Football Player Potential Prediction Model

At Scoutium every scout rates each player they watch across dozens of attributes, then makes one final call: is this player average, or highlighted? The problem is that this final call can differ between two scouts looking at the same score sheet.

The attribute scores are data; the final verdict is judgement. What connects them - which attributes actually drive the "highlighted" decision - had never been measured.

Scoutium didn't want a system to replace the scout. They wanted a model that could read the same score sheet and give a consistent, repeatable second opinion.

I built a classification model that predicts the final potential label from scout attribute scores.

1. I reduced the data to player level. Attribute scores and final labels lived in separate tables; I merged them into a pivot where each row is a player-position and each column an attribute. In total 10,730 attribute scores, 39 attributes, 10 positions, 277 players.
2. I handled the imbalance honestly. The label distribution was skewed: 255 average, 63 highlighted, 4 below average. The below-average class - about 1% of the data - was removed, because there is nothing to learn from four examples and keeping it would distort the results.
3. I didn't trust a single model. Logistic regression, KNN, random forest, gradient boosting, XGBoost and LightGBM were compared on the same data. Measurement rested on 10-fold cross-validation rather than one split, reporting accuracy, precision, recall, F1 and ROC-AUC together.
4. I chose the right metric for imbalanced data. With 80% of the data labelled "average", accuracy alone is misleading - a model that blindly says "average" scores well. So ROC-AUC became the real criterion: the model's ability to separate a highlighted player from an average one.

Result: ROC-AUC 0.86 under 10-fold cross-validation. The model doesn't replace the scout; it reads the same sheet and offers a consistent second opinion. When a scout's verdict diverges sharply from the model's, that divergence is itself a signal worth a conversation.

Stack: Python, pandas, scikit-learn, XGBoost, LightGBM; pivot-table feature setup, LabelEncoder, StandardScaler, 10-fold cross-validation.