A Novel Supervised-Unsupervised Approach for Past-Due Prediction

Abstract

In the current landscape of banking and financial services, a primary concern for industry practitioners revolves around predicting the probability of default (PD) and categorizing raw data into risk classes. This study addresses the challenge of predicting payment past-due for customers of Residential Mortgage-Based Securities (RMBS) and Small and Medium Enterprises (SMEs) within the Italian banking sector, employing an innovative approach that integrates a classification model (Random Forest) with an anomalies detection technique (Isolation Forest). The models are trained on a substantial dataset comprising performing loans from the 2020-2022 period. Notably, this research stands out not only for its novel modeling approach but also for its focus on the arrear status of RMBS and SME customers as the target variable. By concentrating on past-due rather than the broader concept of probability of default, this approach enhances understanding of customers' financial stress levels, enabling proactive monitoring and intervention by decision-makers. The ultimate aim of this experimentation is to develop a robust and effective algorithm applicable in real-world scenarios for predicting the likelihood of past-due among individual customers and companies, thereby supporting management decision-making processes. Empirical results demonstrate that the proposed framework surpasses conventional statistical and machine learning algorithms in credit risk modeling, exhibiting robust performance on new data (validated against 2023 data) and thus proving its operational suitability.

Keywords

• past due
• credit risk