中文摘要：

本文运用基于独立Minnesota-Wishart共轭先验分布的贝叶斯向量自回归模型（BVAR）,并通过Gibbs抽样的马尔科夫链蒙特卡洛模拟方法预测中国银行间国债的收益率。此外,按照固定窗口的滚动预测规则,采用统计性损失函数和经济准则（夏普比率和资产组合效用损失）共同作为评判标准,比较BVAR模型与其他8个常见模型在直接和递归方式上的预测效果。结果表明BVAR模型的短期预测效果不稳定,但中长期直接预测效果显著好于递归预测及其他模型,并且预测步长及收益率的期限越长,预测精度越高,反映了BVAR模型预测中长期国债收益率的优越性。

英文摘要：

Based on independent Minnesota-Wishart conjugate prior Bayesian vector autoregressive model （BVAR）, this paper forecasts China＇s inter-bank bond yield by MCMC of Gibbs sampling. Except for statistical measures of forecast accuracy, we also consider alternative measures according to economic criteria （ Sharpe Ratio ＆ portfolio losses） , and based on fixed rolling forecast window, we compare the direct and recursive forecast accuracy of BVAR with other 8 common models. The result shows that BVAR＇ s short-run prediction may be unstable, while its medium-and long-term direct prediction accuracy is better than that of recursive and other models, and the longer the horizon and the yield maturity the higher prediction precision is, which reflects the BVAR＇ s the significant superiority in the forecast of medium-and long-term yields.