Application of Expected Goals (Xg) as a Predictive Metric in Brazilian Football

Caio Luiz Prates de Freitas; Maria Valeria Espinos Guerra Martins

DOI: 10.22167/2675-6528-202602750

Article derived from a Final Course Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

This study analyzed Expected Goals (xG) as a predictive metric applied to Brazilian football, focusing on the Campeonato Brasileiro Série A, seeking to evaluate its relationship with team performance and its potential and limitations in the national context, as well as its integration with machine learning techniques. The research was developed from a database built with public information from the FootyStats platform, covering five consecutive seasons (2021-2025), with data aggregated by team and season due to the limited availability of detailed information. For predictive analysis, a classification model was developed using the Random Forest algorithm, with xG for and xG against as explanatory variables, and team performance categorized into three classes (high, medium, and low) through quantile-based segmentation. The model was evaluated by accuracy, precision, recall, and F1-score metrics, and compared to a logistic regression baseline model. The results suggested a consistent association between xG metrics and team performance, with the xG balance standing out as a relevant indicator. Random Forest presented an accuracy of 0.50, superior to logistic regression (0.40), indicating moderate performance and a greater ability to identify high-performance teams, although with limitations in distinguishing between medium and low classes. It was concluded that xG has potential as a complementary metric for performance analysis in Brazilian football, and its combination with machine learning can contribute to predictive analyses, despite the limited data granularity and the reduced number of observations restricting the generalization of the results.

Keywords: data analysis; expected goals; football; prediction; random forest.

1. Introduction

Football, as a highly complex sport, presents match results frequently influenced by a series of random factors. Elements such as individual errors, refereeing decisions, and luck itself can significantly impact the outcome of a game. This unpredictable nature of football constitutes a considerable challenge for the objective analysis of team performance (ARAUJO, 2017).

Unlike other high-scoring sports, goals in football are relatively rare events. This characteristic implies that the final score of a match does not always fairly reflect a team’s performance quality. Consequently, traditional analysis approaches, which rely exclusively on match results or basic statistics such as ball possession and number of shots, demonstrate significant limitations in deeply understanding performance.

To overcome the gaps in conventional analyses, applied statistics in sports has evolved considerably in recent years. Among the most advanced metrics, Expected Goals, or xG, stands out, representing a statistical measure of the probability of a shot resulting in a goal. The calculation of xG considers various observable characteristics of the shot and the context of the play, such as the distance and angle of the shot in relation to the goal, the body part used, and the type of assist (EGGELS, 2016).

This metric assigns a probability between zero and one to each goal attempt, quantifying the quality of opportunities created by a team, regardless of the actual outcome of the play. By providing a more accurate view of the quality of goal chances, xG allows for a more objective assessment of team performance, less susceptible to the randomness inherent in football (EGGELS, 2016).

Despite its consolidation in European leagues and international academic literature, the application of xG in Brazilian football is still incipient. National championships have specific characteristics, such as competitive balance between teams, diversity of playing styles, and variations in pitch and climate conditions. These particularities highlight the importance of investigating the applicability and benefits of this metric in the Brazilian context.

However, the use of xG is not without criticism. Some analysts argue that the metric may not fully capture the complex context of plays, such as defensive pressure, the quality of the pass preceding the shot, or field conditions. Furthermore, the dependence on the variables used can lead to divergent results between different xG models, and in close-scoring matches, the aggregated xG may not correctly predict the winner, given that a single play can be decisive.

Another relevant issue for the national scenario lies in the limited availability of detailed public xG data for Brazilian football. Unlike European leagues, where specialized platforms offer granular data at the shot, match, and team level, sources for the Campeonato Brasileiro Série A generally provide only consolidated information per team and per season. This restriction in data granularity imposes methodological challenges and limits the capacity for more in-depth analyses.

Given these considerations, it becomes fundamental to investigate the potential of xG as a complementary tool for performance analysis in Brazilian football, contributing to more precise technical diagnoses and to the improvement of team planning and training processes. The metric offers a superior ability to represent the quality of opportunities created compared to traditional indicators, such as the number of goals, as highlighted by Eggels (2016). Thus, this study is justified by the need to develop a more robust understanding of team performance in the Campeonato Brasileiro Série A.

In this context, the present work aims to analyze Expected Goals (xG) as a predictive metric applied to Brazilian football, focusing on the Campeonato Brasileiro Série A, as well as to evaluate its integration with machine learning techniques in modeling team performance.

