Predicting Football Shot Outcomes in the Turkish Super League: Support Vector Machine and Naive Bayes Models with Feature

Creative Commons License

Genc M. N., Can M., Fidan Kececi N.

RESEARCH IN SPORT EDUCATION AND SCIENCES, cilt.28, sa.3, ss.176-193, 2026 (ESCI)

Özet

This study presents a modeling framework that formulates goal prediction as a binary classification problem (1 for a goal, 0 for no goal), based on 46,238 shot records from the Turkish Super League between the 2020-2025 seasons. In this study, Naive Bayes (NB) and Support Vector Machine (SVM), which are supervised machine learning algorithms, were used for prediction models. Considering the substantial class imbalance between goal and non-goal outcomes in the original dataset, a balanced sample was preferred to ensure a more balanced representation of the classes in the modeling process. The results of machine learning models created for a balanced dataset of 10,081 shots were compared, and the findings showed that SVM models (AUC 0.830 for RBF SVM and 0.824 for Linear SVM) discriminated between goals and non-goals more consistently than the NB models (Gaussian NB 0.801 and Complement NB 0.757). Feature importance analyses based on Permutation Importance and SHAP identified distance to goal as the dominant predictor, followed by the length of the preceding pass, while categorical shot descriptors carried substantially less weight. By combining a comparison of two algorithm families with model-agnostic explainability methods, this study provides league-specific evidence on which shot characteristics distinguish goals from non-goals and offers a transparent framework that analysts working in this league can apply directly.