Position Prediction and Talent Discovery in Football Leagues Using Performance Data

Abstract

Football has always been dependent on the subjective evaluation of scouts and coaches to find and hire players. Although these methods work to some extent, they usually have restrictions due to human biases, irregularity, and the huge volume of football data. As more data on player performance is made available, data analytics and machine learning represent a chance to introduce objectivity, consistency, and scalability in the recruitment process. This research study suggests a machine learning-based classification model along with a clustering model to classify football players in their main positional roles using statistical performance features. The research is based on the development of models that would help to differentiate among defenders, midfielders, and attackers based on their passing efficiency, contributions to defense, won duels, and attacking indicators. For data extraction, Fbref has been used as the source of data. The player-level data of the 2023-24 season of the Top 5 European Leagues has been extracted using the Python programming language. The data involved various statistical categories addressing all the areas of performance. Position labels were merged with the scraped tables to ensure accurate role mapping. This combination resulted in the creation of an entire dataset with both performance and position features. The dataset was cleaned and prepared using data preprocessing techniques, and selected features were then used in the training process. K-Means Clustering was applied to the PCA-transformed data to cluster similar players based on their playing profile.  Different supervised learning algorithms have also been applied, such as Logistic Regression, Decision Tree, K-Nearest Neighbors (KNN), Random Forest, and Voting Classifier. The standard evaluation parameters are used to provide a detailed evaluation of the predictive performance. It was found that ensemble algorithms, in particular Random Forest and the Voting Classifier, performed better than single baseline models and were stronger and more reliable in positional classification. The results suggest the potential of machine learning models when recruiting players in football teams and to facilitate and aid expert judgment. This research sets up a systematic, data-driven framework that helps clubs to screen the enormous number of players effectively in a non-subjective manner.