Accuracy tracker
We hold our own model to account. Here's how often Footyntel's match predictions have been right this season — scored honestly against real results. No other free football model publishes its report card.
Most-likely-outcome accuracy 49.9% vs 43.7% for always backing the home win. Brier 0.6 beats the 0.667 of a no-skill model (lower is better).
By league
| League | Matches | Accuracy | Brier |
|---|---|---|---|
| Brasileirão Série A | 299 | 51.5% | 0.6 |
| Championship | 120 | 43.3% | 0.6 |
| La Liga | 79 | 51.9% | 0.6 |
| Eredivisie | 72 | 54.2% | 0.6 |
| Serie A | 66 | 56.1% | 0.5 |
| Primeira Liga | 63 | 54% | 0.6 |
| Premier League | 60 | 40% | 0.7 |
| Ligue 1 | 54 | 44.4% | 0.6 |
| Bundesliga | 45 | 51.1% | 0.6 |
Calibration
When the model's most likely outcome is given an X% chance, how often does it actually happen? Points on the dashed line are perfectly calibrated — above it the model is too cautious, below it too confident.
| Model confidence | Predictions | Actual hit rate |
|---|---|---|
| 30–40% | 108 | 37% |
| 40–50% | 329 | 44.7% |
| 50–60% | 201 | 48.8% |
| 60–70% | 123 | 56.1% |
| 70–80% | 60 | 71.7% |
| 80–90% | 34 | 82.4% |
How this works
For every completed match we rebuild the model from scratch using only the data available before that matchday — a walk-forward backtest, so there's no hindsight. We take the pre-match win / draw / win probabilities and compare them to what actually happened. Accuracy is how often the single most likely outcome occurred. The Brier score grades the full set of probabilities (0 = perfect, 0.667 = a no-skill model guessing 1/3 each). The sample grows automatically as new matches finish.
It's a statistical model, not betting advice. Football is high-variance — even a strong model is "wrong" often, which is exactly why we publish probabilities, not certainties.
FAQ
How accurate is the Footyntel model?
Across 858 backtested matches this season, the model's most-likely outcome was correct 49.9% of the time, versus 43.7% for always picking the home win. Its Brier score is 0.6 (lower is better; a no-skill 1/3-each model scores 0.667).
How is accuracy measured?
For every finished match we rebuild the model using only data available before that matchday (a walk-forward backtest with no hindsight), take its pre-match win/draw/win probabilities, and compare to the actual result. Accuracy counts how often the single most likely outcome happened; the Brier score grades the full probability calibration.