Prediction Accuracy
Performance of the current model over the last 14 days of completed matches in our dataset. Every fixture in the window is included — nothing is hand-picked — but read the note below on what these numbers are and are not.
See the full dated results log →
What these numbers are: a rolling backtest
These figures come from re-running the current model over completed fixtures and comparing its output with the final score. They are a backtest — a measure of how the model behaves on matches it has data for — not an audited archive of predictions timestamped and published before kick-off. We do not currently store a pre-match snapshot of every prediction, so we do not present this as a live betting track record, and no profit or return figure is claimed anywhere on this site.
Because the underlying form data is refreshed continuously, a re-run can see information that was not available before kick-off. That tends to flatter a backtest. Treat the calibration chart below as the more useful signal: it shows whether the stated probabilities line up with real outcome frequencies.
1X2 hit rate
50.8%
BTTS hit rate
61.7%
Over/under 2.5
58.3%
Brier score
0.618
Sample size
120
Model calibration
For each confidence band: the average probability the model issued, next to how often those picks actually won. Close bars mean the percentages can be taken at face value.
- Under 40%19 picksPredicted37.9%Actual57.9%
- 40–50%55 picksPredicted43.6%Actual41.8%
- 50–60%31 picksPredicted55.6%Actual58.1%
- 60–70%11 picksPredicted63.9%Actual45.5%
- 70%+4 picksPredicted75.6%Actual100%
Recently settled picks
Every finished match in the review window, with the selection the model produces for that fixture.
| Date | Match | Score | Model pick | Prob. | Result |
|---|---|---|---|---|---|
| 2026-10-12 | Gimnasia M. vs Newells Old Boys | 1–0 | Gimnasia M. | 39% | ✓ Correct |
| 2026-10-12 | Lanus vs Velez Sarsfield | 2–0 | Lanus | 37% | ✓ Correct |
| 2026-10-12 | Talleres Cordoba vs Independiente | 0–0 | Talleres Cordoba | 36% | ✕ Missed |
| 2026-10-12 | Chapecoense-sc vs Atletico Paranaense | 1–4 | Atletico Paranaense | 47% | ✓ Correct |
| 2026-10-12 | Coritiba vs Botafogo | 1–1 | Botafogo | 43% | ✕ Missed |
| 2026-10-12 | Famalicao vs Alverca | 1–4 | Famalicao | 43% | ✕ Missed |
| 2026-10-12 | Torino vs Udinese | 2–1 | Udinese | 44% | ✕ Missed |
| 2026-10-12 | Atalanta vs Venezia | 3–1 | Atalanta | 56% | ✓ Correct |
| 2026-10-12 | Levante vs Sevilla | 3–3 | Sevilla | 41% | ✕ Missed |
| 2026-10-12 | Coventry vs Newcastle | 4–1 | Newcastle | 61% | ✕ Missed |
| 2026-10-11 | Nice vs Strasbourg | 2–1 | Nice | 39% | ✓ Correct |
| 2026-10-11 | SC Freiburg vs FC Schalke 04 | 2–1 | SC Freiburg | 74% | ✓ Correct |
| 2026-10-11 | 1. FC Köln vs Borussia Mönchengladbach | 2–1 | 1. FC Köln | 62% | ✓ Correct |
| 2026-10-11 | Cagliari vs Juventus | 1–1 | Juventus | 36% | ✕ Missed |
| 2026-10-11 | Sassuolo vs AC Milan | 1–1 | AC Milan | 42% | ✕ Missed |
| 2026-10-11 | Lecce vs Bologna | 1–2 | Bologna | 46% | ✓ Correct |
| 2026-10-11 | Lazio vs Monza | 3–1 | Lazio | 59% | ✓ Correct |
| 2026-10-11 | Como vs AS Roma | 0–0 | AS Roma | 55% | ✕ Missed |
| 2026-10-11 | Racing Santander vs Valencia | 2–2 | Racing Santander | 44% | ✕ Missed |
| 2026-10-11 | Real Betis vs Osasuna | 5–2 | Real Betis | 58% | ✓ Correct |
| 2026-10-11 | Real Sociedad vs Deportivo La Coruna | 3–2 | Deportivo La Coruna | 45% | ✕ Missed |
| 2026-10-11 | Elche vs Celta Vigo | 0–0 | Celta Vigo | 56% | ✕ Missed |
| 2026-10-11 | Liverpool vs Manchester City | 0–3 | Manchester City | 55% | ✓ Correct |
| 2026-10-11 | Hull City vs Everton | 2–2 | Draw | 35% | ✓ Correct |
| 2026-10-11 | Crystal Palace vs Nottingham Forest | 3–1 | Nottingham Forest | 56% | ✕ Missed |
How to read these numbers
The 1X2 hit rate counts how often the highest-probability outcome was the outcome that occurred. Realistic figures for football sit between 45% and 55%; anything advertised far above that range across a large sample should be treated with caution.
BTTS and over/under are two-way markets, so a coin flip scores 50%. A model needs to beat that meaningfully and consistently to be adding value.
The Brier score measures calibration rather than raw hits: it compares each probability issued against what actually happened, and lower is better. It is the fairest single number for judging a probabilistic model, because it penalises confident mistakes more than cautious ones.
Sample size is shown deliberately. A hit rate over a small number of matches carries large error bars, and short-term swings in either direction are expected even from a well-calibrated model.