Objective: Developed a machine learning algorithm to efficiently estimate player-team compatibility scores, utilizing advanced analytical techniques for accurate assessments.

Conducted comprehensive data reshaping, employing SciPy for anomaly detection; developed targets based on playing time rates, using binary classification for goalkeepers and numerical values for the other 12 positions.

Implemented LightGBM models for each position, utilizing the SHAP library for iterative (RFECV-like) feature selection, which resulted in test error levels ranging from 5-15%.

Refined the model outputs during the prediction pipeline phase by converting them to ranks. This approach helped reduce potential noise and enhanced the intuitiveness of data presentation.

Optimization & Outcome: Initially, the prediction API included over 500 million team-player combinations. However, enriching the SQL pipeline with an affordability calculation reduced the number of combinations to 10x. Additionally, prioritizing certain predictions ensured the delivery of the most critical insights.