[Paper Review] A Machine Learning Approach for Player and Position Adjusted Expected Goals in Football (Soccer)

James H. Hewitt, Oktay Karakuş|arXiv (Cornell University)|Jan 19, 2023

Sports Analytics and Performance7 references4 citations

TL;DR

This paper proposes a machine learning-based Expected Goals (xG) model that incorporates player and positional adjustments using StatsBomb event data, improving accuracy by integrating features like goalkeeper positioning, defender proximity, and shot technique. It demonstrates that Lionel Messi's xG efficiency exceeds even top forwards by 347 goals over the model's predictions, validating the method's ability to capture elite player quality beyond standard xG metrics.

ABSTRACT

Football is a very result-driven industry, with goals being rarer than in most sports, so having further parameters to judge the performance of teams and individuals is key. Expected Goals (xG) allow further insight than just a scoreline. To tackle the need for further analysis in football, this paper uses machine learning applications that are developed and applied to Football Event data. From the concept, a Binary Classification problem is created whereby a probabilistic valuation is outputted using Logistic Regression and Gradient Boosting based approaches. The model successfully predicts xGs probability values for football players based on 15,575 shots. The proposed solution utilises StatsBomb as the data provider and an industry benchmark to tune the models in the right direction. The proposed ML solution for xG is further used to tackle the age-old cliche of: 'the ball has fallen to the wrong guy there'. The development of the model is used to adjust and gain more realistic values of expected goals than the general models show. To achieve this, this paper tackles Positional Adjusted xG, splitting the training data into Forward, Midfield, and Defence with the aim of providing insight into player qualities based on their positional sub-group. Positional Adjusted xG successfully predicts and proves that more attacking players are better at accumulating xG. The highest value belonged to Forwards followed by Midfielders and Defenders. Finally, this study has further developments into Player Adjusted xG with the aim of proving that Messi is statistically at a higher efficiency level than the average footballer. This is achieved by using Messi subset samples to quantify his qualities in comparison to the average xG models finding that Messi xG performs 347 xG higher than the general model outcome.

Motivation & Objective

• To develop a more accurate Expected Goals (xG) model by incorporating player-specific and positional-specific performance adjustments.
• To address limitations in existing xG models that overlook factors like defender positioning, goalkeeper location, and shot technique.
• To quantify and validate the superior shooting efficiency of elite players like Lionel Messi using a data-driven, player-adjusted xG framework.
• To provide a reproducible, transparent xG model using publicly available StatsBomb data, enhancing academic and industry access to advanced football analytics.
• To explore whether positional subgroups (forwards, midfielders, defenders) differ significantly in xG conversion efficiency, and whether individual players exceed these norms.

Proposed method

• The study frames xG prediction as a binary classification problem, using logistic regression and gradient boosting (XGBoost) to predict goal probability from 15,575 shots.
• Key features include shot location (distance and angle), defender pressure radius, goalkeeper positioning, and shot technique (e.g., volley, overhead kick).
• The model is trained and validated using StatsBomb event data, with performance benchmarked against industry standards like StatsBomb, FBRef, and infogol.com.
• Position-adjusted xG is implemented by splitting training data into three subgroups—Forwards, Midfielders, and Defenders—to evaluate positional differences in shooting efficiency.
• Player-adjusted xG is applied specifically to Lionel Messi’s shots, comparing his actual xG performance to the general model output to quantify his elite efficiency.
• Model performance is evaluated using game-level xG timelines and shot maps, with cross-validation against real match outcomes (e.g., 2018 Champions League final).

Figure 1: A freeze frame example: shot location (distance and angle)

Experimental results

Research questions

• RQ1 How does incorporating goalkeeper positioning and defender pressure improve xG prediction accuracy compared to standard models?
• RQ2 Do forwards, midfielders, and defenders differ significantly in their expected goal conversion efficiency when adjusted for position?
• RQ3 To what extent does Lionel Messi’s shooting performance exceed the predictions of a general xG model, indicating superior individual quality?
• RQ4 Can a machine learning-based xG model with enhanced features outperform commercial xG metrics in real match scenarios?
• RQ5 How do different shot techniques (e.g., volleys, overhead kicks) affect xG values, and do current models overestimate their likelihood?

Key findings

• Forwards demonstrated the highest xG conversion efficiency, followed by midfielders and defenders, both in general and position-adjusted models.
• The proposed xG model achieved a total xG of 3.4 for the 2018 Champions League final, closely aligning with StatsBomb (2.7), FBRef (3.4), and infogol (3.59).
• Lionel Messi’s xG performance exceeded the general model’s prediction by 347 goals, indicating a statistically significant efficiency advantage over average players.
• The model predicted a higher xG for Gareth Bale’s overhead kick (0.131) than StatsBomb (0.023), suggesting potential overestimation of such high-skill shots in commercial models.
• Shot technique and defender pressure were significant predictors of xG, with volleys and overhead kicks showing notable discrepancies between commercial and proposed models.
• The inclusion of goalkeeper positioning and opposition between shooter and goal significantly improved model accuracy, addressing a key gap noted in prior literature.

Figure 2: (a) Baseline model results with existing 26 features. (b) Baseline model performance after adding extra features. (c) Proposed xG model performance.