Modeling and Forecasting Serially Dependent Yield Curves
利用收益率曲线的序列依赖性，提出半参数模型进行估计和预测，基于1985-2023年美国国债数据发现动态结构可简化为三维向量过程，且预测效果优于其他三因子模型。
Abstract Yield curves are serially dependent. By leveraging this feature, this article proposes a semiparametric model to estimate and forecast yield curves based on factors driving the serial dependence. In this model, factor loadings are related to the autocovariance functions of the continuous and smooth yield curve subject to unobservable errors; the dynamic evolution is driven by a vector autoregression for a small set of factors, and the yield data determine the number of factors and aggregation of information over different lags. Applying this method to monthly U.S. government bond yields from January 1985 through December 2023, I find that the dynamic structure of yield curves reduces to a vector process lying in a 3-dimensional space, with 1-month lag information. Yield curve residuals from this new model exhibit less autocorrelation than alternative three-factor models. Moreover, this new model provides favorable forecasting results.