The analysis of technical-tactical performance in men’s football began to develop significantly in the late 20th and early 21st centuries (Hughes and Bartlett, 2002), using notational and observational records (Preciado et al., 2019). Later, with the use of new technologies, this analysis started to be conducted on data obtained from positional sensors such as Global Positioning System (GPS) and Local Positioning System (LPM) (Low et al., 2020). In the case of women’s football, the lower participation of women in the sport and a lack of social and research interest delayed the publication of the first studies by more than a decade (Kirkendall, 2007; Mara et al., 2012; Leite, 2013). A significant increase in research occurred starting in 2020, coinciding with the FIFA Women’s World Cup 2019 (Lee and Mills, 2021; Iván-Baragaño et al., 2022; Kubayi, 2022) and, later, with the FIFA Women’s World Cup 2023 (Branquinho et al., 2024; Bradley, 2025a, 2025b; Iván-Baragaño et al., 2025; Oliva-Lozano et al., 2025).

Currently, Artificial Intelligence, and Machine Learning in particular, have become topics of interest for researchers and practitioners (Nassis et al., 2023; Rico-González et al., 2023), who have conducted studies with various objectives, such as establishing differences between men’s and women’s football (Pappalardo et al., 2021), predicting injury risk (Robles-Palazón et al., 2021), or the probability of success of different types of actions, such as entries into the penalty area (Iván-Baragaño et al., 2021; Stival et al., 2023) or shots during set-piece situations (Maneiro et al., 2019). In all of these studies, different regression and/or classification models were trained with the aim of predicting outcomes or future behaviors.

More recently, other studies have attempted to apply more complex strategies, materialized in the use of various techniques based on deep neural networks. Among the different examples of the use and application of Artificial Intelligence in the analysis of high-performance football, the article by AlMulla et al. (2023) trained a deep neural network model (Gated Recurrent Unit) to predict the outcomes of football matches in the Qatari league over 10 consecutive seasons, using data from data providers. Similarly, Wang et al. (2024) trained and evaluated a generative AI model based on deep learning and graph methods, which allowed the generation of execution proposals for set-piece actions. This was part of an unusual collaboration between Google DeepMind and Liverpool FC. Despite this, and in agreement with Claudino et al. (2019) the synergy that Artificial Intelligence needs to create alongside football still requires further development in the coming years.

This gap is even more pronounced in the case of women’s football, with scarce scientific evidence where AI or ML has been applied to female samples. In this regard, some authors have sought to understand the differences between men’s and women’s football (Pappalardo et al., 2021) using supervised ML techniques and applying explainability methods such as SHAP values (Lundberg and Lee, 2017). On the other hand, other studies have conducted analyses of offensive play using supervised techniques such as binary logistic regression (Iván-Baragaño et al., 2022), multinomial logistic regression (Casal et al., 2023), or decision trees (Maneiro et al., 2019). Additionally, some authors (Shen et al., 2024) have proposed models focused on convolutional neural networks and computer vision to determine offensive positioning in women’s football, using images extracted from UEFA Women’s Champions League matches.

In any case, and as a common aspect of studies conducted using supervised machine learning classification techniques, most studies have been carried out using methods characterized by high intrinsic explainability (such as decision trees or logistic regression), but often with moderate performance. In this context, there is a need to improve the performance of predictive models applied to a chaotic and non-linear reality like football, without sacrificing interpretability, to ensure the application of these studies’ results to training and competition.

For the reasons mentioned above, the objective of this study was twofold. First, it aimed to create two binary classification models that would allow the prediction of the outcome of ball possessions in elite women’s football (i.e., whether the possessions end in a Goal or a Shot). Additionally, once these models were trained, the SHAP library was implemented to identify the technical-tactical performance indicators that had the greatest influence on the model.

For both recoding 1 and recoding 2, the Random Forest algorithm demonstrated higher performance compared to the XGBoost algorithm. The combination of hyperparameters that provided the best performance for recoding 1 (Goal or Shot) was Random Forest: (i) n_estimators = 200, (ii) max_depth = None, (iii) min_samples_split = 5, (iv) min_samples_leaf = 1, Bootstrap = False. Similarly, for recoding 2 (Goal), the best performance was achieved with the following Random Forest combination: (i) n_estimators = 300, (ii) max_depth = None, (iii) min_samples_split = 5, (iv) min_samples_leaf = 1, Bootstrap = False.

The objective of this study was, first, to create two binary classification models that could predict the outcome of ball possessions in elite women’s football. Additionally, once the models were trained, the aim was to identify the technical-tactical indicators associated with a higher probability of achieving a goal or a shot during ball possessions. To achieve these objectives, a mapping of the Possession Outcome variable was performed based on the degree of success (Goal or Shot). Following this, oversampling of the imbalanced class was conducted.

Previous studies have employed similar procedures with the aim of predicting the outcome of ball possessions in women’s football. However, in most of these studies, success was defined as reaching the penalty area, reaching the final third, or, more generally, the creation of Goal Scoring Opportunities (Scanlan et al., 2020; Kubayi, 2022; Mitrotasios et al., 2022; Mesquita et al., 2023). This aspect is crucial when training a classification model, as approximately one in four (25%) ball possessions in women’s football ends with a move into the final third or the opponent’s penalty area (Iván-Baragaño et al., 2021; Casal et al., 2023), allowing a balance between correctly classified positive and negative cases. In contrast, in this study, the dataset showed a percentage of positive cases of 1.35 and 8.28%, respectively, which necessitated oversampling of the imbalanced classes to prevent the model from ignoring the minority class (Haller et al., 2023).

