% |
| 0.2–0.3 | 12,494 | 25.3% | 26.0% |
| 0.3–0.4 | 5,462 | 34.4% | 33.5% |
| 0.4–0.5 | 3,910 | 44.8% | 44.7% |
| 0.5–0.6 | 2,627 | 54.7% | 53.4% |
| 0.6–0.7 | 1,474 | 64.6% | 65.1% |
| 0.7–0.8 | 840 | 74.6% | 74.8% |
| 0.8–1.0 | 400 | 85.1% | 84.2% |
Walk-forward = each prediction only used results from before that match. Source: PunditRival backtest, full method on the Method page. Probabilities are estimates, never guarantees.
Hit rate is not accuracy
A tipster who only picks 1.20 favourites can show an 80% hit rate and still lose money. Hit rate depends on what you choose to predict. Two better measures are calibration (do 60% calls come true 60% of the time?) and log loss (how much probability was put on what actually happened, on every match, not just the chosen ones).
Our model's tested record
We tested our model walk-forward on more than ten thousand matches in nine leagues: each week it only saw results from before that week. It is well calibrated: in every probability band the predicted and observed frequencies are close (see the table in the short answer above).
We also compared it with the sharpest betting market. The market was slightly more accurate. That is normal: liquid betting markets absorb team news and huge amounts of information. We publish that comparison on our Method page rather than hiding it.
Warning signs
The warning signs are about proof, not presentation: calls you can only see after the match, no timestamp or hash fixed before kick-off, a record that shows selected wins instead of every call, hit rates with no odds or probabilities attached, 'sure win' or 'guaranteed' language, and leaderboards or testimonials you cannot check.
What to look for instead: every prediction committed before kick-off with a timestamp or hash an outsider can check, every call counted, and the full history open to recompute. Our calls in the 9 model leagues and the three UEFA club competitions are hashed and chained before kick-off for exactly this reason (other competitions are posted before kick-off without a hash); the related checklist guide below goes through each check.
Questions people ask
Can any model beat the bookmakers?+
Rarely, and only by small margins. The sharpest bookmakers are very well calibrated. Treat any claim of large, steady profits with suspicion unless the full record is verifiable.
How can I check a prediction was made before kick-off?+
On PunditRival every prediction in the 9 model leagues and the Champions League, Europa League and Conference League is hashed and chained before kick-off; the Verify page lets you recompute the hash yourself after the salt is published.
Related guide: How to check a football prediction site's track record (6-point checklist)