Performance of Professional Soccer Players: Multilevel Classification Using Machine Learning and SHAP Explainability by Playing Position

Background . Recent advances in data systematization have enabled the development of machine learning models to evaluate performance in elite sports; however, studies are needed to analyze the performance of professional players in relation to their playing position. Objective . To analyze the performance of professional soccer players who competed between 2017 and 2024 by applying a multilevel classification approach that integrates different machine learning algorithms. Materials and Methods . We analyzed 9088 player-seasons from professional players during the 2017–2024 seasons. These data were extracted from standardized databases on sports performance analysis belonging to the following leagues: LaLiga (Spain), Premier League (England), Bundesliga (Germany), Serie A (Italy), and Ligue 1 (France). The final sample, distributed by performance level, was as follows: elite players ( n = 1818; 20%), mid-level players ( n = 2727; 30%), and low-level players ( n = 4543; 50%). The average age of the players analyzed was 26.1 ± 4.0 years, distributed across three outfield playing positions: center backs (CB), central midfielders (CM), and strikers (ST); goalkeepers were excluded because the dataset contains no goalkeeper-specific performance metrics. Results . Under a leakage-controlled protocol (the label-defining indicators were excluded from the predictors and the train/test split preceded all preprocessing), a linear model (logistic regression) achieved the best overall performance (mean macro-F1 = 0.729), ahead of ensemble and kernel-based methods; this indicates that, once target leakage is removed, the classification does not require non-linear models. Strikers (ST) were the most separable position (best macro-F1 = 0.818, area under the receiver operating characteristic curve [AUC-ROC] = 0.948) and center backs (CB) the most difficult (macro-F1 = 0.602, AUC-ROC = 0.791), with midfielders (CM) intermediate (macro-F1 = 0.770, AUC-ROC = 0.916). Conclusions . The results confirmed that performance structures differ substantially depending on the position in the field, supporting the use of position-specific analytical strategies. In this context, the combination of position-stratified dimensionality reduction, handling of imbalance, and explainable artificial intelligence allowed for the identification of interpretable performance patterns associated with the profiles of elite, mid-level, and low-level players.