This dissertation aims to establish an objective analytical framework that supports on-field decision-making beyond conventional performance evaluation in soccer. It develops structural performance indices that quantify interactions among players usin...
This dissertation aims to establish an objective analytical framework that supports on-field decision-making beyond conventional performance evaluation in soccer. It develops structural performance indices that quantify interactions among players using large-scale sports data and proposes time-series forecasting models to predict the temporal evolution of these indices. The dissertation consists of two complementary studies, and the main findings are summarized as follows.
Study 1 proposes a Set Transformer–based set learning model to explain team performance across attacking, defensive, and tactical domains, and compares its predictive performance with conventional team-aggregated machine learning models, including Random Forest, XGBoost, and Support Vector Regression. The proposed model consistently outperformed all comparison models across domains. In the attacking domain, it achieved the highest predictive accuracy (MAE = 0.3351, RMSE = 0.4426, R² = 0.7244). In the defensive domain, it recorded an RMSE of 0.6468 and an R² of 0.4116, while in the tactical domain it demonstrated very high accuracy (RMSE = 0.0375, R² = 0.8913). These results indicate that team performance emerges from structural interaction patterns among players participating in the same match rather than from simple aggregations of individual statistics.
Variable importance analysis revealed clear domain-specific characteristics. Attacking performance indices were primarily driven by goal-related events such as shots, goals, and shots on target. Defensive performance indices were most strongly influenced by critical events that sharply increase concession risk, including penalty concessions and defensive errors. Tactical performance indices were mainly explained by repetitive and cumulative ball-involvement variables, such as pass attempts, receptions, and open-play touches. Match-level visualizations further demonstrated that the proposed structural centrality indices capture role differentiation, tactical imbalance, and team operational structure beyond conventional box-score–based evaluation.
Study 2 develops Transformer-based multi-horizon time-series forecasting models using the structural performance indices derived in Study 1 and examines domain-specific temporal patterns by varying forecasting horizons (H) and observation window lengths (L). While all three domains exhibited meaningful predictability, distinct structural differences were observed across domains.
In the attacking domain, optimal observation windows varied across forecasting horizons, with test RMSE ranging from 0.0596 to 0.0610 and MAE from 0.0437 to 0.0442. Prediction errors gradually increased with longer horizons, indicating that attacking centrality is sensitive to recent match context and short-term event dynamics. Performance landscape analysis showed that mid-length observation windows (L = 4–6) were most effective for prediction.
In the defensive domain, longer observation windows (L = 7–8) were consistently optimal regardless of forecasting horizon. Although overall prediction errors were higher than in other domains (RMSE = 0.1643–0.1668), MAE slightly decreased as the forecasting horizon increased. This pattern suggests that defensive centrality exhibits high short-term variability but converges toward an average contextual state over mid- to long-term horizons.
The tactical domain exhibited the most stable predictive performance. Long observation windows (L = 8) consistently yielded optimal results, with test RMSE ranging from 0.0145 to 0.0148 and MAE from 0.0112 to 0.0114. This stability indicates that tactical centrality is shaped primarily by long-term structural factors, such as spatial occupation and sustained ball circulation patterns, rather than short-term events.
Permutation importance analysis further revealed strong autoregressive dependence in the attacking and tactical domains, whereas defensive centrality displayed relatively weak autoregressive effects. Instead, defensive predictions relied more heavily on attacking and tactical centrality indices, highlighting the reactive nature of defensive performance to overall match dynamics and opponent strategies.
Overall, this dissertation presents an integrated analytical framework that jointly accounts for structural interactions among players and their temporal evolution. The proposed performance indices provide interpretable insights into team structure and role differentiation, while the forecasting models reveal domain-specific temporal behaviors. Together, these findings extend soccer performance analysis beyond post hoc evaluation toward proactive monitoring and support mid- to long-term decision-making in tactical assessment, player role management, and team strategy planning.