Our client is one of India’s prominent private sector banks with a growing national footprint, offering specialized services across five key business sectors: Corporate Banking, Commercial Banking, Branch & Business Banking, Retail Assets, and Treasury & Financial Markets Operations. The client sought to establish a methodology for assessing expected credit losses. In this case study, we explore the development of a credit risk model using a synthetic dataset of Indian bank customers. The primary goal is to predict whether a customer is likely to default on their loan.
Overview
Solution
Data Sources: The assessment of creditworthiness draws from three main data sources: financial statements, credit bureau data, and alternate data. The latter includes unconventional sources like social media activity and online purchasing behavior, providing additional insights, particularly for individuals or businesses with limited traditional credit history.
Data Source Details: The dataset is divided into training and testing sets, with variables such as employment status, debt-to-income ratio, monthly expenses, and number of dependents. The binary “Default” variable indicates loan default (1) or no default (0).
Data Preprocessing: Essential to model development, data preprocessing involves encoding categorical variables, scaling and normalization of numerical variables, and the division of data into training and testing sets.
Feature Engineering: Feature engineering includes calculating Weight of Evidence (WOE) and Information Value (IV), derived from logistic regression, to aid in variable selection and ranking. WOE measures the predictive power of independent variables, while IV ranks variables based on importance.
Model Selection and Training: Logistic regression is chosen for its simplicity, interpretability, and effectiveness in credit risk assessment, especially in credit scoring and probability of default (PD) modeling.
Model Evaluation and Performance Metrics: Performance metrics such as the Gini Coefficient, Area Under the Curve (AUC), and Kolmogorov-Smirnoff Statistic (KS) are employed. The Gini coefficient measures predictive power, AUC assesses overall model performance, and KS identifies differences in cumulative event and non-event distributions.
Model Validation: Using the testing dataset, the model’s predictions are compared to true outcomes, and performance metrics are calculated to validate its effectiveness.
Output
This project outlines the meticulous process of building a credit risk model using a synthetic dataset. The journey involves data preparation, model selection, feature engineering, model training, and evaluation. Logistic regression emerges as a robust choice for credit risk modelling. Performance metrics, including the Gini coefficient, AUC, and KS statistic, offer valuable insights for lenders in making informed decisions about loan applicants’ creditworthiness.
