Predicting Defensive Formation Effectiveness in Football Using Random Forest and LSTM Models

Abstract

Defensive strategies are fundamental to football success, yet the evaluation of formation effectiveness often remains subjective. This study proposes a data-driven approach to predict the most effective defensive formations by integrating machine learning models. Using tracking-derived features from 150 professional European matches (2018–2023), Random Forest (RF) and Long Short-Term Memory (LSTM) models were applied to assess defensive outcomes. The results indicate that the 5-3-2 formation consistently achieved the highest predicted defensive success across direct, wing, and central attacks, followed by 4-4-2, while the 4-3-3 formation exhibited the weakest defensive stability. RF identified key static features such as line height, block width, and compactness, while LSTM captured temporal dynamics of coordinated player movements, yielding superior predictive performance. This study concludes that combining interpretable ensemble models with sequence-based neural networks offers a robust framework for tactical analysis. The findings provide actionable insights for coaches and analysts, supporting evidence-based decision-making in optimizing defensive strategies in modern football.