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.