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

01 ABSTRACT

Using a daily panel of 73,806 observations across 100 A-share convertible bonds from 2022 to 2025, the paper evaluates cross-bond neural architectures for next-day price forecasting under strict temporal splitting, boundary purging, fold-specific preprocessing, and dependence-aware inference. The facts are: on the 29,514-observation holdout, the random walk achieves an RMSE of 2.3577 CNY, while 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 and non-significant HAC p-values, and block-bootstrap intervals spanning zero. The authors' view is that without contemporaneous equity market information or explicit contractual event conditioning, deep panel architectures do not deliver incremental next-day predictability over persistence.

02 KEY FINDINGS

1. The random walk achieves an RMSE of 2.3577 CNY on the holdout, while 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).
2. Target-date clustered HAC p-values are 0.453 and 0.248, and ten-day circular block-bootstrap intervals span zero, so the null of no advantage is not rejected.
3. Cross-bond pooling lowers test RMSE for 79 of 90 individual bonds, showing a variance-stabilizing shrinkage effect, but fails to produce an aggregate forecast advantage; five highly volatile, clause-sensitive bonds account for 74.4% of the model's total excess squared error.
4. Direct price-level training degrades performance substantially.