Athens University Economics and Business (AUEB)
Discrete, continuous and machine learning models with applications in credit risk
Abstract
dc:descriptionCredit risk modelling is a versatile and dynamic area of financial mathematics, with important practical implications, as has been historically established. In particular, the last financial crisis made it clear that credit risk models had to become more rigorous. For this reason, the recent International Financial Reporting Standards (IFRS) 9 have introduced a framework making credit risk modelling forward-looking, thereby increasing the need for robust mathematical tools. The goal of this thesis is to develop and explore such mathematical tools and models, motivated by specific open problems that arise due to these regulations, and develop frameworks that are both mathematically-sound and can be efficiently applied by practitioners. We begin with discrete models, specifically Markov chains which are well established in the field of credit risk, and develop a framework that can be implemented by financial institutions for credit rating reporting and compliance purposes under IFRS 9. Subsequently, we consider continuous-time stochastic models and study how these can be used for probability of default estimation and provisioning calculations. Specifically, we use a general family of models to incorporate various latent variables on which the credit exposure may depend, and use approaches relying on Integral and Partial Integro-differential Equations to describe and prove important mathematical properties of the resulting probability of default process. To show how these mathematical tools can be implemented by practitioners, we develop and study numerical methods that can be used for the estimation of default probabilities. We use the well-known Finite Difference methods, which we apply to the equations that arise under various types of stochastic models, to illustrate the variety of uses these approaches can find. Lastly, we draw from recent research to also consider how the modern field of machine learning models, and particularly the family of Deep Neural Networks, can be used to estimate the default probabilities, and discuss important theoretical and practical considerations that should be taken into account when comparing these to the Finite Difference methods.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Georgiou, Kyriakos-Kristofer
- Γεωργίου, Κυριάκος-Κρίστοφερ
Subjects
dc:subject × 13Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/54809