Ciência de Dados aplicada na sistematização da avaliação do potencial esportivo

Abstract

Sports talent is a complex and multifactorial construct, and its identification is fundamental for athlete development. In this context, the Projeto Atletas de Ouro (Gold Athletes Project), developed by the Laboratory for Exercise and Sport Studies and Research (LABESPEE), proposed the Gold Score, a predictive index that adopts a multidimensional approach to estimate the sports potential of young individuals. However, the current model may introduce biases and limit its applicability in certain contexts due to its reliance on subjective evaluations performed by coaches. Therefore, the aim of this dissertation was to develop and evaluate machine learning models capable of estimating the Gold Score, reducing or eliminating the reliance on subjective assessments. To achieve this, four progressive experiments were conducted wherein 13 machine learning algorithms were tested, including boosting techniques (CatBoost, XGBoost, Gradient Boosting, LightGBM), linear regression, and tree- and distance-based methods. The models were optimized using the Particle Swarm Optimization (PSO) algorithm and evaluated via 10-fold cross-validation on a dataset of over 3,000 student assessments collected between 2015 and 2023.

The results demonstrated that gradient boosting algorithms (especially CatBoost and XGBoost) exhibited the best predictive performance. In the complete scenario (including subjective variables), the models achieved an R² > 0.99 and successfully identified all medalists among the 20 students with the highest predicted potential. Even after removing the coaches subjective evaluations, the models maintained satisfactory explanatory power (R² ≈ 0.87) and preserved the ability to differentiate selected from non-selected individuals (large effect size, d > 0.88). In addition to optimizing resource allocation for in-depth individual assessments, the findings broaden the discussion on the use of computational methods as decision-support tools in sports talent identification processes. The proposed approach can be replicated in different school contexts, contributing to assessment equity by reducing reliance on subjective factors. These results point to promising directions for future research, including the prospective validation of the models, the expansion of the dataset to include other institutions, and the development of accessible computational tools (e.g., spreadsheets and web applications) for teachers and coaches.