Cascade of Deep Neural Network And Support Vector Machine for Credit Risk Prediction

Abstract

One of the core financial services that banks render to their customers is granting of loans with interest over a period. To minimize the risk of loan default which eventually may lead to bad debt; the banks use statistical models to determine the customer loan eligibility. There is a transition from the statistical models for predicting eligibility for bank loans to the use of machine learning models and several pieces of research have been carried out in this direction, but the accuracy is still a challenge. In our research work, we adopted a cascade of a pre-trained Deep Neural Network (DNN) and a Support Vector Machine (SVM) to realize a loan eligibility model. An 11-layer DNN with a sigmoid output layer was trained with a loan credit dataset obtained from Kaggle and the output layer was removed which then makes SoftMax with 64 outputs a new output layer. The DNN is then used to transform the original 11-feature dataset into a 64-feature high dimension dataset. An SVM with a polynomial kernel was trained on the original dataset and achieved an accuracy of 87% but the same SVM achieved an accuracy of 97.05% when trained with the transformed high dimension dataset obtained from the pre-trained DNN. In our study, our proposed prediction model has the best performance with regards to related reviewed works having accuracy of 97%. Our proposed prediction model has the best performance with regards to related reviewed works and it can be concluded that our machine learning mix-strategy is effective and can be adapted for a similar task.