Sim2Win predicts football match outcomes without knowing team names
This AI system beats ELO and Pi-Rating using only tactical playstyles...
Sim2Win reframes football match outcome prediction as a tactical decision-support problem, eliminating reliance on team names or scouting. The system processes StatsBomb open event data from eleven competitions (178 teams, 1,411 matches), building rolling five-match tactical profiles. It engineers four interpretable tactical feature ratios, clusters team behaviors into eight distinct playstyles via K-Means, and trains 13 classifiers to estimate win, draw, and loss probabilities purely from tactical matchup representations.
In a rigorous Leave-One-Competition-Out (LOCO) evaluation, Sim2Win achieves a mean ROC-AUC of 0.704 and mean accuracy of 55.4% on completely unseen teams—outperforming ELO, Pi-Rating, and GAP baselines on all 21 ROC-AUC comparisons and 19 of 21 accuracy comparisons. CatBoost produced the best in-distribution accuracy at 60.90%. The authors argue that behavioral tactical representations provide transferable predictive signal, offering a viable alternative to identity-dependent football prediction systems. Code is available on GitHub.
- Trained on StatsBomb data from 1,411 matches across 178 teams and 11 competitions
- Achieves 55.4% accuracy and 0.704 ROC-AUC on unseen teams, beating ELO and Pi-Rating
- Uses 8 playstyle clusters and 13 classifiers without any team identity features
Why It Matters
Proves tactical behavior alone can outperform traditional team-based prediction, enabling scouts to prepare for unfamiliar opponents.