The classification models yielded excellent results on the resampled datasets, with recall and specificity exceeding 93% in both models. However, their performance on the original datasets was poor. When predicting shots or goals, the model had a recall of 13%, and in the case of goal prediction, the model did not predict any positive outcomes. These results highlight the difficulty of predicting infrequent events in football, such as shots and goals, and underscore the need for incorporating a larger number of predictor variables, as well as further tuning the hyperparameters during model training. Similarly, as seen in injury prediction, where different studies have shown recalls between 10 and 15% (Haller et al., 2023; Majumdar et al., 2024), the holistic nature of the sport contributes to the challenge of accurately predicting such events.

In relation to the SHAP technique (Lundberg and Lee, 2017) applied in this study, it was found that a large number of indicators associated with ball possessions contributed to increasing the probability of a favorable outcome for the executing team. The performance indicators associated with successful ball possessions in elite women’s football observed in this work largely align with previous studies on this topic. In this regard, Maneiro et al. (2022) demonstrated that developing ball possessions in the opponent’s half increased the likelihood of the possessions ending with a delivery into the penalty area. Similarly, the offensive tactical intent once ball possession was initiated, or the number of passes made in the offensive sequence, were variables that significantly altered the outcome of ball possessions in women’s football (Scanlan et al., 2020; Iván-Baragaño et al., 2021; Casal et al., 2023).

However, considering that the level of success analyzed in this study was higher than in previous studies, this work also demonstrated the existence of variables that had not previously shown a multivariate influence on the outcome of ball possessions. For example, in the study by Iván-Baragaño et al. (2022), it was observed that the current match score had an influence on the development and outcome of ball possessions. Thus, it is interesting to note that, while success in delivering the ball into the penalty area can be influenced by the flow of the game, when it comes to taking a shot or scoring a goal, this variable does not have sufficient influence. This insight may have significant implications for the sport, as it could suggest that when teams are losing, they tend to deliver the ball into the penalty area more often but are less successful in converting these deliveries into shots or goals.

Similarly, it is interesting to analyze the influence of the variables Time (5Q and 6Q) and Start Zone Width (Central). According to the SHAP values generated for these variables, the following insights can be drawn. When predicting a shot, the likelihood increases if the possession occurs in the last 15 min of the match (6Q). However, when analyzing the SHAP values for the positive outcome “Goal,” the probability increases between the 60th and 75th minutes of the match. This contradicts the findings from the 1999, 2003, and 2007 World Cups, where a higher number of goals were observed in the final 15 min of the match (Armatas et al., 2007), as well as the results from the most recent Women’s Euro 2022 (Sanmiguel-Codina et al., 2025).

Additionally, the observation that starting an attack in the central lane (Start Zone Width = Central) increases the probability of success had not been noted in previous studies (Scanlan et al., 2020; Iván-Baragaño et al., 2021; Maneiro et al., 2022). This may suggest that while starting attacks from wide areas may facilitate successful entries into the penalty area, shots and goals are more likely to result from attacks initiated in central zones.

This study presents several limitations that should be addressed in future research. First, while the classification models achieved excellent performance on oversampled datasets using the SMOTE technique, their ability to detect true positives in the original dataset was notably poor. From a football perspective, this suggests that the actions leading to dangerous situations may follow highly specific patterns that generic classification models, such as Random Forest, are unable to effectively capture. In this context, future studies might benefit from the implementation of advanced statistical techniques like T-Patterns, which have proven effective in identifying offensive patterns and sequences in other sports (Pic and Jonsson, 2021; Pic et al., 2021). Additionally, exploring alternative tools to mitigate overfitting during model training is essential. Expanding the dataset by analyzing additional championships could also enhance the robustness of the identified patterns related to goal scoring. Furthermore, the inclusion of certain predictor variables, such as Match Outcome, was found to influence model performance, not due to their predictive capability, but because of their retrospective causal relationship (e.g., the winning team scored more goals). This introduces data leakage during model training. Consequently, future research should consider excluding such variables from the training process to ensure more reliable and generalizable results.

The models trained and tested in this study showed excellent performance on the resampled datasets using the SMOTE technique (Last et al., 2017). However, when these models were evaluated on the original dataset, their performance was low or non-existent. In the case of predicting Goals or Shots, the model achieved a recall of 13%, which slightly increased the relative frequency of the positive class but fell far short of an acceptable performance. For goal prediction, the model was unable to output the positive class at all. Based on this, it can be stated that such events in elite women’s football possess very specific characteristics and patterns that cannot be clearly defined or that, at least, involve variables not analyzed in this study.

On the other hand, the SHAP explainability techniques applied in this study allowed for the identification of various variables associated with the achievement of goals and shots. Some of these variables showed similarities to previous studies, where success was categorized as entries into the penalty area or similar metrics. However, other variables such as start zone width, timing, or defensive intent had a significant influence on the model when analyzing a higher degree of success, enabling a tactical understanding of how these types of actions occur.