A review of imbalanced data techniques with application to loan default

Abstract

Customer loan default occurs when a borrower does not honour the loan repayment programme agreed with the bank. In accounting standards terms and banking regulation, a loan is deemed to be in default when repayment of capital and/or interest fall in arrears by 90 days or more. In the case of personal loans, which is the basis of this thesis, a borrower can default due to loss of income caused by, for example, redundancy and loss of income generating assets. The lenders’ mitigation of the risk of loss depends on its ability to adequately evaluate credit risk. The lender’s objective is to increase revenue by holding a loan portfolio containing predominantly well performing loans at an interest rate reflecting the relative risk, while at the same time minimising losses resulting from defaulted loans. Tools which can be used for risk mitigation purposes are tree-based methods. However, one of the disadvantages of these methods is that of imbalanced data. The focus of this dissertation is the application of techniques to overcome this issue and thus generate accurate predictions from these tree-based methods for classification. The theory of classification trees will also be discussed where decision trees will be discussed as the foundation to understand bagged trees. Bagged trees will be applied on a real-life dataset after applying various techniques used to remove class imbalance within a dataset, namely SMOTE, Borderline SMOTE, ADASYN, Safe-Level SMOTE and SMOTE-NC and the relative results will be compared. Logistic regression will also be applied as it is a benchmark of statistical models for classification.