Comparison of the Temporal Fusion Transformer and XGBoost Models for Multi-Horizon Time-Series Forecasting of Competitive Rankings of Football Clubs

Sammanfattning

The use of data and machine learning (ML) has revolutionized numerous industries, andprofessional football is no exception. While research has traditionally focused on short-termmatch predictions or historical evaluations, predicting the long-term strength and relative ranking of clubs across multiple time horizons remains underexplored. This study utilizes Time-Series Forecasting (TSF) to evaluate the competitiveness of football clubs, measured by Elo rankings, over 1, 3, 6, and 12-month intervals. A central focus of this research is the comparisonbetween advanced Transformer-based architectures and more established machine learningmodels, specifically XGBoost, alongside basic linear baselines. Furthermore, the study identifiescritical underlying indicators for long-term success, analyzing the interplay between advancedon-field performance data and off-field strategic and financial factors. The empirical results demonstrate that the linear-calibrated XGBoost models consistentlyoutperform the complex Temporal Fusion Transformer architecture across all forecastinghorizons, delivering superior point-prediction accuracy. Furthermore, game-theoretic interpretability reveals that while immediate standings rely heavily on historical ratings, long-term trajectories are predominantly driven by underlying on-field performance metrics, with structural financial factors scaling in importance.