How our predictions work
No secret sauce, no "insider tips", no 92% win-rate claims. A classical statistical model, fitted on years of results, published with its own scorecard —the track record — attached.
The model
Our probabilities come from a Dixon-Coles model (Dixon & Coles, 1997): each team gets an attack and a defence rating, goals follow Poisson distributions built from those ratings plus home advantage, and a correction term fixes the well-known bias in low-scoring results (0-0, 1-0, 0-1, 1-1). From the resulting score matrix we read off every market we publish: match result (1X2), Over/Under 2.5 goals, and both teams to score.
Recent matches matter more: the fit weights each result by an exponential time decay, so a team's rating tracks what it is now, not what it was two seasons ago. Alongside it we maintain a football-adjusted Elo rating as a second, independent signal — when the two disagree about a match, the shown confidence drops.
The data
- Fixtures, results and standings: football-data.org, refreshed three times a day.
- Five seasons of historical results per league (2021-22 to 2025-26) to calibrate ratings before day one.
- Historical closing odds, used only to benchmark the model in backtests — never shown as our own prediction.
The honest numbers
Before launch we ran a walk-forward backtest over three full seasons (2023-2026, 2,071 predictions across the Premier League and Primeira Liga): refit the model weekly on everything known before each cutoff, predict the following week, grade against reality.
| 1X2, 2,071 matches | Accuracy | Brier score ↓ |
|---|---|---|
| Our model (dc-0.1.0) | 53.5% | 0.578 |
| Naive baseline (league frequencies) | — | 0.652 |
| Betting market (closing odds, de-vigged) | 55.7% | 0.555 |
Read that honestly: the model is clearly better than naive guessing, and clearly behind the closing-odds market — which aggregates the information of thousands of professionals and is the hardest benchmark in existence. We publish that gap instead of hiding it, and the live track record shows whether it closes.
Why we don't sell "value bets" (yet)
The same backtest showed that naively betting every apparent model edge against closing odds would have lost about 18% flat-stake. Apparent edges from an uncalibrated model are mostly miscalibration, not value. Until calibration demonstrably clears zero in simulation, we will not dress predictions up as betting advice.
Known limits
- Promoted teams start with league-average ratings — no top-flight history exists for them, so their early-season probabilities are the model's weakest.
- No injuries, suspensions or lineups in this version; those signals arrive later.
- Cup competitions are out of scope for now — the model is built for league play.
The audit trail
Every prediction is written to our database before kick-off with a model version and timestamp, and graded automatically after full time. The database role that writes them has no permission to update or delete a prediction — append-only by construction, not by promise.