Enhancing Credit Risk Models with DFS and MIV

This paper is a technical report that discusses the enhancement of credit risk models at Revolut through the integration of Deep Feature Synthesis (DFS) and Marginal Information Value (MIV). It outlines the importance of generating predictive features from complex relational datasets for accurate risk assessment, highlighting the limitations of traditional manual feature engineering methods. The report explains how DFS serves as an automated technique to construct features from relational and temporal data, addressing the challenges faced by financial institutions in handling vast transactional datasets. Furthermore, the combination of DFS with MIV for feature selection is presented as a method to derive an automated credit acquisition scorecard generation process. The paper details how this approach enables the identification of complex patterns and interactions within the data, leading to improved model performance and more informed business decisions regarding lending and risk management, particularly for individuals with limited credit history.