Club football · AI probability model
Football AI predictions — match probabilities for every market
GoalPulse publishes pre-match probabilities for 87 upcoming fixtures across 15 competitions, blending a Dixon-Coles bivariate Poisson model with the bookmaker consensus — where prices exist, the published figures are the margin-free market consensus. Every market on a match page — 1X2, handicap, correct score, total goals, half-time/full-time — comes from one score-probability matrix. Out of sample (6249 matches, weekly walk-forward refits) the published forecast calls the 1X2 outcome correctly 51.4% of the time, with an RPS of 0.203, 11.2% better than league base rates.
Since 2026-11-07, matches with prices blend the model with the bookmaker consensus, and Brasileirão, J1 League, K League 1, MLS and the three UEFA club competitions are covered (see Methodology).
Data updated: · Market & China Sports Lottery guide · Open data JSON
Leagues
- Premier League8 upcoming#1 Arsenal +0.961X2 accuracy 48% · n=92
- La Liga8 upcoming#1 Barcelona +1.771X2 accuracy 56% · n=111
- Serie A7 upcoming#1 Inter +1.061X2 accuracy 63% · n=103
- Bundesliga8 upcoming#1 Bayern Munich +1.921X2 accuracy 53% · n=73
- Ligue 18 upcoming#1 Paris Saint-Germain +1.421X2 accuracy 56% · n=82
- Eredivisie8 upcoming#1 PSV +1.291X2 accuracy 52% · n=100
- Primeira Liga7 upcoming#1 Benfica +1.561X2 accuracy 54% · n=91
- EFL Championship8 upcoming#1 Middlesbrough +0.731X2 accuracy 51% · n=172
- Brasileirão Série A2 upcoming#1 Flamengo +1.161X2 accuracy 54% · n=339
- J1 League4 upcoming#1 FC Tokyo +0.871X2 accuracy 54% · n=136
- K League 14 upcoming#1 FC Seoul +0.371X2 accuracy 41% · n=206
- MLS15 upcoming#1 Vancouver Whitecaps FC +0.721X2 accuracy 50% · n=495
- Champions League0 upcoming#1 Bayern Munich +2.341X2 accuracy 68% · n=72
- Europa League0 upcoming#1 Benfica +1.321X2 accuracy 54% · n=72
- Conference League0 upcoming#1 Brighton +0.661X2 accuracy 59% · n=54
Upcoming match predictions
Saturday, November 7
AI picks: Borussia Dortmund · 2:1 · Over 2.5
AI picks: Bayern Munich · 1:2 · Over 2.5
AI picks: SC Paderborn/Eintracht Frankfurt · 1:1 · Over 2.5
AI picks: RB Leipzig · 2:1 · Over 2.5
AI picks: Arsenal · 2:0 · Over 2.5
AI picks: Birmingham City · 1:1 · Over 2.5
AI picks: West Ham United · 1:1 · Over 2.5
AI picks: Cardiff City/Burnley · 1:1 · Over 2.5
AI picks: Draw/Sheffield United · 1:1 · Over 2.5
AI picks: Fulham/Newcastle United · 1:1 · Over 2.5
AI picks: Southampton · 1:1 · Over 2.5
AI picks: QPR · 1:1 · Over 2.5
AI picks: Middlesbrough · 0:2 · Over 2.5
AI picks: West Brom/Draw · 1:1 · Over 2.5
AI picks: Espanyol/Deportivo La Coruña · 1:1 · Under 2.5
AI picks: Fortuna Sittard/Telstar · 1:1 · Over 2.5
AI picks: Bologna · 0:1 · Under 2.5
AI picks: Manchester City · 1:1 · Over 2.5
AI picks: Draw/Athletic Club · 1:1 · Under 2.5
AI picks: VfB Stuttgart · 1:1 · Over 2.5
AI picks: ADO Den Haag/Sparta Rotterdam · 1:1 · Over 2.5
AI picks: Alverca · 1:1 · Under 2.5
AI picks: Paris FC · 1:1 · Over 2.5
AI picks: Roma · 2:0 · Over 2.5
+12 more:MLS · 8Eredivisie · 1La Liga · 1Ligue 1 · 1Primeira Liga · 1
Model track record (public, reproducible)
- Matches
- 6249
- 1X2 accuracy
- 51.4%
- League base rates 43.7%
- RPS
- 0.203
- League base rates 0.229
- Skill vs baseline
- +11.2%
- RPS
On the same 5606 test matches: model alone RPS 0.2032, market consensus 0.2007, blend 0.2007 (lower is better).
