Brier score
0.147
Live evaluations · N=10,558 · as of 1 Jun 2026
Lower is better · vs 0.222 uniform baseline
Model Track Record
MatchMind publishes how its calibrated football model actually performs against real outcomes. This page is public — no signup required — because trust in a probability model has to be earned with numbers, not promises.
Log-loss
1.055
Live evaluations · N=10,558 · as of 1 Jun 2026
Lower is better · vs 1.099 uniform baseline
Calibration error (ECE)
0.039
Validation set
Lower is better
Model version
v20260521022940
Active champion
Calibration — Reliability Diagram (home-win 1×2)
n = 10,558 · last 365 daysEach dot is a probability bin — x-position is the average predicted home-win probability, y-position is how often the home team actually won. The dashed diagonal is perfect calibration. Dots above = model is under-confident in that range; below = over-confident.
Performance Breakdown
How the model has performed across leagues, confidence levels, and predicted outcomes — all derived from the same pre-match evaluation set.
By Competition
| League | N | Accuracy | Brier | Log-loss |
|---|---|---|---|---|
| Premier League | 864 | 20.1% | 0.1601 | 0.6750 |
| La Liga | 1,510 | 34.6% | 0.1565 | 0.7042 |
| Bundesliga | 686 | 43.7% | 0.1625 | 0.6797 |
| Serie A | 1,047 | 54.6% | 0.1600 | 0.6208 |
| Ligue 1 | 893 | 41.5% | 0.1608 | 0.6230 |
By Confidence Bucket
Derived from the model's highest predicted probability: High ≥ 60%, Medium 45–60%, Low < 45%.
High confidence
37.9%
accuracy
n = 2,668
Brier 0.1884
Medium confidence
38.2%
accuracy
n = 1,539
Brier 0.1312
Low confidence
43.1%
accuracy
n = 793
Brier 0.1168
Precision by Predicted Outcome
When the model's highest-probability pick was Home / Draw / Away — what fraction of those turned out correct.
Home win
46.5%
precision
predicted 860×
Draw
22.5%
precision
predicted 80×
Away win
37.5%
precision
predicted 4,060×
Breakdown covers the last 365 days of pre-match evaluations (n = 5,000). Slices with fewer than 30 samples are shown as insufficient. Accuracy = fraction where model's highest-probability class matched the actual outcome.
How to read this page
- Brier score measures the squared error of probabilistic predictions. Perfect = 0; predicting 1/3 each every time = 0.222. The model has to beat that baseline to be useful.
- Log-loss penalises confident wrong calls more heavily. Predicting 1/3 each gives log-loss = 1.099.
- ECE (Expected Calibration Error) measures how close stated probabilities are to observed frequencies. Lower means a stated 60% actually means 60%.
- Live vs validation: when we have fewer than 50 live pre-match evaluations, this page shows training-validation metrics. The card above tells you which source the numbers come from.
Analytical estimates only.The numbers on this page describe the model's historical performance — they do not guarantee future outcomes, and MatchMind does not provide betting tips, picks, or guaranteed predictions. Past calibration is the best evidence available for trusting probabilities, but it is not a promise.