2. Material and Methods

This study was characterized as a quantitative research, with a methodological approach focused on the predictive analysis of football team performance. The central objective was to investigate the applicability of the Expected Goals (xG) metric in the Brazilian Championship Series A and its integration with machine learning techniques. To this end, a research strategy based on the collection, organization, and analysis of historical data was employed, aiming at the construction and evaluation of predictive models.

The unit of analysis consisted of the teams participating in the Brazilian Championship Series A. The database was built from public information made available by the FootyStats platform (FOOTYSTATS, 2025). The period covered by the data collection comprised five consecutive seasons of the competition, from 2021 to 2025, seeking to capture consistent patterns over time and increase the robustness of the analysis.

Due to the limited availability of detailed data for Brazilian football, the decision was made to use aggregated information per team and per season. Data collection was performed manually, extracting the metrics of interest directly from the tables available on the FootyStats platform’s web environment. After extraction, the data were converted to CSV (Comma-Separated Values) format, facilitating their manipulation and integration into analysis tools. The infeasibility of automated “web scraping” techniques (MITCHELL, 2018) was verified due to the website’s structure and protection mechanisms.

The collected variables included Expected Goals for (xG), Expected Goals against (xGA), goals scored, goals conceded, and the overall performance of the teams throughout the competition. These metrics were selected for their theoretical relevance in football performance analysis and the consistency of their availability in the data source. The xG metric, according to Eggels (2016), quantifies the probability of a shot resulting in a goal, considering characteristics such as distance and angle of the shot.

After collection, the data were subjected to a rigorous preprocessing procedure, which included steps for variable standardization, inconsistency treatment, and integrity verification. This phase was crucial to ensure the quality and reliability of the final database, as recommended in data science projects (PROVOST; FAWCETT, 2013). The data were consolidated into a single structured database, with approximately 100 observations, corresponding to twenty teams per season over the five seasons.

Column names were standardized to ensure a unique format, eliminating spaces and unifying nomenclature. Team identifications were standardized through a manual mapping dictionary, correcting variations such as abbreviations or acronyms. Season information was explicitly incorporated, and bases were integrated using the standardized team name and season as keys, with internal cross-referencing to keep only observations present in both sources and removal of duplicate records.

The dependent variable, representing team performance, was constructed from the final position in each season’s league table. To transform this continuous variable into a categorical one, the `qcut` function from the pandas library was used (MCKINNEY, 2017). This function allowed segmenting the teams into three groups of similar size, classified as “high”, “medium”, and “low” performance, based on quantiles, ensuring balance between classes and adapting the segmentation to the distribution of data for each season.

For predictive analysis, a classification model was developed using the Random Forest algorithm, a supervised machine learning method (BREIMAN, 2001). The explanatory variables used were xG for and xG against. The data were divided into training and testing sets, using the `train_test_split` function from the Scikit-learn library, with a proportion of 80% for training and 20% for testing (test_size = 0.2). The `random_state` parameter was set to 42 to ensure the reproducibility of the results.

The hyperparameters of the Random Forest algorithm were adopted in their default settings, with the exception of `random_state`. For comparative evaluation purposes, a baseline logistic regression (Logistic Regression) model was implemented, widely used in classification problems (JAMES et al., 2013). This model was built with the same explanatory variables and the same split between training and testing sets, allowing for a direct comparison with the performance of Random Forest.

The performance evaluation of the predictive models was carried out using standard metrics in multi-class classification problems. These included accuracy, which measures the proportion of correct predictions; precision, which indicates the proportion of correct positive predictions among all positive classifications; sensitivity (recall), which assesses the model’s ability to correctly identify true positive cases; and the F1-score, which represents the harmonic mean between precision and sensitivity (JAMES et al., 2013; BISHOP, 2006; MANNING; RAGHAVAN; SCHÜTZE, 2008).

3. Results and Discussion

The consolidation of the database, which gathered information aggregated by team and by season of the Brazilian Championship Series A between 2021 and 2025, allowed for the development and evaluation of the proposed predictive model. The final database, structured from data extracted from the FootyStats platform and converted to CSV format, demonstrated consistency and integrity suitable for the application of machine learning techniques. No critical inconsistencies were identified that could compromise subsequent analysis, ensuring the reliability of the data used in the study.

Although the use of aggregated data per team and season represents a limitation in terms of granularity, this approach proved effective in identifying structural performance patterns over time. This methodology allowed for the analysis of more stable trends, mitigating the influence of isolated events inherent to football, such as occasional goals or isolated refereeing decisions. Such a strategy was fundamental to the robustness of the analysis, considering the availability of data for the Brazilian football context.

