Goal and shot prediction in ball possessions in FIFA Women’s World Cup 2023: a machine learning approach

Abstract

[Abstract]: Introduction: Research in women’s football and the use of new game analysis tools have developed significantly in recent years. The objectives of this study were to create two predictive classification models to forecast the occurrence of a shot or a goal in the FIFA Women’s World Cup 2023 and to identify the associated technical-tactical indicators to these outcomes. Methods: A total of 2,346 ball possessions were analyzed using an observational design, mapping two different target variables (Success = Goal and Success2 = Goal or Shot) with a relative frequency of 1.28 and 8.35%, respectively. The predictive capacity was tested using Random Forest and XGBoost and finally and SHAP values were calculated and visualized to understand the influence of the predictors. Results: Random Forest technique showed greater efficacy, with recall and sensitivity above 93% in the resampled dataset. However, recall on the original test sample was 13% (Success = Shot or Goal) and 0% (Success = Goal), demonstrating the models’ inability to predict rare events in football, such as goals. The indicators with the greatest influence on the outcome of these possessions were related to the possession zone, attack duration, number of passes, and starting zone, among others. Conclusion: The results highlight the need to incorporate a greater number of predictive variables in the models and underline the difficulty of predicting events such as goals and shots in women’s football.