Abstract Energy poverty (EP) is considered an urgent challenge, intensified by rising energy costs, economic inequality, and the transition toward green energy, which involves many Western countries. By referring to Italy, this study employs machine learning algorithms (MLAs) to predict and classify EP using official Household Budget Survey (HBS) data. To evaluate EP, the study compares several MLAs alongside three expenditure-based indicators proposed in three seminal articles by Hills, Faiella and Lavecchia, and Betto et al. Among these, the indicator developed by Betto et al., which accounts for regional and socioeconomic disparities, consistently outperforms the others across all MLAs, demonstrating higher accuracy, precision, and recall. This ensures a more comprehensive identification of energy-poor households. The analysis highlights the significant impact of data imbalance on model performance, emphasizing the need for techniques such as SMOTE and undersampling. The superior performance of the Betto et al. indicator underscores its potential as a benchmark for EP measurement, providing a valuable tool for policymakers to design targeted interventions, allocate resources effectively, and support a just and sustainable energy transition. The study reinforces the importance of dynamic, data-driven approaches to address EP, and calls for improved data collection to enhance prediction accuracy and policy effectiveness.
Scandurra et al. (2026) studied this question.