Exploratory data analysis revealed significant patterns in performance metrics, particularly in variables related to Expected Goals. A strong association was observed between “xG” for and the overall performance of teams throughout the seasons. This finding indicates that teams with a greater capacity to generate volume and quality of goal-scoring opportunities tend to achieve better competitive results, reinforcing the importance of creating clear chances in football.

Complementarily, the “xG” against, or “xGA”, demonstrated an inverse relationship with team performance. Teams that presented lower “xGA” values, meaning they conceded fewer high-quality goal opportunities to opponents, generally exhibited more consistent campaigns. This behavior underscores the relevance of the balance between offensive and defensive performance, suggesting that competitive success depends not only on the ability to create opportunities but also on the skill to prevent dangerous chances from the opponent.

The Expected Goals balance, calculated as the difference between “xG” for and “xG” against, emerged as a particularly relevant indicator, synthesizing the overall performance of teams into a single metric. Throughout the analyzed seasons, it was observed that teams with a consistent positive balance tended to occupy higher positions in the league table. In contrast, teams with a negative balance showed a higher probability of inferior campaigns, a result that aligns with international literature and validates the applicability of “xG” as an effective performance evaluation metric (EGGELS, 2016).

The analysis of offensive efficiency was conducted by the ratio between the number of goals scored and “Expected Goals” (“xG”), allowing for an evaluation of the degree to which teams capitalize on the opportunities they create. This metric sought to capture the relationship between observed performance and expected performance, providing a direct measure of finishing effectiveness. Teams that consistently outperform their “xG” can be considered more efficient in converting chances, while those that fall short may indicate problems with finishing or decision-making.

Based on the selected variables, the classification model was developed using the Random Forest algorithm, aiming to identify patterns associated with team performance over the seasons. The choice of this algorithm proved suitable due to its ability to handle the non-linear nature of the relationships between variables and the presence of complex interactions among performance indicators, common characteristics in sports data (JAMES et al., 2013).

The model validation was performed by dividing the database into training and testing sets, ensuring the independence of observations. Despite the limitations arising from the use of aggregated data, the model demonstrated a moderate generalization capacity, as evidenced by the results obtained on the test set. This validation process is crucial to ensure that the model is not excessively adjusted to the training data, a phenomenon known as overfitting (BREIMAN, 2001).

In terms of global performance, the Random Forest model presented an accuracy of 0.50, indicating that half of the observations were correctly classified. Considering the multi-class nature of the problem, this result represents a performance superior to what would be expected in random classification, although it still reveals limitations in the model’s predictive capacity. Accuracy, as a general measure, offers an initial view of the model’s effectiveness in predicting team performance.

The analysis of performance metrics by class revealed notable differences in model behavior. The high-performance class achieved the best results, with a precision of 0.70, recall of 0.78, and an F1-score of 0.74. These values indicate a high capacity of the model to correctly identify teams with superior performance, demonstrating its effectiveness in recognizing patterns associated with top teams in the Brazilian Championship Series A.

On the other hand, the low-performance class presented reduced values for the precision, recall, and F1-score metrics, all equal to 0.25. This result highlights the model’s difficulty in correctly identifying teams classified in the low-performance category. The intermediate class, in turn, exhibited moderate performance, with an F1-score of 0.31, reflecting a greater overlap between the performance patterns of teams in this range, which made precise distinction difficult.

The aggregate metrics reinforced this pattern of behavior, with a macro average F1-score of approximately 0.43 and a weighted average close to 0.49. These values indicate a moderate and relatively balanced overall performance between classes, although classes with a larger number of observations exerted a more significant influence on the weighted averages. The distribution of classes, with approximately 33 observations for “high”, 33 for “low”, and 34 for “medium”, shows a balance in the sample used for model training and testing.

The confusion matrix provided a detailed analysis of the model’s behavior, revealing not only the number of correct predictions but also specific error patterns between classes. It was observed that the model demonstrated better performance in classifying the “high” class, with seven correct classifications out of a total of nine actual observations. This high accuracy rate reinforces the model’s ability to consistently identify top-performing teams, a relevant finding for sports analytics.

However, the matrix showed significant difficulties in classifying the “medium” and “low” classes. For the “medium” class, only two out of seven observations were correctly classified, while most were confused with the “high” class. This suggests considerable overlap between these categories in the feature space. In the “low” class, the model showed limited performance, with only one correct prediction out of four observations, and the errors were distributed among the other classes.

