Expected Goals (xG) models have become a central tool in modern football analytics, yet their exclusive focus on shot quality imposes structural limitations when used to evaluate team performance or predict match outcomes. This study conducts a three-stage analysis aimed at critically reassessing the role and reliability of xG using publicly available data. A first highly effective xG model was developed by expanding the engineered feature set, leveraging freeze-frame positional context and contextual descriptors, to achieve top-tier accuracy using publicly available data. Once it was ensured that the model performed better than current models, the operational framework was made public to facilitate the model's interpretability and the model's limitations were then analyzed and discussed. To go beyond these limitations, a complementary event-based performance metric was then introduced, aggregating attacking and defensive behaviours over the full match and capturing dimensions of performance that go beyond shot-based information. Finally, a combined approach was evaluated by integrating team-level xG features into the event-based model. Using historical data from men's elite competitions, the event-based model was shown to slightly outperform xG in predicting match outcomes, while the combined model achieves the highest overall effectiveness. These results indicate that xG alone cannot fully explain match dynamics, but remains highly informative when embedded within broader representations of team behaviour. The findings highlight the need for more comprehensive performance metrics and clarify the complementary roles of shot-based and event-based information in football analytics.
I modelli di Expected Goals (xG) sono diventati uno strumento centrale nella moderna analisi calcistica, tuttavia il loro focus esclusivo sulla qualità del tiro impone limitazioni strutturali quando vengono utilizzati per valutare la performance di squadra o prevedere gli esiti delle partite. Questo studio conduce un’analisi in tre fasi, finalizzata a rivalutare criticamente il ruolo e l’affidabilità degli xG utilizzando dati pubblicamente disponibili. Un primo modello xG ad alte prestazioni è stato sviluppato ampliando l’insieme di feature ingegnerizzate, sfruttando il contesto posizionale dei freeze-frame e descrittori contestuali, per ottenere prestazioni di livello top utilizzando dati pubblicamente disponibili. Una volta assicurato che il modello ottenesse prestazioni migliori rispetto ai modelli attuali, il framework operativo è stato reso pubblico per facilitare l’interpretabilità del modello e le sue limitazioni sono quindi state analizzate e discusse. Per andare oltre queste limitazioni, è stata poi introdotta una metrica di performance complementare basata sugli eventi, che aggrega i comportamenti offensivi e difensivi sull’intera partita e cattura dimensioni della prestazione che vanno oltre l’informazione basata sui tiri. Infine, è stato valutato un approccio combinato integrando feature xG a livello di squadra nel modello basato sugli eventi. Utilizzando dati storici delle competizioni maschili d’élite, è stato mostrato che il modello basato sugli eventi supera leggermente gli xG nella previsione degli esiti delle partite, mentre il modello combinato raggiunge la migliore performance complessiva. Questi risultati indicano che gli xG da soli non possono spiegare completamente le dinamiche di una partita, ma rimangono altamente informativi quando inseriti all’interno di rappresentazioni più ampie del comportamento di squadra. I risultati evidenziano la necessità di metriche di performance più complete e chiariscono i ruoli complementari delle informazioni basate sui tiri e sugli eventi nell’analisi calcistica.
A critical analysis of the Expected Goals (xG) metric in professional soccer
ROMANO, LUCA
2024/2025
Abstract
Expected Goals (xG) models have become a central tool in modern football analytics, yet their exclusive focus on shot quality imposes structural limitations when used to evaluate team performance or predict match outcomes. This study conducts a three-stage analysis aimed at critically reassessing the role and reliability of xG using publicly available data. A first highly effective xG model was developed by expanding the engineered feature set, leveraging freeze-frame positional context and contextual descriptors, to achieve top-tier accuracy using publicly available data. Once it was ensured that the model performed better than current models, the operational framework was made public to facilitate the model's interpretability and the model's limitations were then analyzed and discussed. To go beyond these limitations, a complementary event-based performance metric was then introduced, aggregating attacking and defensive behaviours over the full match and capturing dimensions of performance that go beyond shot-based information. Finally, a combined approach was evaluated by integrating team-level xG features into the event-based model. Using historical data from men's elite competitions, the event-based model was shown to slightly outperform xG in predicting match outcomes, while the combined model achieves the highest overall effectiveness. These results indicate that xG alone cannot fully explain match dynamics, but remains highly informative when embedded within broader representations of team behaviour. The findings highlight the need for more comprehensive performance metrics and clarify the complementary roles of shot-based and event-based information in football analytics.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Romano_Executive Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary della tesi
Dimensione
413.75 kB
Formato
Adobe PDF
|
413.75 kB | Adobe PDF | Visualizza/Apri |
|
2026_03_Romano_Thesis.pdf
accessibile in internet per tutti
Descrizione: Testo tesi
Dimensione
1.07 MB
Formato
Adobe PDF
|
1.07 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/250581