Predicting repayment delays in Italian real estate crowdfunding: Evidence from a machine learning approach

This paper develops a machine learning model for early risk detection in real estate crowdfunding (RECF), focusing on repayment delay. Drawing on adverse selection and signaling theories, this study interprets borrower characteristics, deal-level attributes, and platform-generated risk indicators as informative mechanisms that may reduce information asymmetries. Using projectlevel data from an Italian lending-based platform, we compare four classification algorithms to predict whether a project will experience delayed capital repayment. Based on out-of-fold predictive performance and threshold-free discrimination metrics, Gradient Boosting is selected as the primary model. The predictive model is complemented by a continuous risk scoring system that supports portfolio-level monitoring and targeted intervention. Model interpretability is ensured through SHapley Additive exPlanations (SHAP), which clarify the contribution of key project variables to delay risk predictions. The proposed model contributes to the fintech lending and real estate crowdfunding literature by reframing credit risk assessment around repayment timing and by providing a valuable decision-support tool for platforms and investors, enhancing transparency, proactive risk governance, and data-driven resource allocation.