A novel hybrid ensemble model based on tree-based method and deep learning method for default prediction

• A novel hybrid ensemble model for default prediction is proposed. • LightGBM is used to build new feature interactions to enhance feature expression. • CNN is used to build new feature interactions to reflect deeper information. • Ensemble model combining deep learning and tree-based classifiers are used. • The proposed model outperforms comparative methods in four evaluation metrics. Default prediction plays an important role in emerging financial market, so it has attracted extensive attention from financial industry and academic community. A slight improvement in default prediction performance can avoid huge economic losses. Many existing studies have used feature selection to improve the performance of default prediction models but paid limited attention to feature generation. Additionally, deep learning methods have been gradually explored for classification problems. In this study, a novel hybrid ensemble model is proposed to improve the performance of default prediction. First, a tree-based method (i.e., LightGBM) is used to learn new feature interactions and enhance the representation of original features. Second, a deep learning method (i.e., Convolutional Neural Network) is used as feature generation method to generate deeper feature interactions. Moreover, the structure of Inner Product-based Neural Network (IPNN) is used as deep learning classifier to learn feature interactions and reach a good trade-off between predictive accuracy and complexity. Third, ensemble learning method is used to combine the deep learning classifier with tree-based classifiers to obtain superior predictive results. Finally, two default datasets and four evaluation metrics are used to measure the predictive performance. The experimental results show that each component of the proposed model has significant improvement on overall performance.