Using AI to assess credit risk; developing a model

Credit scoring methodologies have been widely used since the 1960s, to measure the probability of default of the applicants, in particular in consumer credit market. Traditionally, approaches based on regressive techniques have been used, but more recently models based on artificial intelligence have seen considerable diffusion due to their greater discriminating capacity. PSD2 has enabled the creation of a single data infrastructure usable by all the institutions authorized by the current account owner. The PSD2 purpose is current accounts aggregation and reconciliation, however leveraging on GDPR data portability, current account information uniformed thanks to PSD2, may be used to evaluate credit default probability. Moreover, data categorization necessary to address PSD2 objectives, represents a significant boost for model accuracy enabling the opportunity to move model philosophy from a pure financial analysis standpoint to a wider lifestyle behavior credit risk related analysis. This paper illustrates an operative use case of credit scoring model development using AI algorithms, comparing it with traditional regressive model performance. In particular, many methodologies have been tested in order to find the optimal model searching among methods in literature and also trying to figure-out the non-linear correlation differences with respect to the target variable. So, at the end of this process, the improvements of AI algorithms have been quantified, knowing that they boost the ability to keep improving thanks to their capacity to overcome the classic static model view.