Predicting bond risk premiums with machine learning: Evidence from China

This study evaluates the ability of machine-learning algorithms to forecast bond risk premiums in the Chinese market. Using a comprehensive set of macro-, firm- and bond-level predictors, we find that machine learning, especially neural network, delivers markedly higher out-of-sample performance than traditional linear benchmarks. The local per-capita fiscal expenditure (EXPEND), bond credit ratings (CREDIT), and profitability- and intangible-related firm characteristics emerge as the most informative variables. Predictive gains are especially pronounced for low-rated issues, non-state-owned enterprises, and periods of heightened economic policy uncertainty. Incorporating machine-learning-based forecasts also helps to enhance credit rating accuracy. Collectively, our findings highlight the value of non-linear machine learning modeling techniques for bond pricing in emerging markets.