Cross-League Football Data Conversion: How Machine Learning Solves the Comparability Challenge of Player Statistics in the Big Five Leagues

This project, developed by Mohammad Arshan Shaikh and open-sourced on GitHub, aims to achieve cross-league conversion of per-90-minute player statistics in Europe's Big Five Leagues (Premier League, La Liga, Bundesliga, Serie A, Ligue 1) through a machine learning framework. It also incorporates conformal prediction methods to quantify prediction uncertainty, addressing the long-standing league barrier issue that plagues analysts.

League Barrier Issue

Differences in competitive level, playing style, and referee standards across leagues lead to misleading direct data comparisons. Traditional methods rely on subjective judgment or simple coefficient adjustments, lacking systematicity and uncertainty assessment.

Core Problem

Focuses on predicting changes in player statistics after transfers between the Big Five Leagues, collecting data from the 2017/18 to 2023/24 seasons, covering 25 league conversion directions.

1. Transfer Bridge Strategy

Track cross-league transferred players and pair their pre- and post-transfer data (requiring ≥5 full 90-minute appearances). The final dataset includes samples for forwards (202 training / 82 test), midfielders (255 / 109), and defenders (257 / 135).

2. Model Zoo

Compare 12 algorithms: linear models (Ridge/Lasso, etc.), tree ensembles (RandomForest/CatBoost, etc.), robust regression (GBM-Huber, etc.), and neural networks (MLP).

3. Conformal Prediction Framework

• V1: 20% validation, coverage rate of 52.5%
• V2: 30% validation + interaction features, coverage rate of 83.3%
• Mondrian: League-stratified calibration, average coverage rate of 93.2% (meeting the 90% nominal requirement) Conformal prediction is distribution-free and provides finite sample guarantees.

Overall Performance

Average MAE of 0.3325 (log ratio scale), 6.1% improvement over the mean baseline and 7.6% improvement over the paired mean baseline.

Model Ranking

• CatBoost dominates 5 metrics with balanced performance
• Progressive passes for midfielders: GBM-Huber is optimal
• Interceptions for defenders: LightGBM with interaction features is optimal

Feature Importance

Original statistics from the source league and differences in UEFA coefficients are key predictors.

• Transfer Decisions: Scientifically evaluate the expected performance and confidence level of overseas players
• Gambling/Fantasy Football: Build fair valuation models
• Academic Research: Submitted to the Journal of Quantitative Analysis in Sports and under review

• Code structure: MLCLSTData.ipynb (data pipeline), MLCLST.ipynb (modeling process)
• Reproducibility: Fixed random seed (42), dependencies managed via requirements.txt
• Data: FBref raw data not included, but bridge CSV can be used directly

Limitations

• Small sample size for some league combinations
• Simplified position classification (forward/midfielder/defender)
• No consideration of temporal dynamics in league level

Future Directions

• Incorporate physical fitness data
• Refine position classification
• Explore deep learning architectures
• Regularly update the model

This study not only solves the practical problem of cross-league player comparison but also introduces conformal prediction to quantify uncertainty (rare in sports analysis). The open-source code provides a scientific tool for industry practitioners and demonstrates the application value of conformal prediction to machine learning researchers.