Comparative Analysis of Expected Goals Models: Evaluating Predictive Accuracy and Feature Importance in European Soccer

Expected Goals (xG) is a widely used metric in soccer analytics that estimates the probability of a shot resulting in a goal based on various characteristics of the shot. This study compares the predictive accuracy and feature importance of two prominent xG models: Opta and Understat. Using data from the top five European leagues from the 2017-2018 to the 2023-2024 seasons, we evaluate the predictive accuracy of each model using L1 and L2 loss metrics. Our findings indicate that Understat outperforms Opta in terms of lower prediction errors in the Bundesliga, Premier League, and Serie A, while Opta yields more stable predictions in La Liga and Ligue 1. We further analyze the factors influencing xG predictions through feature importance techniques using Random Forest and XGBoost models, complemented by SHAP (SHapley Additive exPlanations) analysis. Results reveal that goal exposure angle, shooting angle, and shot distance are key features in predicting goal probability, with differences in how categorical variables are weighted between the models. The study concludes with a discussion of the strengths, limitations, and league-specific applications of both models, highlighting the need for standardized data collection practices and expanded contextual features to enhance xG model utility and accuracy.