Predicting corporate bond returns in the US bond market via machine learning

Abstract

We perform a comparative analysis of two machine learning methods to predict corporate bond return in the US bond market. In contrast to previous studies, we find that the most influential variables are associated with size risk and past return. However, credit and liquidity risks are more prominent when negative externalities impact the market. Further, high predictability at short horizons combined with the investment strategy employed translates into highly significant alphas. We identify the best-performing method to be a decisiontree- based model utilizing boosting. The out-of-sample performance for this method remains statistically significant after accounting for transaction costs.