Deep Dense Convolutional Networks for Repayment Prediction in Peer-to-Peer Lending

Abstract

In peer-to-peer (P2P) lending, it is important to predict default of borrowers because the lenders would suffer financial loss if the borrower fails to pay money. The huge lending transaction data generated online helps to predict repayment of the borrowers, but there are limitations in extracting features based on the complex information. Convolutional neural networks (CNN) can automatically extract useful features from large P2P lending data. However, as deep CNN becomes more complex and deeper, the information about input vanishes and overfitting occurs. In this paper, we propose a deep dense convolutional networks (DenseNet) for default prediction in P2P social lending to automatically extract features and improve the performance. DenseNet ensures the flow of loan information through dense connectivity and automatically extracts discriminative features with convolution and pooling operations. We capture the complex features of lending data and reuse loan information to predict the repayment of the borrower. Experimental results show that the proposed method automatically extracts useful features from Lending Club data, avoids overfitting, and is effective in default prediction. In comparison with deep CNN and other machine learning methods, the proposed method has achieved the highest performance with 79.6%. We demonstrate the usefulness of the proposed method as the 5-fold cross-validation to evaluate the performance.