Survival analysis of bank loans and credit risk prognosis
Date
2015-03-29
Authors
Marimo, Mercy
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Standard survival analysis methods model lifetime data where cohorts are tracked from
the point of origin, until the occurrence of an event. If more than one event occurs, a
special model is chosen to handle competing risks. Moreover, if the events are defined
such that most subjects are not susceptible to the event(s) of interest, standard survival
methods may not be appropriate. This project is an application of survival analysis in a
consumer credit context. The data used in this study was obtained from a major South
African financial institution covering a five year observation period from April 2009 to
March 2014. The aim of the project was to follow up on cohorts from the point where
vehicle finance loans originated to either default or early settlement events and compare
survival and logistic modeling methodologies. As evidenced by the empirical Kaplain
Meier survival curve, the data typically had long term survivors with heavy censoring
as at March 2014. Cause specific Cox regression models were fitted and an adjustment
was made for each model, to accommodate a proportion p of long term survivors. The
corresponding Cumulative Incidence Curves were calculated per model, to determine
probabilities at a fixed horizon of 48 months. Given the complexity of the consumer
credit lifetime data at hand, we investigated how logistic regression methods would
compare. Logistic regression models were fitted per event type. The models were
assessed for goodness of fit. Their ability to differentiate risk were determined using
the model Gini Statistics. Model assessment results were satisfactory. Methodologies
were compared for each event type using Receiver Operating Characteristic curves
and area under the curves. The Results show that survival methods perform better than
logistic regression methods when modelling lifetime data in the presence of competing
risks and long term survivors.
Description
A dissertation submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in fulfilment of requirements for the degree of Master of Science. Johannesburg, 2015.