Does the model actually work?
Before showing you a single probability, we tested the model on real, already-played matches it had never seen. This page is the honest answer to one question: how often is it right? The numbers come straight from the test files — nobody edits them.
Tested on 30 real matches from the 2025 season. None were left out. · Model version dc-2026.07.29-2336
Before kick-off
The model makes its forecast before the match starts. How often was its most likely outcome the one that actually happened?
Match result (win / draw / win)
60%Right in 18 of 30 matches
Over or under 2.5 goals
73%Right in 22 of 30 matches
Both teams score — yes or no
57%Right in 17 of 30 matches
Worth saying out loud: on "both teams score" the model is currently only a little better than a coin flip. We would rather tell you that than hide it.
During the match
The model keeps re-computing as the match unfolds — goals, red cards, the clock. Its call of the final result gets sharper with every minute:
By minute 85 the live model was calling the final result correctly in 25 of 30 matches (83%). Before kick-off it managed 18 of 30.
It also predicts which team scores next: right 53% of the time across 300 situations. On rare near-certain calls it runs slightly overconfident — a known limit we keep watching.
Where it was wrong
No model sees everything, so we publish the misses too. A few matches it badly misjudged:
Liverpool vs Bournemouth (E0 2025-08-15) finished 4-2. At minute 85 the score was 2–2 and the model gave the actual final result only a 11.5% chance — the late goals proved it wrong (88' home, 90' home).
Rennes vs Marseille (F1 2025-08-15) finished 1-0. At minute 85 the score was 0–0 and the model gave the actual final result only a 8.3% chance — the late goals proved it wrong (90' home).
Fiorentina vs Milan (I1 2026-01-11) finished 1-1. At minute 85 the score was 1–0 and the model gave the actual final result only a 8.9% chance — the late goals proved it wrong (90' away).
Technical details for the statistically curious
Log loss scores the full probability, not just the pick (lower is better), compared against an uninformed guesser.
| Market | Accuracy | 95% CI | Log loss | Uninformed baseline |
|---|---|---|---|---|
| 1X2 | 60.0% | 42%–75% | 0.966 | 1.099 |
| O/U 2.5 | 73.3% | 56%–86% | 0.691 | 0.693 |
| BTTS | 56.7% | 39%–73% | 0.683 | 0.693 |
| Minute | Live log loss | Pre-match log loss |
|---|---|---|
| 0' | 0.963 | 0.966 |
| 10' | 1.015 | 0.966 |
| 20' | 0.919 | 0.966 |
| 30' | 0.939 | 0.966 |
| 40' | 0.900 | 0.966 |
| 50' | 0.925 | 0.966 |
| 60' | 0.804 | 0.966 |
| 70' | 0.788 | 0.966 |
| 80' | 0.667 | 0.966 |
| 85' | 0.591 | 0.966 |
How we test: the model is trained only on matches BEFORE each test match (strict date cut-offs — it can never peek at the future), then scored against what really happened. The in-play replay uses recorded match events. How the model works
Statistical model output for informational purposes only. Not betting, financial, or investment advice. Probabilities are model estimates and carry uncertainty. Past data does not predict outcomes. 18+