From 2026-11-07 the record scores the published blend; earlier matches are back-filled with the closing prices that were available at the time.
2026-27 · Matches 2626 · 1X2 accuracy 52.0% · RPS 0.203
AI picks · Pick result
| Play | Pick result · n=6249 | 2026-27 · n=2626 |
|---|---|---|
| 1X2 (match result) | 68.6% | 68.5% |
| 1X2 (match result) · Confidence High | 72.4% | 72.9% |
| Handicap 1X2 | 51.3% | 51.2% |
| Correct score (top 2) | 23.4% | 21.9% |
| Total goals (top 2) | 46.3% | 44.1% |
| Half-time / full-time | 31.7% | 31.8% |
| Over / under 2.5 | 57.4% | 57.6% |
The top choice for each market — for reference only, no guarantee; any bet you place on it is at your own risk. For matches with bookmaker prices the pick equals the top option of the margin-free market consensus; without prices it is the statistical model.
Calibration
| Probability bucket | Predicted | Observed | n |
|---|---|---|---|
| 0–10% | 7.6% | 3.2% | 314 |
| 10–20% | 16.0% | 14.5% | 2197 |
| 20–30% | 25.7% | 26.1% | 7574 |
| 30–40% | 34.3% | 33.6% | 3656 |
| 40–50% | 44.7% | 45.5% | 2378 |
| 50–60% | 54.6% | 53.1% | 1411 |
| 60–70% | 64.5% | 67.7% | 770 |
| 70–80% | 74.4% | 79.5% | 356 |
| 80–90% | 83.4% | 84.4% | 90 |
Market guide
| Play | Options | GoalPulse probability |
|---|---|---|
| 1X2 (胜平负) | 3: home / draw / away | Match-page 1X2 bar; the three probabilities sum to 100%. |
| Handicap 1X2 (让球胜平负) | 3 | Handicap table with every integer line from −3 to +3, plus the model's most balanced line to compare with the published one. |
| Correct score (比分) | 31: 13 home-win cells (1:0…5:2 + other), 5 draw cells, 13 away-win cells | Three-column correct-score table matching the 31 Jingcai options. |
| Total goals (总进球) | 8: 0–6 and 7+ | Total-goals (0–7+) distribution. |
| HT/FT (半全场) | 9 combinations | 3×3 HT/FT grid, rescaled to agree with the 1X2. |
Method
- Model: Dixon-Coles (1997) bivariate Poisson with the low-score correction ρ, exponential time decay (faster for matches with closing prices, slower for score-only matches) and Gamma-prior shrinkage on attack/defence.
- Each country's top flight and second tier are fitted jointly, so a promoted club is rated from its second-tier results rather than a guess.
- One score matrix yields 1X2, handicap, correct score, total goals, HT/FT and goal parity; HT/FT splits goals by first-half share and is rescaled to match the 1X2.
- The record is out of sample: refit every Monday on earlier results, then predict that week. Hyper-parameters were chosen on 2024-25 only; the test set is all of 2025-26 plus 2026-27 to date.
- Since 2026-11-07 the published probabilities blend the statistical model with the bookmaker consensus: a logarithmic pool whose weight was fitted on 2024-25 closing prices and tested on 2025-26 onwards.
- The backtest is blunt: closing prices are already more accurate than our model, so the best weight goes almost entirely to the market. Where prices exist the page is essentially the margin-free consensus; where they do not (e.g. fixtures still weeks away) the model is used on its own.
- The blended 1X2 and over/under 2.5 are refitted into one Dixon-Coles score matrix, so handicap, correct score, total goals and HT/FT stay mutually consistent.
- Champions League, Europa League and Conference League clubs come from dozens of leagues, so cups use a cross-league goal Elo (every stored domestic and European match, in date order), then blend with the market.
- Since 2026-11-07 the team ratings learn not only from scores: the closing prices of past matches are turned into market-implied expected goals and fitted alongside them, plus corrections that Claude Fable 5.1 found by iterating on the validation season and that held up on a locked test set (data recency, league reliability, strength-gap and draw calibration). Out of sample the model alone improved from RPS 0.2087 to 0.2051, which mainly helps fixtures that have no prices yet.
Probabilities include the bookmaker consensus, which already reflects public team news; the statistical-model part uses results, scores and the closing prices of past matches, not team news directly. Check official club news before kick-off.
Claude pre-match intel is a public shadow test and changes no published probability: 93 matches recorded, 16 scored after full time. The comparison with the market consensus is published once 150 matches are scored (rule fixed in advance).