The cross-section yielding of bonds based on machine learning

Abstract

With the rapid development of China's bond market, credit risk and the systemic financial risk derived from it are increasingly exposed. The market shock brought by the normalization of bond default events reflects the role of bond mispricing in fueling the financial crisis. Scientific prediction of bond yield is of great significance for financial product pricing and financial risk control. Based on the perspective of public fund manager, this paper mainly studies the prediction problem of cross-section excess returns of credit bonds in China. In the aspect of forecasting factors, this paper creatively selects micro influencing factors of credit spread and term spread to forecast the annual excess return of credit bonds. In terms of research model, besides the traditional logistic regression model, this paper also adopts six machine learning algorithms to construct the bond yield prediction model, including three single classifiers (neural network, support vector machine and k-nearest neighbor) and three ensemble algorithms (random forest, XGBoost and Adaboost).The results show that, compared with the traditional logistic regression model, machine learning algorithms significantly improve the out-of-sample prediction ability of bond returns. The prediction performance difference between the ensemble algorithm and the single classifier is mainly reflected in the recall rate. Among them, random forest algorithm is the ensemble algorithm with the best performance and the strongest stability. At the same time, the predictor system constructed in this paper effectively improves the situation that the explanatory power of prediction factors is not strong in previous studies, and shows a good degree of differentiation in the prediction of bond yield.