A novel hybrid method for credit risk prediction based on modified Kohonen SOM and cox proportional hazards model

The banking sector is one of the many economic areas where recent research has shown the potential applicability of deep learning techniques. Repayment of loans by borrowers is one of the most significant issues that banks handle, and it is linked to issues with credit scores. Predictions of default are essential for influencing banks’ approaches to risk management and improving overall financial stability. Through precise estimation of the probability of default (PD), financial institutions can execute focused strategies aimed at reducing lending-related risks. Kohonen Self-Organizing Maps are a type of unsupervised learning method that has been used in numerous studies to provide a dynamic perspective on the probability of default. This method depends on grid clustering, and the decrease in dimensionality makes it simple to spot patterns in the data. The weight vectors of Kohonen Self-Organizing Maps are determined based on data and require a long time. Feature subset selection is an important issue in credit scoring to create models with a small number of characteristics that improve the performance of the classifier. Nevertheless, the large complexity and class imbalance of real credit datasets offer significant challenges to the development of accurate and successful credit scoring algorithms. Kohonen SOM is effective in clustering high-dimensional data and identifying patterns in borrower behavior. It is particularly useful in visualizing complex datasets. This study aims to build a hybrid method based on modifications of the self-organizing map and the Cox proportional hazards model. The hybrid method integrates self-organizing maps for unsupervised clustering and Cox proportional hazards for survival analysis, enhancing interpretability and adaptability. This approach addresses limitations of existing techniques by examining data structure and survival outcomes simultaneously, providing a more nuanced understanding of credit risk. The modified Kohonen method, which assigns density-based weight initialization, significantly enhances the model’s accuracy, precision, and recall, enabling better clustering and classification outcomes by better capturing the data’s underlying structure. The suggested method was applied to a real dataset based on a Gulf Commercial Bank in Iraq for three years, from 2017 to 2019. The proposed method showed a minimum classification time of 60.54 seconds for all features and 45.43 seconds for Cox feature selection, respectively. Banks are enhancing decision-making by reducing classification times, enabling quicker identification of high-risk borrowers, enhancing resource allocation, and improving customer satisfaction.