Comparison of Logistic Regression and Classification Trees to Forecast Short Term Defaults on Repeat Consumer Loans

Naicker, Keeland

Comparison of Logistic Regression and Classification Trees to Forecast Short Term Defaults on Repeat Consumer Loans

Master Thesis

2021

Abstract

This dissertation highlights the performance comparison between two popular contemporary consumer loan credit scoring techniques, namely logistic regression and classification trees. Literature has shown logistic regression to perform better than classification trees in terms of predictiveness and robustness when forecasting consumer loan default events over standard twelve-month outcome periods. One of the major shortcomings with classification trees is its tendency to overfit data eroding its robustness, making it vulnerable to underlying population characteristic shifts. Classification trees remains a popular technique due to its ease of application (algorithm machine learning basis) and model interpretation. Past research has found classification trees to perform marginally better than logistic regression with respect to predictiveness and robustness when modelling short term consumer credit default outcomes related to previously unseen new customer credit loan applications. This dissertation independently tested this finding on reloan consumer loan data, repeat customers who renewed loan facilities at a significant South African micro lender. This dissertation tests the finding if the classification tree technique would outperform logistic regression when modelling this new type of loan data. Credit scoring models were built and tested for each respective technique across identical data sets with the intent to eliminate bias. Robustness tests were constructed via careful iterative data splits. Performance tests measuring predictiveness and robustness were conducted via the weighted sums of squared error evaluation approach. Results reveal logistic regression to outperform classification trees on predictiveness and robustness across the designed uniform iterative data splits, which suggests that logistic regression remains the superior technique when modelling short term credit default outcomes on reloan consumer loan data.

Keywords

Corporate Finance and Valuations

Reference:

Collections

Masters

Full item page