New perspectives on recovery rates modeling and prediction

The main requirement for effective credit risk management is the sound quantification of default and recovery risk. This amounts to modeling the default probability of the counterparty as well as its recovery rate, the percentage of exposure that can be recovered upon default. As opposed to default probabilities however, recovery rates modeling remains widely unexplored and fixed recovery rate parameters are often used in practice. The scope of this thesis is to go beyond this unrealistic assumption and propose a new framework for modeling recovery rates on defaulted exposures, such as corporate bonds and non-performing loans (NPL). We begin by investigating the determinants of bond recovery rates, with the objective of identifying the drivers of recovery rate fluctuations in time. In this first part, we revisit a key result of previous studies and identify economic uncertainty as the most important systematic determinant of recovery rate distributions. We then tackle the problem of recovery rate prediction in the second part of the thesis. We first undertake a large scale benchmark study of machine learning methods for forecasting recovery rates on defaulted credit cards. We show how to derive behavioral predictors of recovery potential from bank recovery data and further identify a superior set of models for prediction. Finally, and building on the findings of the first two sections, we design a new modeling framework for bond recovery rates based on meta-learning, which is the combination of multiple machine learning algorithms. We provide industry participants with a set of best practice indications for recovery rates modeling and with methods associated to better performances, higher interpretability and lower model risk compared to those of traditional approaches.(ECGE - Sciences économiques et de gestion) -- UCL, 202