Essays on credit risk

• Ano de defesa: 2023
• Autor(a) principal: Silva, Rafael Scopel
• Orientador(a): Martins-da-Rocha, Victor Filipe, Chela, João Luiz
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Palavras-chave em Português: IFRS 9; Bancos; Risco de crédito; Perdas esperadas de crédito; Provisões para perdas de crédito; Probabilidade de inadimplência; Projeção; Florestas sobrevivência; Dados truncados e com censura; Perda dado default; Modelos de retomada; Modelos de ajuste de preço
• Palavras-chave em Inglês: Banks; Credit risk; Expected credit losses; Loan loss provisions; Loan loss reserves; Probability of default; Forecasting; Random survival forests; Left-truncated and right-censored data; Loss given default; Mortgages; Repossession; Haircuts
• Link de acesso: https://hdl.handle.net/10438/33457

This thesis consists of three independent chapters. The first chapter is a post- implementation review of IFRS 9 with a focus on the loan loss reserves (LLR) level and volatility by analyzing quarterly financial reports of two systemically important financial institutions in Brazil. First, we investigate the impact on loan loss reserves level, finding statistically significant evidence that IFRS 9 resulted in a rise of LLR for our sample. Next, we analyze the impact on loan loss reserves volatility. The results for LLR volatility show average volatility increased after IFRS 9 implementation, albeit at different levels, for each bank under analysis indicating that the incorporation of lifetime losses and macroeconomic expectations led to increased volatility. In chapter two, we inspected the survival analysis approach to credit risk modeling with the incorporation of time-varying covariates in the presence of truncation and censoring using survival forests. More specifically, we investigate the performance and stability of the probability of default (PD) forecasting utilizing three ensemble models of survival analysis: Relative Risk Forests (RRF-TV), Conditional Inference Forests (CIF-TV), and Transformation Forests (TSF-TV). We proceed by comparing these models to the classical extended Cox models in a dataset of 3.626 US mortgages covering quarters starting in Q3 – 2000 up to Q1 – 2015 which includes the period of the Global Financial Crisis (GFC). We verify the improved performance of the ensemble models over the traditional extended Cox models both in terms of forecasting precision and stability. In chapter three we designed and tested models for forecasting the loss given de- fault (LGD) of mortgage loans using a dataset on 5.000 US mortgages provided by the International Finance Database starting in Q3 – 2000 up to Q1 – 2015 which includes the period of the Global Financial Crisis (GFC). This chapter had two objectives. On the first objective, we evaluated the gains in accuracy of a repos- session model using a novel survival forest analysis approach. Our results indicate that all survival forests outperform the prevailing extended COX models in terms of accuracy. Secondly, we propose a "two-stage" model with a survival analysis prob- ability of the repossession component and non-linear haircut component to produce estimates of LGD. Our results demonstrate that our models perform at similar lev- els to the classical "single-stage" with the additional benefit of providing insights into the repossession and haircut factors of mortgage loss as well as a framework for stress testing. For instance, our findings include insights into the connection of macroeconomic and behavioral variables with the LGD estimates.