Table 2.
Accuracy of the machine learning models for all offensive and defensive performance scores. Offensive performance is most accurately predicted by the random forest model, and XGBoost model is the best model for defensive actions. Models that only use action type as predictors are only slightly less accurate than our best machine learning models, suggesting that adding predictors related to external and internal training load does not result in large improvements in the prediction of offensive and defensive performance.
| Action Type | Model | MAE (95% CI) |
Difference in MAE | p-Value | Cohen’s d
(95% CI) |
Effect Size |
|---|---|---|---|---|---|---|
| Offense | Random Forest | 0.91 (0.62–1.19) |
−46.8% | p < 0.001 | 0.79 (0.47–1.18) |
Medium |
| XGBoost | 1.09 (0.78–1.41) |
−36.3% | p < 0.01 | 0.58 (0.23–0.99) |
Medium | |
| Action Model | 1.04 (0.75–1.32) |
−39.2% | p < 0.001 | 0.66 (0.35–1.04) |
Medium | |
| Baseline | 1.71 (1.38–2.05) |
|||||
| Defense | Random Forest | 1.15 (0.79–1.51) |
−59.4% | p < 0.001 | 1.47 (1.10–1.98) |
Large |
| XGBoost | 0.75 (0.50–1.00) |
−73.5% | p < 0.001 | 2.09 (1.63–2.78) |
Large | |
| Action Model | 0.79 (0.58–1.00) |
−72.1% | p < 0.001 | 2.14 (1.63–2.88) |
Large | |
| Baseline | 2.83 (2.47–3.20) |