Adaptation of current scoring techniques to the needs of a credit institution: CFCAL-Banque.

Summary Financial institutions are confronted with various risks in the performance of their functions, including credit risk, market risk and operational risk. The instability of these factors weakens these institutions and makes them vulnerable to financial risks which, for their survival, they must be able to identify, analyze, quantify and manage appropriately. Among these risks, the one linked to credit is the most feared by banks, given its capacity to generate a systemic crisis. The probability of an individual moving from a non-risky state to a risky state is thus at the heart of many economic questions. In credit institutions, this issue is expressed in terms of the probability that a borrower will move from a "good risk" state to a "bad risk" state. For this quantification, credit institutions are increasingly using credit scoring models. This thesis focuses on current credit-scoring techniques adapted to the needs of a credit institution, the CFCAL-bank, specialized in mortgage-backed loans. In particular, we present two non-parametric models (SVM and GAM) whose performance in terms of classification is compared with that of the logit model traditionally used in banks. Our results show that SVMs perform better if we are only interested in the global predictive capacity. However, they exhibit lower sensitivities than the logit and GAM models. In other words, they predict defaulting borrowers less well. In the current state of our research, we advocate GAM models, which admittedly have lower overall predictive ability than SVMs, but give more balanced sensitivities, specificities, and predictive performance. By highlighting targeted credit scoring models, applying them to real mortgage data, and comparing them through their classification performance, this thesis makes an empirical contribution to the research on credit scoring models.