Author(s): ABDELMALAK, M., BASHA, T., Institution: CAIRO UNIVERSITY , Country: EGYPT, Abstract-ID: 1679

INTRODUCTION:
Forecasting how players reposition right after a pass is central to understanding team tactics, judging decisions, and shaping realistic training in soccer[1,2]. Existing approaches often focus on pass difficulty or value (e.g., expected pass) without providing coordinated forecasts of full team formations or individual movements[3,4]. In this work, we propose a transformer model that predicts post-pass locations at three granularities -single player, possession team (11), and both teams (22)- using FIFA World Cup data. The aim is to provide a scalable tool for reconstructing immediate post-pass formations to support performance analysis, scouting, and training design.
METHODS:
Event and tracking data from matches of the 2022 FIFA World Cup were processed into three-event pass sequences. For single-player models, high-participation French players (>50%) were selected. Sequences encoded 66 features, including player positions (ordered arbitrarily or by formation role), event descriptors, ball coordinates, and pitch-aware spatial indicators. The proposed Pitch-Aware Temporal Transformer model integrates short- and medium-term branches with multi-head attention and LSTMs to predict single-player or team-level coordinates. Baselines trained on 37 non-French matches (80/10/10 split) were compared to models fine-tuned on France's 7 matches and France-only baselines. Performance was quantified using MAE, RMSE, and R2.
RESULTS:
The transformer consistently improved forecasts, with benefits from role-ordering and fine-tuning. For unordered inputs, fine-tuning reduced both-teams MAE from 8.1 m to 5.8 m (28% reduction) and possession-team MAE from 6.3 m to 4.7 m (25% reduction). Fine-tuning outperformed direct France-only training by 14-15%, indicating that adapting a multi-team model yields better tactical forecasts than training from scratch. With ordered inputs, fine-tuning improved both-teams MAE to 5.1 m (R2 0.85) and possession-team MAE to 4.8 m (R2 0.87). Ordered variants outperformed unordered by approximately 10-15%. The single-player model achieved the highest precision (MAE 3.4 m, R2 0.9, RMSE 4.8 m), improving 29-37% over team-level models.
CONCLUSION:
The proposed transformer model accurately reconstructs post-pass formations, particularly when player roles are ordered and models are fine-tuned. Notably, our single-player model (MAE 3.4 m, RMSE 4.8 m) outperforms previously reported related tasks such as Agent Imputer (MAE 4.5 m and RMSE 6.9 m) [5]. Robust gains at the team level highlight the framework's ability to capture coordinated tactical reorganization. Unlike approaches that only estimate pass value, this work delivers a variable-granularity tool to simulate likely post-pass configurations. For sports scientists and coaches, this offers a scalable way to link event data to realistic on-pitch reorganization, supporting informed decision-making in scouting and session planning.References: [1]Teixeira2025 /[2]Ashford2021 /[3]Fernandez2021 /[4]StatsBomb2023 /[5]Everett2023