Convertible bond return predictability with machine learning

We employ 13 machine learning algorithms and construct 56 convertible bond predictors to predict the cross-sectional returns of Chinese convertible bonds, a financial instrument that combines debt and equity characteristics. The out-of-sample tests reveal that neural networks (e.g., NN1) notably outperform other models. All methods identify the same set of dominant predictors, primarily including convertible bond returns, prices, conversion premium, and yield to maturity. More importantly, the machine learning portfolio performance remains economically significant even after accounting for transaction costs. Machine learning also enhances out-of-sample trading performance compared to traditional valuation models.