SURVIVAL ANALYSIS AS A TOOL FOR BETTER PROBABILITY OF DEFAULT PREDICTION

Michal Rychnovský

doi:10.18267/j.aop.594

Abstract

This paper focuses on using survival analysis models in the area of credit risk and on the modelling of the probability of default (i.e. a situation where the debtor is unwilling or unable to repay the loan in time) in particular. Most of the relevant scholarly literature argues that the survival models produce similar results to the commonly used logistic regression models for the development or testing of samples. However, this paper challenges the standard performance criteria measuring precision and performs a comparison using a new prediction-based method. This method gives more weight to the predictive power of the models measured on an ex-ante validation sample rather than the standard precision of the random testing sample. This new scheme shows that the predictive power of the survival model outperforms the logistic regression model in terms of Gini and lift coefficients. This finding opens up the prospect for the survival models to be further studied and considered as relevant alternatives in financial modelling.

Highlights

This paper deals with the field of consumer credit underwriting and the mathematical models that are used
I concluded that in this sample both models have a similar performance on the random training and testing sample, which is in line with the existing research of Stepanova and Thomas (2002), Cao et al (2009), Bellotti and Crook (2009) and Tong et al (2012) and the regional specific Czech fix-term unsecured loan banking data make no exception
When compared with the new performance criteria measuring the predictive power of the model, the Cox model notably outperforms the logistic regression model

Summary

Introduction

This paper deals with the field of consumer credit underwriting and the mathematical models that are used. The loan approval process is usually very complicated, it could be said that the main parts of the process could be the evaluation of client’s ability to repay the loan and the verification of income and other information provided. The riskiness of the client is typically established based on the estimation of the probability of default (PD) conditional to the client’s characteristics. The probability of default is usually estimated using the logistic regression models. The regression model, called the scoring model, assigns a score to each client, which is used as a key factor for automated approval or rejection of the loan application in the process or as one of the main inputs for the subsequent manual underwriting

Objectives

Results

Conclusion