Model record · independent validation
How accurate are our football predictions?
On 1,140 Premier League matches it had not seen, the model's most likely result happened 51.8% of the time. On the same matches, simply backing the club higher in the table was right 49.0% of the time: the model gets about 3 more results right per 100 matches.
The online record below grows with every match played. It needs many more results before it says much on its own.
Check the published online record ↓ · Historical test and baseline ↓
Separate recent-data sample
Backtest · 10 matches
Forecasts fitted without seeing these results. Separate from the published online record.
Log loss 1.103 · Brier 0.666.
Small sample; descriptive results only.
Premier League · published online
Live prediction record
- Published
- 50
- Settled
- 50
- Result hits
- 26 / 50
0 awaiting results · fixed pre-match snapshots
How this record is scored
Predictions are stored from published pre-24h snapshots and never rewritten. Result hits count the most likely 1X2 outcome. This online sample is separate from the historical test.
Prediction accuracy by league
Each league has its own model, tested on matches it had not seen. The hit rate is how often the most likely result happened; log loss scores the full probabilities, and lower is better.
| League | Held-out matches | Most likely result happened | Backing the higher club | Log loss |
|---|---|---|---|---|
| Premier League | 1,140 | 51.8% | 49.0% | 0.993 |
| Bundesliga | 1,140 | 51.7% | 48.8% | 0.992 |
| Championship | 1,140 | 44.3% | 43.5% | 1.056 |
| La Liga | 1,140 | 53.6% | 49.7% | 0.972 |
| Ligue 1 | 1,140 | 52.4% | 47.1% | 1.000 |
| Serie A | 1,140 | 52.7% | 52.0% | 0.981 |
European competitions · shared model
Walk-forward log loss 0.942 on 2,817 European matches (2016-17 to 2026-27); historical outcome baseline 1.051. This is one shared test, not three independent validations.
How to check a prediction record
Check timestamps before kickoff, an unedited record of hits and misses, and probability scores against a baseline on the same matches. Hit rate alone does not measure probability quality.
When the model says 60%, does it happen 60% of the time?
Every match has three possible results, each with a probability. Grouping all 3,420 of them from 1,140 held-out matches by what the model said shows how often each group actually happened. Well-calibrated probabilities sit close to the figure the model gave.
Bar length is how often results in each group actually happened. Groups with fewer than 30 results are left out.
Historical test · 1,140 matches
| Forecast | Most likely result happened |
|---|---|
| Our model | 51.8% |
| Back the club higher in the table | 49.0% |
On the same matches, simply backing the club higher in the table was right 49.0% of the time: the model gets about 3 more results right per 100 matches.
Test method
None of these matches were used to choose the model's settings. The table rule uses the league table as it stood before each kickoff; when two clubs are level, as on the opening weekend, it backs the home side. It never predicts a draw, which is part of why it is hard to beat by much.
The model also gives probabilities, not just a pick. Its log loss was 0.993 and Brier score 0.593 (lower is better); see the calibration above for whether those probabilities can be taken at face value.
Probabilities were produced by fits refreshed every 28 days using only results available before each match.
About the accuracy scores
Log loss and Brier score measure how much probability a forecast gave to what actually happened; lower is better for both. They reward honest probabilities rather than picking winners, so a model can pick the favourite every time and still score badly if it is overconfident. See how log loss and Brier score work.
- Results up to
- 2023-24, 2024-25, 2025-26, 2026-27
- Matches used
- 1,140
- Method
- Walk-forward Dixon–Coles, checked for calibration and compared with backing the club higher in the league table.
- Data source
- Loaded match database
- Limitations
- Score-based model only. Historical accuracy does not guarantee future accuracy.