Analyzing the Risk of Loan Default Using Different Coding Methods

Credit scoring is a mere for rating creditworthiness of a debtor. A credit score is based on credit history, which includes information like number of accounts, total levels of debt, repayment history, and other factors. Historically, investment experts evaluated all these deciding factors. However, with availability of loans, the demand increased drastically and manual assessment became almost impossible. Therefore, the task of making informed decisions on issuing loans was passed to machine learning algorithms. Though the algorithm performs scoring on a client, its other important responsibility is to select the most important features that influence the final decision. The goal of the work is to examine how different approaches to data preprocessing, especially varying encoding methods, can affect the interpretation of features with the strongest predictive power. More specifically, predictive power of features is as important as prediction of default in credit scoring. List of keywords: Credit Default, Credit Scoring, Binary Classification, Target Encoding, One-Hot Encoding, Frequency Encoding, Random Forest, Logistic Regression, Extreme Gradient Boosting, Neural Network, SHAP