Comparação de modelos preditivos para campeonatos de futebol: uma análise de seis ligas mundiais (2003-2023)

Abstract

This study systematically compares three predictive modeling approaches for football championships: (i) a purely quantitative model based on Poisson distribution, (ii) a purely categorical model based on performance profiles, and (iii) a hybrid model combining both approaches. The analysis was applied to six major world leagues (Brazil, England, Spain, Italy, Germany, and France) over 21 seasons (2003 to 2023), totaling 126 season-championships and 46,503 matches. Using temporal cross-validation with 5 folds, the memory parameter $h$ was optimized independently for each model, league, and season. The results revealed better performance of the Pure Poisson model for the analyzed dataset, which presented the best metrics across all six leagues (global MAE: 3.37 versus 3.43 for categorical models) and higher categorical accuracy (82.5% champion prediction versus 79.4%), empirically validating the adequacy of the Poisson distribution for modeling goals in football within the scope of this research. All combinations showed high performance inertia (h > 0.75), indicating that accumulated history is more informative than recent form. Systematic differences between leagues were identified: Germany more predictable (MAE = 2.96), England less predictable (MAE = 3.71). Hierarchical clustering analysis identified three distinct groups of leagues with similar dynamics, demonstrating that inertia and predictability are partially independent dimensions. As a practical application, although the Pure Poisson model presented superior metrics, the Hybrid model was used to generate predictions for the 2025 Brazilian Championships Series A and B due to its methodological completeness in generating realistic scores while capturing match outcome dynamics (W/D/L). The application indicated a balanced title dispute in Series A between Flamengo (46%) and Palmeiras (43%) and Coritiba's favoritism in Series B (72% title probability), in addition to detailing the 35 possible promotion combinations in the latter.