This error pattern indicates a tendency of the model to overestimate team performance, more frequently classifying observations into the upper categories. Such behavior can be attributed to the lower separation between the explanatory variables of the intermediate and lower classes, as well as to the reduced number of observations in certain categories. In the context of football, where small performance differences can lead to distinct rankings, this difficulty in distinguishing between close classes is consistent with the nature of the sport.

In general, the confusion matrix analysis reinforces that the model is more efficient in identifying extreme performance patterns, while presenting limitations in classifying intermediate scenarios. This observation points to opportunities for model improvement, possibly through the incorporation of more detailed variables that capture specific nuances of each team, such as tactical and situational aspects, which are not fully encompassed by season-aggregated data.

As a form of comparative evaluation, a baseline model was implemented using logistic regression. Both models were trained with the same explanatory variables (“xG” and “xGA”) and the same split between training and testing sets. The logistic regression model achieved an accuracy of 0.40, lower than the accuracy of 0.50 achieved by the Random Forest model. This result indicates that the use of a non-linear model, such as Random Forest, contributes to a better capture of the patterns present in the data, even if the performance gain is moderate.

The relatively small difference between the models suggests that part of the relationship between the explanatory variables and team performance can be explained by linear patterns. However, the superiority of Random Forest highlights the existence of more complex interactions that are better represented by ensemble-based methods. This corroborates the literature that emphasizes the ability of ensemble models to handle non-linear relationships and complex interactions between variables (JAMES et al., 2013).

It is fundamental to interpret the obtained results in light of the limitations of the database used. The absence of data at the play and event level restricts the model’s ability to capture more detailed game dynamics, which may impact its predictive accuracy. However, even with this limitation, the model demonstrated the ability to identify relevant structural patterns, reinforcing the utility of “xG” as an aggregate metric for performance analysis in Brazilian football.

The use of multiple seasons for the construction of the database contributed to mitigating some of the granularity limitations, increasing data variability and allowing for a more robust analysis. This approach is consistent with the recommendations in the data science literature for improving the generalization capability of predictive models (PROVOST; FAWCETT, 2013).

The results of this study reinforce the relevance of Expected Goals (“xG”) as an advanced metric for performance analysis in football, corroborating findings from international literature. As proposed by Eggels (2016), “xG” shows a greater capacity to represent the quality of chances created when compared to traditional metrics, such as the absolute number of goals. This characteristic was evidenced in both the descriptive analysis and the predictive modeling, where the variables derived from “xG” stood out as the most relevant.

The importance of the “xG” balance observed in the results is also aligned with the literature, which suggests that performance in football is more clearly explained by indicators that combine offensive and defensive aspects. In this sense, the fact that “xG” against (“xGA”) has high relevance in the model reinforces the idea that defensive solidity plays a fundamental role in the success of teams, often being as or more decisive than offensive capacity.

Additionally, the results related to offensive efficiency highlight an aspect widely discussed in studies on sports performance analysis: the existence of variability between expected performance and observed performance. As discussed by Spearman et al. (2017), football is a complex system, in which multiple factors influence the outcome. In this context, “xG” proves particularly relevant by incorporating probabilistic elements, reducing the influence of chance and providing a more objective view of performance.

The study’s findings present possibilities for practical application in the context of professional football, especially in activities related to “scouting” and performance evaluation. Although the model was built with data aggregated by team and season, its application demonstrates how “Expected Goals” metrics can complement traditional analyses based exclusively on goals scored, goals conceded, and league standing.

Performance analysis and professional club “scouting” departments can use “xG” metrics as additional indicators to evaluate the quality of opportunities created and conceded by teams, contributing to a more in-depth analysis of sporting performance. This approach allows for the identification of teams that exhibit superior or inferior performance compared to what is observed in their traditional results, aiding in opponent evaluation, match planning, and identification of performance trends.

Another possible application, although dependent on individualized data, is related to the process of player identification and evaluation. The comparison between goals scored and “xG” could be used to analyze athletes’ finishing efficiency. Players who consistently show goal production superior to their “xG” could be identified as potentially efficient finishers, while persistent differences could indicate performance characteristics that deserve further investigation for “scouting” and recruitment processes.

However, it is crucial to highlight that “xG” should not be used in isolation for hiring decisions or athlete evaluation. The efficiency observed in relation to “xG” can be influenced by several factors, such as the quality of opportunities, opponent characteristics, and variations associated with the analyzed period. Therefore, the practical use of the metric should occur in a complementary manner to other information available in performance analysis and scouting departments, integrating into a holistic view of performance.

