Sim2Win: A Team-Agnostic, Event-Based Pre-Match Outcome Prediction and Tactical Profiling System for Football

Overview

In professional football, pre-match tactical decision-making has traditionally relied on subjective expert analysis and identity-based scouting systems that fail to generalize to unseen teams. In response, this paper introduces Sim2Win, a team-agnostic, event-based pre-match tactical recommendation framework. Unlike conventional approaches that treat match outcome prediction as a pure forecasting task, Sim2Win reframes it as a tactical decision-support problem, offering actionable insights without requiring team identifiers.

By 2026, the growing availability of fine-grained event data has opened new avenues for data-driven sports analytics. Sim2Win leverages this trend to build robust machine learning models that prioritize behavioral patterns over team identity.

Methodology

Sim2Win utilizes StatsBomb open event data from eleven competitions, covering 178 teams and 1,411 team-match records. The framework proceeds through several stages:

1. Tactical Profile Construction: Five-match rolling tactical profiles are created for each team, capturing recent performance and playing style.
2. Feature Engineering: Four interpretable tactical feature ratios are derived, representing key aspects of team behavior (e.g., attacking intensity, defensive compactness, possession style).
3. Behavioral Clustering: K-Means clustering groups team behaviors into eight distinct playstyles, enabling direct tactical matchup comparisons.
4. Outcome Prediction: Thirteen classifiers are trained to estimate win, draw, and loss probabilities from tactical matchup representations. Models include logistic regression, random forests, gradient boosting machines (including XGBoost and CatBoost), and neural networks.



A critical design choice is the omission of team names or identity features, forcing the model to rely solely on tactical signatures. This ensures generalization to teams never encountered during training.




Evaluation and Results




The system is rigorously evaluated using a Leave-One-Competition-Out (LOCO) protocol, simulating the most challenging scenario: predicting outcomes for completely unseen teams across different competitions. Sim2Win achieves:



• Mean ROC-AUC: 0.704
• Mean accuracy: 55.4% on unseen teams



These results outperform traditional football rating systems—ELO, Pi-Rating, and GAP—on all 21 ROC-AUC comparisons and 19 out of 21 accuracy comparisons. Among the 13 models tested, CatBoost achieved the strongest in-distribution performance with 60.90% accuracy.




Contributions and Implications




The findings demonstrate that behavioral tactical representations contain transferable predictive signal under distribution shift. This is a significant departure from identity-dependent models (e.g., team rankings or historical head-to-head records), which fail when encountering new opponents or competitions. Sim2Win offers a viable, scalable alternative for:



• Pre-match analytics across leagues with limited historical data
• Talent scouting and opponent profiling for less-resourced teams
• Tactically informed decision-making for coaches and analysts



Conclusion




Sim2Win establishes that event-based, team-agnostic modeling can produce reliable outcome predictions and tactical insights in football. As sports analytics continues to evolve into 2026, approaches like Sim2Win—rooted in behavioral patterns rather than team identity—are likely to become standard in both research and practice.




---

Code Availability: The source code is publicly available at GitHub.




Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)




ACM Classes: I.2.6; I.5.3; H.2.8

via ArXiv LG