Advanced survival modelling for consumer credit risk assessment: addressing recurrent events, multiple outcomes and frailty

A thesis submitted in partial fulfillment of the requirements for the degree of Doctor in Information Management, specialization in Statistics and EconometricsThis thesis worked on the application of advanced survival models in consumer credit risk assessment, particularly to address issues of recurrent delinquency (or default) and recovery (cure) events as well as multiple risk events and frailty. Each chapter (2 to 5) addressed a separate problem and several key conclusions were reached. Chapter 2 addressed the neglected area of modelling recovery from delinquency to normal performance on retail consumer loans taking into account the recurrent nature of delinquency and also including time-dependent macroeconomic variables. Using data from a lending company in Zimbabwe, we provided a comprehensive analysis of the recovery patterns using the extended Cox model. The findings vividly showed that behavioural variables were the most important in understanding recovery patterns of obligors. This confirms and underscores the importance of using behavioural models to understand the recovery patterns of obligors in order to prevent credit loss. The findings also strongly revealed that the falling real gross domestic product, representing a deteriorating economic situation significantly explained the diminishing rate of recovery from delinquency to normal performance among consumers. The study pointed to the urgent need for policy measures aimed at promoting economic growth for the stabilisation of consumer welfare and the financial system at large.Chapter 3 extends the work in chapter 2 and notes that, even though multiple failure-time data are ubiquitous in finance and economics especially in the credit risk domain, it is unfortunate that naive statistical techniques which ignore the subsequent events are commonly used to analyse such data. Applying standard statistical methods without addressing the recurrence of the events produces biased and inefficient estimates, thus offering erroneous predictions. We explore various ways of modelling and forecasting recurrent delinquency and recovery events on consumer loans. Using consumer loans data from a severely distressed economic environment, we illustrate and empirically compare extended Cox models for ordered recurrent recovery events. We highlight that accounting for multiple events proffers detailed information, thus providing a nuanced understanding of the recovery prognosis of delinquents. For ordered indistinguishable recurrent recovery events, we recommend using the Andersen and Gill (1982) model since it fits these assumptions and performs well on predicting recovery.Chapter 4 extends chapters 2 and 3 and highlight that rigorous credit risk analysis is not only of significance to lenders and banks but is also of paramount importance for sound regulatory and economic policy making. Increasing loan impairment or delinquency, defaults and mortgage foreclosures signals a sick economy and generates considerable financial stability concerns. For lenders and banks, the accurate estimation of credit risk parameters remains essential for pricing, profit testing, capital provisioning as well as for managing delinquents. Traditional credit scoring models such as the logit regression only provide estimates of the lifetime probability of default for a loan but cannot identify the existence of cures and or other movements. These methods lack the ability to characterise the progression of borrowers over time and cannot utilise all the available data to understand the recurrence of risk events and possible occurrence of multiple loan outcomes. In this paper, we propose a system-wide multi-state framework to jointly model state occupations and the transitions between normal performance (current), delinquency, prepayment, repurchase, short sale and foreclosure on mortgage loans. The probability of loans transitioning to and from the various states is estimated in a discrete-time multi-state Markov model with seven allowable states and sixteen possible transitions. Additionally, we investigate the relationship between the probability of loans transitioning to and from various loan outcomes and loan-level covariates. We empirically test the performance of the model using the US single-family mortgage loans originated during the first quarter of 2009 and were followed on their monthly repayment performance until the third quarter of 2016. Our results show that the main factors affecting the transition into various loan outcomes are affordability as measured by debt-to-income ratio, equity as marked by loan-to-value ratio, interest rates and the property type. In chapter 5, we note that there has been increasing availability of consumer credit in Zimbabwe, yet the credit information sharing systems are not as advanced. Using frailty survival models on credit bureau data from Zimbabwe, the study investigates the possible underestimation of credit losses under the assumption of independence of default event times. The study found that adding a frailty term significantly improved the models, thus indicating the presence of unobserved heterogeneity. The major policy recommendation is for the regulator to institute appropriate policy frameworks to allow robust and complete credit information sharing and reporting as doing so will significantly improve the functioning of the credit market