Essays on credit risk modeling.

Summary Whether they are managers, employees, business partners, lenders, or investors, default prediction is a topic of paramount importance for all actors inside or outside the company. The first chapter of this thesis proposes a model for predicting the default of French SMEs based on their financial statements. Its discriminating power, measured by an accuracy ratio of 93.46%, gives it a real potential usefulness for an internal rating system by creditors. In the same vein, the second chapter looks at the predictive value of certain so-called "unconventional" data for anticipating default. It appears that the use of a chartered accountant to validate the forecast of SMEs is a guarantee of solvency, associated with a significantly lower default rate. The last two chapters focus on modeling the default of large international firms. On such portfolios with so few defaults, specific methodologies must be used. We then present a Shadow Rating model in the third chapter. With a replication rate close to 90%, we explain and predict the external ratings of large firms based on their financial statements and their sector. Finally, the last chapter focuses on the optimization of an internal model using machine learning. Combining artificial intelligence and human judgment, the proposed approach allows us to overcome their drawbacks, respectively lack of interpretability and subjectivity, to obtain an optimized model that is understandable, explainable, and compliant with banking regulations.