We study machine-learning forecasts and portfolio allocation for Chinese government bonds. Penalised linear regressions and gradient-boosted trees predict next-month total returns for eight maturity buckets. Their forecasts feed mean–variance, Black–Litterman, and risk-budgeting implementations with weight, duration, and trading-cost controls; convexity, tail-risk, and attribution constraints are analysed separately as theoretical extensions. From January 2019 through December 2025, the three optimised cash portfolios earn net excess-return Sharpe ratios of 1.94–2.26 and maximum drawdowns of 0.18–0.25%. Three long-only bond-selection variants attain Sharpe ratios of 1.39–1.70. A futures-hedged arbitrage variant attains 1.08 under an 80% hedge-P&L scaling and basis-cost assumption. A stylised 10-year futures overlay reduces rather than improves cash-portfolio Sharpe ratios.

In this study, we adjust the hierarchical risk parity (HRP) model by introducing hierarchical information on assets to help manage portfolio risk. The adjusted HRP model with hierarchical information considers both correlation and hierarchical information. Compared with other models, the HRP model with hierarchical information has better out-of-sample robustness for simulation data. Moreover, this model achieves better out-of-sample performance using Chinese industry indices data. The results reveal that the adjusted HRP model is an efficient tool to control out-of-sample portfolio risk.