Samenvatting

Background. Between 20% and 35% of registered youth players disengage from amateur football clubs within any given season, most without formally withdrawing. Volunteer coaches, who average four to six hours per week on coaching activity, identify individual player feedback as the task they most frequently skip. The Super Star Academy (SSA) platform, developed by Epesi B.V., records player attendance, match selection, playing time, and disciplinary events as part of normal club operations, but this data had not been used to generate automated player risk assessments or personalised development feedback at the time this research began.

Objective. This study investigates whether structured platform data, collected through routine club operations, can support the early detection of player stagnation in amateur youth football and whether that detection can drive personalised development recommendations at a scale that volunteer coaching cannot achieve manually. The resulting system is called the Player Growth Engine.

Method. The research followed five CRISP-DM iterations on a static export of SSA platform data covering August 2024 to February 2026, comprising 129,000 activity rows across multiple Dutch amateur football clubs. A LightGBM classifier was trained on a 14-feature behavioural player profile constructed through relational data integration and 28-day rolling-window feature engineering. The target variable was a heuristic stagnation label derived from attendance drop, non-selection, and disciplinary event patterns. A leakage audit was conducted before any performance was reported. A retrieval-augmented generation (RAG) architecture was implemented for the recommendation component, combining a domain-specific SSA knowledge base with curated external coaching content, retrieved using the all-MiniLM-L6-v2 sentence transformer and filtered by player position and development pattern. The system was evaluated by a licensed football coach using a structured scoring instrument.

Key finding. Quartile analysis of the fair play participation ratio revealed a previously unnamed player subpopulation: ghost-benchwarmers, defined as players who attend matches consistently but receive little or no selection. Of the 754 players identified in this group, 64.1% met the stagnation criterion, compared to a platform-wide stagnation rate of 6.1%. Five dedicated ghost-benchwarmer features were engineered to represent this subpopulation. Adding these features resolved a precision ceiling that six other optimisation strategies had failed to break.

Results. The final LightGBM classifier (v7), trained on authenticated player accounts only, achieved precision 0.839, recall 0.927, F2-score 0.934, and AUC 0.998 on a held-out test set, meeting all five pre-defined success criteria. The RAG recommendation layer achieved Precision@3 of 0.73 and a position match rate of 0.80. Expert evaluation of the AI Coach component produced a content relevance score of 4.6/5 and an automatic response quality score of 4.0/5. The evaluating coach recommended a Socratic advisory structure, in which the system poses diagnostic questions before prescriptive advice, as the primary architectural change for a production deployment.



Conclusion. A unified player profile constructed from structured SSA platform data can support the early detection of stagnation and the delivery of personalised development feedback in amateur youth football. The classifier meets the precision threshold identified in the coaching tool adoption literature as the minimum required for volunteer coach trust.

The ghost-benchwarmer archetype is an original contribution to the sports engagement prediction literature; the pattern has not been named or systematically studied in prior research. Full production deployment requires longitudinal validation of the stagnation label against confirmed player dropout, a Dutch-language embedding model to reduce the current retrieval gap for Dutch-language content, and a formal privacy impact assessment before the system processes personal data in a club environment.