Credit Risk Assessment Algorithm Using Deep Neural Networks with Clustering and Merging

Abstract

A reliable assessment model can help financial institutions to increase profits and reduce losses. In credit data, classes of the data are extremely imbalanced owing to the small sample size of bad customers. In this paper, we propose a credit risk assessment algorithm using deep neural networks with clustering and merging, to achieve a balanced dataset and judge whether customer can be granted loans. In the algorithm, the majority class samples are divided into several subgroups by k-means clustering algorithm, each subgroup is merged with the minority class samples to produce several balanced subgroups, and these balanced subgroups are classified using deep neural networks respectively. In the experiments, we analyze influences of the model parameters and data sampling methods on the model performance, and compare classification ability of different models. The experimental results show that the proposed algorithm has a higher prediction accuracy in credit risk assessment.