M.S. Applied Data Science - Capstone Chronicles 2025

22

Figure 23 Top Predicted Outfield Performers (Next 4 Weeks)

Figure 24 Biggest Predicted Performers Gains (Outfield Players)

To highlight the model’s sensitivity to dynamic player trends, we visualized the most significant forecasted score changes. As illustrated in Figure 25, the model detects both upward and downward trends across midfielders, defenders, and forwards. This capability supports actionable insight for rotation strategies, scouting, and tactical preparation. Figure 25 Biggest Predicted Performance Declines (Outfield Players)

5.2 Tuning and Iterative Improvement The model was refined across four iterative forecast windows, retraining weekly using newly accumulated player match data. Hyperparameters such as learning rate, max depth, and tree count were manually adjusted using observed performance consistency and SHAP stability. This approach, although not fully automated via grid search, allowed for targeted optimization of prediction accuracy and robustness. 5.3 Holdback Validation and Predictive Insights We simulated holdback validation by excluding the most recent match week and using it for forward looking prediction. Forecasts were compared to qualitative domain expectations, confirming alignment in most cases. For example, Éder Militão’s predicted increase of over 2.1 points from his current performance reflects anticipated growth in his match influence. Conversely, Nicolás Paz was forecasted to decline by −1.25 points, potentially reflecting reduced involvement or form (see Figure 24).

69