Machine learning framework for expected goals inference: bridging statistical modeling and causal understadning in football analytics

Football analytics has evolved from descriptive statistics to sophisticated machine learning models, yet much of the existing research remains limited to isolated predictive or exploratory approaches. This study proposes an integrated analytical framework that combines supervised, unsupervised, and causal inference methodologies to advance the modeling of Expected Goals (xG) and deepen the understanding of shot quality and decision-making in football. Supervised models, including logistic regression, decision tree, random forests, gradient boosting, xgboost and neural network were employed to predict goal probabilities, while unsupervised methods such as Principal Component Analysis and K-Means clustering uncovered latent shot structures and tactical typologies. Isolation Forest identified rare or anomalous events, enhancing sensitivity to high-value but infrequent actions, and a Counterfactual Autoencoder introduced a causal layer that simulated “what-if” scenarios, quantifying how controlled changes in shot context, could affect the expected outcome. Experimental results demonstrate that ensemble models improve discriminative power, while unsupervised analyses reveal interpretable clusters that align with tactical behaviors on the pitch. The counterfactual component extends the framework beyond association, enabling causal insight into the determinants of chance creation. Collectively, these findings illustrate a progression from purely predictive modeling toward interpretable, context-aware, and prescriptive analytics. By bridging statistical accuracy with tactical relevance, this work contributes to the emerging paradigm of data-driven causal understanding in sports analytics, transforming raw performance data into actionable football intelligence.