A central aspect to consider is the trade-off between model complexity and its generalization capacity. Although the Random Forest algorithm presents mechanisms that reduce the risk of overfitting, the use of a relatively limited database in terms of the number of observations can restrict the predictive potential of the model. In this sense, the inclusion of multiple seasons contributed to increasing data variability and improving the robustness of the analysis, as recommended in the data science literature (PROVOST; FAWCETT, 2013).

In summary, the results obtained indicate that Expected Goals is configured as a relevant and informative metric for the analysis and prediction of performance in Brazilian football. Its consistent association with teams’ competitive performance, combined with the ability to identify extreme performance patterns, demonstrates the potential of “xG” as a complementary tool. The integration between probabilistic metrics and machine learning techniques, such as Random Forest, proved effective, paving the way for practical applications in sports analysis, scouting, and strategic decision-making, despite the limitations of data granularity and the difficulty in distinguishing intermediate and low performance classes.

4. Conclusion

This study aimed to analyze Expected Goals (xG) as a predictive metric in Brazilian football, focusing on the Campeonato Brasileiro Série A, and to evaluate its integration with machine learning techniques in modeling team performance. A consistent association was found between xG metrics and team performance over the seasons, with the xG difference standing out as a relevant indicator for analysis. For prediction, a classification model was developed using the Random Forest algorithm, which achieved an accuracy of 0.50, surpassing the baseline logistic regression model, which obtained 0.40. It was observed that Random Forest demonstrated a greater capacity in identifying high-performance teams, with a precision of 0.70 and recall of 0.78, although it exhibited limitations in distinguishing between medium and low performance classes. The analysis of offensive efficiency, by the ratio between goals scored and xG, reinforced the metric’s utility in identifying discrepancies between expected and observed performance, contributing to a deeper understanding of game dynamics. These findings reinforce the relevance of xG as a complementary tool for performance analysis in professional football, offering support for scouting activities and strategic planning.

Despite the contributions, the study faced limitations due to data granularity, which was aggregated by team and season, restricting the model’s ability to capture tactical and situational nuances at the match level. The reduced number of observations also impacted the model’s generalization, especially in distinguishing intermediate and lower performance classes. For future research, it is suggested to use more detailed databases, at the match or event level, and to incorporate additional variables, such as tactical and contextual metrics. It is also recommended to explore other machine learning algorithms and more robust validation techniques to enhance predictive capability. In summary, the integration between probabilistic metrics like xG and machine learning techniques has shown to be a promising path for the development of sports performance analysis in Brazilian football, broadening the understanding of the game and opening new opportunities for strategic decision-making.

Bibliographic References

ARAUJO, G. B. Uma caixinha de surpresas: como o componente aleatório afeta a tomada de decisões no futebol. Revista de Gestão e Negócios do Esporte (RGNE), São Paulo, v. 1, n. 1, p. 1-17, jan./jun. 2017.

BISHOP, 2006 [Referência completa não encontrada no documento original]

BREIMAN, L. Random Forests. Machine Learning, v. 45, n. 1, p. 5–32, 2001.

EGGELS, H. G. H. Expected Goals in Soccer: Explaining Match Results using Predictive Analytics. Eindhoven: Eindhoven University of Technology, 2016.

FOOTYSTATS. Campeonato Brasileiro Série A – Expected Goals (xG). Disponível em: https://footystats.org/pt/brazil/serie-a/xg. Acesso em: 20 jan. 2026.

JAMES, Gareth; WITTEN, Daniela; HASTIE, Trevor; TIBSHIRANI, Robert. An Introduction to Statistical Learning: with Applications in R. New York: Springer, 2013.

MANNING, Christopher D.; RAGHAVAN, Prabhakar; SCHÜTZE, Hinrich. Introduction to Information Retrieval. Cambridge: Cambridge University Press, 2008.

MCKINNEY, 2017 [Referência completa não encontrada no documento original]

MITCHELL, Ryan. Web Scraping with Python: Collecting More Data from the Modern Web. 2. ed. Sebastopol: O’Reilly Media, 2018.

PROVOST, Foster; FAWCETT, Tom. Data Science for Business: What You Need to Know about Data Mining and Data-Analytic Thinking. Sebastopol: O’Reilly Media, 2013.

SPEARMAN, W.; BASYE, A.; DICK, G.; FOGARTY, J.; FOSSDAL, T. et al. Physics-based modeling of pass probabilities in soccer. Journal of Quantitative Analysis in Sports, v. 13, n. 3, p. 1–10, 2017.

Article originating from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform