How SportScore predicts football matches
Every SportScore prediction comes from one statistical model, measured openly against real results. Here is exactly how it works, what it looks at, and how accurate it has actually been.
1What the model predicts
For each upcoming fixture the model produces a full pre-match picture, not just a single tip:
2Cara kerjanya
- Data. The model reads every finished match in each league — thousands of games of history, in many leagues going back decades — with goals, home/away, dates and recent form.
- Model. A time-weighted Dixon–Coles model estimates each team's attacking and defensive strength plus home advantage, weighting recent matches more heavily and applying the Dixon–Coles correction for low-scoring results. It is refit continuously as new results come in.
- Scoreline grid. Those strengths produce expected goals for each team, which become a coherent grid of every scoreline's probability. The win/draw/win, over/under, both-teams-to-score and exact-score numbers are all read from that one grid, so they always agree with each other.
- No peeking (walk-forward). To measure it honestly, the model is tested walk-forward: for every past matchday it is refit using only the games that had already happened, then graded on what it could not see. There is no hindsight in the numbers below.
3How accurate it has been
Walk-forward backtest across 9 leagues, 2,210 graded matches · September 2025 – November 2026.
Hit rate by confidence
The higher the probability the model gives its pick, the more often that pick wins — which is exactly what you want from an honest model.
| When the pick is rated | Picks | Menang | Hit rate |
|---|---|---|---|
| 55%+ menang | 952 | 645 | 68% |
| 60%+ menang | 711 | 511 | 72% |
| 65%+ menang | 515 | 391 | 76% |
| 70%+ menang | 380 | 292 | 77% |
| 75%+ menang | 273 | 219 | 80% |
| 80%+ menang | 176 | 141 | 80% |
Calibration: what it said vs what happened
Grouping every backtested probability by what the model said, against how often those outcomes actually occurred. Close numbers mean the probabilities are trustworthy.
Berdasarkan liga
| Liga | Pertandingan | Pick acc. | Baseline | RPS | Brier |
|---|---|---|---|---|---|
| Internationals | 346 | 63% | 45% | 0,162 | 0,491 |
| Portuguese Primera Liga | 237 | 56% | 43% | 0,19 | 0,562 |
| Netherlands Eredivisie | 239 | 52% | 40% | 0,196 | 0,594 |
| Italian Serie A | 240 | 57% | 44% | 0,198 | 0,555 |
| Spanish La Liga | 237 | 54% | 49% | 0,202 | 0,577 |
| Bundesliga | 238 | 55% | 45% | 0,202 | 0,59 |
| French Ligue 1 | 239 | 52% | 44% | 0,205 | 0,601 |
| Belgian Pro League | 239 | 55% | 44% | 0,206 | 0,581 |
| UEFA Europa League | 195 | 51% | 45% | 0,24 | 0,635 |
Pick accuracy is how often the single most likely outcome was correct; baseline is always backing the division's most common result. RPS and Brier score the full probability distribution (lower is better) — the standard measures for rating forecasts. These are the same numbers shown live on the AI Predictor. Updated 13 November 2026.
4Honest limits
- These are probabilities, not certainties. A 75% favourite still loses one time in four — that is the model working, not failing.
- Coverage varies. Predictions are strongest in major leagues with deep history; smaller or newer competitions are less reliable, and some fixtures are too close to call.
- No bookmaker edge is claimed. Beating the closing market is a much higher bar than being accurate, and we do not claim to do it. The numbers here measure accuracy and calibration, not profit.
- Not betting advice. Predictions are for information and entertainment only. 18+.
5Frequently asked
How accurate are SportScore's football predictions?
In walk-forward backtesting across 9 leagues and 2,210 matches, the model's top pick was correct about 55% of the time, against a 44% base rate. When it rated a team 70% or higher to win, that team went on to win 77% of the time. Accuracy is lower in less predictable leagues and for closely matched teams.
What model does SportScore use to predict matches?
A time-weighted Dixon-Coles model - a statistical model of football goal-scoring that estimates each team's attacking and defensive strength plus home advantage from years of results, weighting recent form more heavily. Those strengths become a grid of scoreline probabilities, which gives the win/draw/win, expected-goals, over/under and both-teams-to-score numbers.
Are SportScore's football predictions free?
Yes. Every prediction is free with no sign-up, and the same underlying data is available through SportScore's free API and MCP server.
Is this betting advice?
No. The predictions are statistical estimates for information and entertainment only, not betting advice. 18+.
How often are the predictions updated?
Continuously. The model refits as new results come in and regenerates predictions for upcoming matches through the day, so each page reflects the latest form.
What does a calibrated prediction mean?
It means the probabilities are honest: across many matches, teams given around a 70% chance really do win about 70% of the time. SportScore publishes its calibration table so anyone can check it.
6Sources & method
The core model follows Mark Dixon and Stuart Coles, “Modelling Association Football Scores and Inefficiencies in the Football Betting Market” (Journal of the Royal Statistical Society, Series C, 1997), the standard time-weighted Poisson approach to football scorelines. Match data is SportScore's own results database. The backtest metrics (log-loss, Brier score and the ranked probability score) are standard scoring rules for probabilistic forecasts. The figures on this page are read directly from the live prediction engine, so they match the AI Predictor.
SportScore predictions are generated by a statistical model for information and entertainment only — not betting advice. 18+. If gambling affects you or someone you know, please seek help from a local support service.