Walk-Forward Cross-Validation Results
Model performance on historical data (2014-2023 train, 2024-2025 test)
Mean Absolute Error
3.69
wins per team
RMSE
5.09
wins per team
R² Score
0.880
variance explained
Test Years
4
walk-forward folds
Mean Absolute Error by Year
Year-by-Year Performance
| Year | MAE | RMSE | R² | Teams |
|---|---|---|---|---|
| 2021 | 4.55 | 6.45 | 0.819 | 30 |
| 2022 | 3.96 | 4.98 | 0.907 | 30 |
| 2023 | 2.97 | 4.15 | 0.915 | 30 |
| 2024 | 3.28 | 4.79 | 0.868 | 30 |
Model Calibration
The model uses walk-forward cross-validation to simulate real-world prediction scenarios. For each test year (2024-2025), the model is trained only on prior years' data (2014-2023).
Target Accuracy: MAE < 5.5 wins per team. Current: 3.69 wins.
Lower MAE = better predictions. An MAE of 5 wins means the model is, on average, within ±5 wins of the actual outcome.