Advancements in Credit Score Analytics using Deep Learning and Predictive Modeling Techniques

Abstract

The credit score is an important factor for the institutions while making decisions related to loan granting. Due to high dimensional data structure and constantly changing relationships within the different features, traditional statistical methods are less effective to understand the credit scoring problem. The advent of data mining and machine learning techniques allows easier implementation to discover latent but relevant features in the large and complex datasets inhospitable to traditional statistical techniques and also enable efficient handling of deeper analysis. Among many machine learning techniques, Artificial Neural Networks have been the most promising and with rapid advancements in the building blocks of these networks, many techniques like Convolutional Neural Networks or Deep Reinforcement Learning are being tested with popular use cases in multiple domains. In this chapter, we will perform comparative research on the popular supervised learning algorithms for the credit scoring domain. We also experiment with the deep learning technique called the Contrastive Divergence ANN, Energy Based Neural Network which is trained with approximate Inference methods. We argue that the Energy based model for supervised learning is more relevant to the problem statement, therefore, applied more complicated and dedicated architecture than the other models i.e. Multilayer Perceptron, Hybrid Neuro-Fuzzy model, and also Fuzzy Expert System and discuss their performance versus the complexity of each model. Finally, we briefly touch upon the credit policy decisions that were made based on the models’ commercial applications, then elaborate on the conclusion of the research findings. We also suggest probable new domains of research related to the combination of energy-based learning and other popular self-supervised learning methods.