Prediction a credit score with analysis of debt payment transactions using machine learning algorithms

With the rapid expansion of the financial sector, minimizing financial risks has become increasingly crucial. Among these risks, credit use, the most common form of individual borrowing, takes the centre stage. Credit providers use systems that predict the repayment risk of the applicant to decide whether to grant credit to the applicant or not. In Turkey, the Credit Registry Bureau (KKB) provides the widely used credit score. However, as this score primarily evaluates risks based on past payment behaviour, it can sometimes fall short in offering a comprehensive assessment. This study introduces an alternative credit scoring model using machine learning (ML) and explainable artificial intelligence (XAI) techniques. Unlike previous studies, this thesis does not rely on credit or credit card history for credit score estimation. Instead, it analyses individual bill payment habits using real transaction data from a bill payment institution in Turkey, combined with actual KKB scores. The methodology includes data cleaning, scaling, artificial sampling, classification algorithms, cross-validation, and XAI techniques. The model was validated through multiple evaluation metrics, resulting in a novel approach to credit scoring. The proposed machine learning model holds significant importance, particularly for the application of machine learning models in the banking and finance sector. It is also important for users to understand and interpret the complex models. The model presented in this study offers an alternative scoring method for individuals without a traditional credit score. For those with an existing credit score, it can serve as a complementary tool to enhance the current score or as a means of verification.