Credit risk prediction with and without weights of evidence using quantitative learning models

The credit risk assessment process is necessary for maintaining financial stability, cost and time efficiency, model performance accuracy, comparability analysis and future business implications in the commercial banking sector. By accurately predicting credit risk, highly regulated banks can make informed lending decisions and minimize potential financial losses. The purpose of this paper is to assess the power of conventional predictive statistical models with and without transforming the features to gain better insights into customer's creditworthiness. The findings of the predicted performance of the logistics regression model are compared to the performance results of machine learning models for credit risk assessment using commercial banking credit registry data. Each model has its strengths and weaknesses, and where one model lacks, another performs better. The article reveals that simpler credit risk assessment techniques delivered outstanding performance while consuming less processing power and have given insights into the most contributing feature categories. Improving a conventional predictive statistical model using some of the feature transformations reduces the overall model performance, specifically for credit registry data. The logistics regression model outperformed all models with the highest F1, accuracy, Jaccard Index and AUC values, respectively. Financial institutions, specifically banks have questioned whether transformations using Weights of Evidence (WoE) have been significant in quantifying the relationship between categorical independent variables for various types of credit data. This study provides insights when considering the usage of feature transformation for credit risk modelling in commercial banking. The transformation technique is particularly useful in situations where statistical predictive modelling techniques are employed. The results revealed that not only can the logistic regression models perform similarly to the machine learning models but can also outperform them. The best performance is attributed to the simplicity, interpretability, and access to understanding features of individual clients within a portfolio of credit products. The logistic regression model without transformation turned out to perform the best out of the five machine learning models. Considering the business impact, enhancing the logistic regression model by using a WoE transformation did not improve the model's performance for commercial banking data considered. However, the transformation did provide insights regarding each binned categorical independent variable. Therefore, our findings in this article contribute towards assisting banks in managing the impact and interpretability of each binned feature category on the discriminatory power of credit scoring.