PunditRivalPundit RivalAI football predictions
Log inRegister free

Open data · free to cite

AI vs bookmakers: 10,930 matches scored against Pinnacle

Prediction sites rarely publish a score. This page does, on the benchmark that settles the argument: Pinnacle's closing line, the sharpest price in football. Every number below comes from a walk-forward test of our Dixon-Coles model over 10,930 matches in nine European leagues, 2022-08-05 to 2026-01-14. The market won. We publish it anyway, with the per-league breakdown and the method, because a track record that only appears when it flatters the model is not a track record.

Key findings

  • Over 10,930 matches in nine European leagues (2022-08-05 to 2026-01-14), the model's 1X2 log loss was 0.9866 against 0.9628 for Pinnacle's closing line. The market is sharper by 0.0239.
  • The model's most likely outcome came true 52.3% of the time, against 53.9% for Pinnacle's closing favourite and 44.1% for always backing the home side.
  • The Championship is the hardest league in the sample to price: even Pinnacle's closing favourite won only 46.8% of the time there, against 58.9% in the Primeira Liga.
  • The model gets closest to the market in the Championship (0.0178 behind) and is furthest away in the Bundesliga (0.0314).

Model vs Pinnacle's closing line, by league

1X2 log loss, lower is better. “Gap” is how far the model sits behind the closing line, so a smaller gap means the model is closer to the market. Sorted by gap.

LeagueMatchesModelPinnacle closeGapModel hitPinnacle hit
Championship · England1,8941.05221.0343+0.017845.6%46.8%
Scottish Premiership · Scotland7320.95220.9320+0.020254.9%55.6%
Serie A · Italy1,3290.99090.9679+0.023052.4%53.3%
La Liga · Spain1,3190.97980.9565+0.023353.2%55.2%
Premier League · England1,3410.97450.9503+0.024253.2%55.7%
Ligue 1 · France1,1391.00120.9769+0.024352.4%53.6%
Primeira Liga · Portugal1,0630.92040.8945+0.025956.9%58.9%
Eredivisie · Netherlands1,0530.94880.9201+0.028754.7%57.1%
Bundesliga · Germany1,0601.00020.9688+0.031452.5%54.6%
All nine leagues10,9300.98660.9628+0.023952.3%53.9%

Is the model calibrated?

A probability is honest when the things given a 30% chance happen about 30% of the time. That, not hit rate, is the test a forecast has to pass.

Model saidOutcomesAverage chance givenActually happened
0% – 10%1,1137.0%7.5%
10% – 20%4,47015.9%15.4%
20% – 30%12,49425.3%26.0%
30% – 40%5,46234.4%33.5%
40% – 50%3,91044.8%44.7%
50% – 60%2,62754.7%53.4%
60% – 70%1,47464.6%65.1%
70% – 80%84074.6%74.8%
80% – 100%40085.1%84.2%

Method, in full

  • Walk-forward, never in-sample. Each league is refitted once per calendar week on the previous two seasons of that league only, then predicts that week's fixtures. Nothing after a match influences its own prediction.
  • The model. A Dixon-Coles goal model: attack and defence ratings per team with recent matches weighted more heavily, turned into home, draw and away probabilities.
  • The benchmark. Pinnacle's closing 1X2 price with the bookmaker's margin removed, so model and market are compared as probabilities that sum to one. Pinnacle is used because it is the sharpest publicly available line in football.
  • The score. Multiclass log loss on the actual result of every match in the sample, plus hit rate and calibration. Log loss punishes confident wrong calls, which is why it is the honest score for a probability.
  • Sample. 10,930 matches, 2022-08-05 to 2026-01-14, in the Premier League, Championship, La Liga, Serie A, Bundesliga, Ligue 1, Eredivisie, Primeira Liga and Scottish Premiership. Every match with a Pinnacle closing price is in it; none are dropped afterwards.
  • Live predictions are separate. This page is the historical walk-forward test. Predictions published going forward are frozen before kick-off, hashed and kept in the public archive with their results, so the live record can be checked independently of this backtest.

Use this data

These figures are free to republish, including tables and charts built from them, under CC BY 4.0. No permission needed and no need to contact us; attribution with a link is the only condition.

Suggested citation

PunditRival, "AI vs Bookmakers: 10,930 matches scored against Pinnacle's closing line" (2026-11-19), https://punditrival.com/research/ai-vs-bookmakers

Download the per-league table as CSV · last updated 2026-11-19. How the model works in detail: method.

Questions

Does the model beat the bookmakers?
No. Over 10,930 matches its 1X2 log loss was 0.9866 against 0.9628 for Pinnacle's closing line, so the market was sharper by 0.0239. Almost no public model beats a sharp closing line; the useful questions are by how much it trails and whether its probabilities are calibrated.
How accurate are football prediction models?
In this 10,930-match test the model's most likely outcome came true 52.3% of the time, against 53.9% for Pinnacle's closing favourite and 44.1% for simply backing the home side every week. Good 1X2 models land roughly in the low 50s; anything claiming far more is usually counting only the games it chose to call.
Which league is hardest to predict?
The Championship. Even Pinnacle's closing favourite won only 46.8% of the time there, against 58.9% in the Primeira Liga, and its closing-line log loss (1.0343) is the highest in the sample.
Can I republish these numbers?
Yes, under CC BY 4.0, with attribution and a link to this page. No permission is needed and the per-league table is downloadable as CSV.