Loading...
Game Outcome Model
Transparent training, accuracy, market, and methodology reporting
Data source
Live API
Model accuracy
63.5%(181/285)
Report date
Mar 23, 2026
Model evidence
Accuracy & Calibration
Review out-of-sample winner accuracy, probability alignment, confidence stratification, and individual results.
Model diagnostic result
01 · Classification result
What the holdout test showed
The model selected 181 of 285 winners correctly in the 2025 full season holdout sample.
Winner accuracy
63.5%
181/285 correct · 104 incorrect
Chance baseline+13.5 pts
Majority-class baseline 53.3%+10.2 pts
Classification
63.5%
Overall winner accuracy
Ranking quality · AUC
Not reported
Unavailable from source
Probability alignment
ECE 0.090
285 games in reported bins
Test design
2025 full season
Trained on 2019–2020–2021–2022–2023–2024 · 1675 games
High confidence: 32/33 correct (97.0%)Calibration and confidence are evaluated separately
Confidence tiers are not strictly ordered: medium-confidence accuracy is below low-confidence accuracy. Treat confidence stratification as diagnostic evidence, not proof of calibration.
02 · Performance context
Predictive performance
Classification results and neutral references for the selected test sample.
Accuracy
63.5%
181 of 285 correct
AUC-ROC Score
Not reported
Not reported by source
Error Rate
36.5%
104 of 285 incorrect
Training Data: 2019, 2020, 2021, 2022, 2023, 2024
Test Data: Full season (2025)
Chance baseline
50.0%
Two-outcome reference, not a betting break-even rate.
Majority-class baseline
53.3%
Accuracy from always choosing the more common test outcome.
Accuracy alone does not establish profitability. Prices, selection rules, and staking assumptions are evaluated in Market Performance.
03 · Calibration instrument
Probability calibration
Whether predicted home-win probabilities match observed outcomes across reported bins.
Expected calibration error
0.090
Lower means predicted and observed rates are closer.
Brier score
Not reported
Lower indicates smaller probability error.
Calibration sample
285
Games represented across reported bins.
Above the diagonal: outcomes occurred more often than predictedBelow: probabilities were higher than observed outcomes
Absolute calibration error by binLonger bars indicate larger gaps
- 10–20%7.0 pts · 10 games
- 20–30%1.0 pts · 30 games
- 30–40%23.5 pts · 43 games
- 40–50%8.4 pts · 54 games
- 50–60%2.9 pts · 57 games
- 60–70%3.9 pts · 38 games
- 70–80%11.4 pts · 30 games
- 80–90%15.4 pts · 23 games
| Probability bin | Games | Predicted | Observed | Absolute error |
|---|---|---|---|---|
| 10–20% | 10 | 17.0% | 10.0% | 7.0 pts |
| 20–30% | 30 | 25.7% | 26.7% | 1.0 pts |
| 30–40% | 43 | 34.6% | 58.1% | 23.5 pts |
| 40–50% | 54 | 45.4% | 37.0% | 8.4 pts |
| 50–60% | 57 | 54.9% | 57.9% | 2.9 pts |
| 60–70% | 38 | 64.4% | 60.5% | 3.9 pts |
| 70–80% | 30 | 74.7% | 63.3% | 11.4 pts |
| 80–90% | 23 | 84.6% | 100.0% | 15.4 pts |
Swipe the table horizontally to inspect every calibration column.
04 · Decision diagnostics
Error pattern and confidence
Where home/away classifications failed and how confidence tiers separated the sample.
Home and away decision board
Read across each actual outcome to see where the model’s home or away choice was correct.
Model picked awayModel picked home
Actual
away win
away win
Correct away pick
True negative
83
29.1% of all games
Incorrect home pick
False positive
50
17.5% of all games
Actual
home win
home win
Incorrect away pick
False negative
54
18.9% of all games
Correct home pick
True positive
98
34.4% of all games
Home-pick precision
66.2%
Correct among predicted home wins
Home-win recall
64.5%
Found among actual home wins
Away-win specificity
62.4%
Found among actual away wins
Interpretation: Incorrect home picks overstate the home team; incorrect away picks understate it. The board describes classification errors, not probability calibration.
Confidence accuracy ladder
Whether higher predicted-winner confidence separated easier and harder picks in this sample.
High confidence
≥80% predicted-winner probability
97.0%
32/33 correct
11.6% of the test sample
Medium confidence
60–80% predicted-winner probability
58.2%
82/141 correct
49.5% of the test sample
Low confidence
<60% predicted-winner probability
60.4%
67/111 correct
38.9% of the test sample
Higher confidence did not consistently correspond to higher accuracy in this sample. This is a stratification warning, not a calibration score.
| Tier | Correct | Accuracy | Range |
|---|---|---|---|
| High confidence | 32/33 | 97.0% | ≥80% predicted-winner probability |
| Medium confidence | 82/141 | 58.2% | 60–80% predicted-winner probability |
| Low confidence | 67/111 | 60.4% | <60% predicted-winner probability |
Confidence measures how strongly the model favored its selected winner. Calibration separately tests whether predicted probabilities match observed frequencies.
05 · Game audit
Test-game results
285 test games available
Open the explorer to filter and review 20 results at a time.