Cross-Bond Panel Learning for Next-Day Price Forecasting of Chinese Convertible Bonds: A Reality Check

Abstract

Recent advances in financial machine learning have prompted growing interest in applying deep neural networks to convertible bond valuation and forecasting. However, existing studies often focus on in-sample valuation fit or cross-sectional return rankings, leaving out-of-sample next-day price forecasting insufficiently benchmarked against persistence. This paper provides a comprehensive empirical reality check on cross-bond panel learning for Chinese convertible bonds. Using a daily panel of 73,806 observations across 100 A-share convertible bonds from 2022 to 2025, we evaluate whether cross-bond neural architectures reliably outperform the random walk under strict temporal splitting, boundary purging, fold-specific preprocessing, and dependence-aware inference. On the 29,514-observation holdout, the random walk achieves an RMSE of 2.3577 CNY. Five-seed ensembles of small and large price-change MLPs yield RMSEs of 2.3619 and 2.3645 CNY, with negative out-of-sample R^2 values (-0.0035 and -0.0057) and target-date clustered HAC p-values of 0.453 and 0.248. Ten-day circular block-bootstrap intervals span zero, while direct price-level training degrades performance substantially. Although cross-bond pooling provides a variance-stabilizing shrinkage effect over single-bond estimation (lowering test RMSE for 79 of 90 individual bonds), it fails to produce an aggregate forecast advantage over persistence. Error decomposition reveals that failure is not uniform: five highly volatile, clause-sensitive bonds account for 74.4% of the model's total excess squared error. Without contemporaneous equity market information or explicit contractual event conditioning, deep panel architectures do not deliver incremental next-day predictability over persistence.