Bond Default Prediction with Temporal Graph Convolutional Neural Network and Weakly Supervised Learning

Abstract

Bond defaults prediction is an important task in financial market. Many researchers have also tried to use machine learning methods for bond default prediction in recent years, but the existing methods are threatened by the class imbalance problem, and frequent change of market conditions. In this paper, we propose a framework for bond default prediction based on weakly supervised learning method that could solve the problems. Firstly, the risk feature data of bonds are embedded through a supervised model based on a temporal deep learning method and graph neural network. Secondly, based on the obtained embedding vector, combined with a reasonable estimate of the current bond market, a weakly supervised learning approach is used to perform bond default prediction. The model can make use of both bond-related information and market-related information, which significantly improves the effectiveness. In the experimental results on a real dataset from the Chinese bond market, when predicting defaulted bonds, our study achieved a recall of 0.83, a precision of 0.65, and an F1 score of 0.72, outperforming existing models.