Title: Deep time-series feature extraction in credit scoring models
Authors: Huei-Wen Teng - National Chiao Tung University (Taiwan)
Jui-Yu Lin - National Chiao Tung University (Taiwan) [presenting]
Yu-Huai Yu - National Chiao Tung University (Taiwan)
Kai-Shiang Fan - National Chiao Tung University (Taiwan)
Yi-Chia Lin - Chinatrust Financial Holdings (Taiwan)
Abstract: Credit scoring models predict the default of a credit card or loan holder and are of considerable importance in the banking system. We investigate the use of time-series extracted feature using deep learning neural network for predicting default in credit scoring models. With borrowers' payment history records, we extract time-series patterns and predict default risk by using Convolutional Neural Networks and Recurrent Neural Networks. We compare a set of machine learning methods with and without our deep extracted time-series features using two data sets, one is from an open source and the other is form a major bank in Taiwan. Our numerical results show that the performance can be improved with time-series extracted feature in terms of AUC.