Empirical study on next-day price prediction of Chinese convertible bonds: deep models fail to outperform random walk

Derun Zhong, Guolian Wang, Shiqing Wang, and Hongjie Feng released an empirical paper examining the predictive power of cross-bond panel learning for next-day prices of Chinese convertible bonds. Results show that deep neural network models fail to outperform the random walk benchmark. The study is based on 73,806 daily observations of 100 A-share convertible bonds from 2022 to 2025, comparing the random walk benchmark with cross-bond neural network models on a held-out set of 29,514 observations. The random walk benchmark has an RMSE of 2.3577 yuan, while the five-submodel ensemble RMSEs of small and large price change multilayer perceptrons are 2.3619 yuan and 2.3645 yuan, respectively. Out-of-sample R² is negative, and target-date clustered HAC p-values are 0.453 and 0.248. The paper states that without using contemporaneous stock market information or explicit contractual event conditions, deep panel architectures cannot provide incremental next-day predictability beyond the persistence benchmark.

AI-generated from reports · 2 hr agoUpdated

Not enough consecutive observations yet; trend is not drawn.