League Winner Predictor

Each season, only one team wins each league — and their forwards tend to have significantly higher goal and shot counts than everyone else. This project trains a K-Nearest Neighbors (KNN) classifier to predict whether a given forward's team won the league that season, using only two features: goals scored and shots on target. Trained on 88,310 player-season records across seven European leagues and five seasons, the model achieves 95.45% accuracy.

Dataset

Player-season records from the Premier League, La Liga, Serie A, Bundesliga, Ligue 1, Eredivisie, and Primeira Liga. The model is restricted to forwards with at least five 90-minute appearances, where the goal-scoring signal is strongest. The 2019–20 season is excluded due to the Eredivisie being suspended mid-season.

Why this works

League-winning teams dominate possession and create more chances — so their forwards rack up noticeably more goals and shots than forwards on lower-finishing sides. This gap is large enough that a simple classifier can pick up on it reliably.

Group	Avg Goals	Avg Shots on Target
Non-winners	6.60	16.96
Winners	13.59	29.37

Data cleaning

Missing goals filled with 0; mislabelled columns corrected; squad-total rows removed
Records with goals but zero shots on target — indicating missing data rather than true zeros — imputed using a linear regression model fitted per league
Primeira Liga had the most affected rows (452), followed by Ligue 1 (61)

Model

KNN classifies each player by looking at the 11 most similar player-seasons in the training data (by goals and shots) and taking a majority vote on whether those neighbors' teams won the league. Hyperparameters were tuned via GridSearchCV across 150 combinations and 5-fold cross-validation.

Features:     Shots On Target, Goals, Season (one-hot encoded)
Scaler:      StandardScaler on numerical features
Search:      GridSearchCV — 5-fold CV, accuracy scoring

Best params:
  n_neighbors  11
  metric       euclidean
  weights      uniform
  CV score     0.9545
    

Key findings

Baseline (goals + shots on target only) already achieved 94.35% — reflecting how cleanly elite forwards separate from the rest of the distribution
Adding one-hot encoding on season pushed accuracy to 95.45% — year-to-year variation in league competitiveness is important
Optimal k=11 suggests the decision boundary benefits from moderate smoothing
Euclidean distance with uniform weights outperformed distance-weighted variants