Previsão de rebaixamento no campeonato brasileiro Série A: uma comparação de modelos de machine learning com validação walk-forward

Abstract

Relegation to Série B is one of the greatest sporting and financial losses a Brazilian football club can suffer, directly affecting broadcasting rights revenues, sponsorships, and squad market value. The objective of this study is to develop and compare Machine Learning models capable of predicting, before the start of the season, which clubs face the highest relegation risk in the Brazilian Série A championship. Four binary classification algorithms were evaluated — Logistic Regression, Random Forest, XGBoost, and LightGBM — over a set of 15 features combining squad financial data, collected from the Transfermarkt portal, with sliding windows of performance from the previous three and five seasons. Validation used the walk-forward scheme, which respects the chronological order of the data, and hyperparameter optimization was performed with RandomizedSearchCV combined with TimeSeriesSplit. On the independent test set (2023–2024), LightGBM achieved the best AUC-ROC (0.877), while Logistic Regression showed the highest temporal stability in walk-forward validation (mean 0.794 ± 0.058). The forecast for the 2025 season pointed to Juventude, Sport Recife, Vitória, and Mirassol as the clubs at highest risk; retrospective validation against the official league results confirmed two of the four predicted relegations and an AUC-ROC of 0.922 for the season, with all four actually relegated clubs ranked among the seven highest predicted risks. The study contributes a reproducible pipeline for sports prediction with temporal data, empirically demonstrates the inadequacy of accuracy as a sole metric in imbalanced classification problems, and delivers the final model as an openly accessible web application.