Predicting buildings' EPC in Italy: a machine learning based-approach

EU member states have committed to achieving carbon neutrality by 2050. Given that building-related activities contribute to almost a quarter of EU greenhouse gas emissions, the implementation of new regulations to decarbonize this sector is paramount. However, policymakers must carefully evaluate policies to mitigate transition risks associated with these regulations, as buildings represent a significant portion of household wealth and bank assets. Accurate metrics regarding buildings' energy efficiency, including the energy class reported in Energy Performance Certificates (EPCs), are essential for such evaluations. In this study, we developed a machine learning-based model to predict the energy class of Italian buildings using publicly accessible data. The model, trained on a geographic subset of the Italian territory, achieves a 37% accuracy rate, which increases to 74% when allowing for a one-class margin of error (1-class accuracy). Further testing against a mortgage portfolio provided by a commercial bank yielded a 69% 1-class accuracy. Comparison with statistics reported by the official EPC register (SIAPE) suggests a potential discrepancy in the representation of the worst energy efficiency class.