Modeling and Forecasting the Yield Curve Under Model Uncertainty

Abstract

We propose a methodology that permits to investigate and forecast the behavior of a variable and its determinants in real time, both in the time and in the frequency domain, starting from a model designed in the time domain, which makes the presentation and evaluation of the results straightforward. This paper applies the methodology to the yield curve. We extract all the shocks affecting the forward rates and the yields and we divide them into three disjoint classes: 1) long-run shocks giving rise to possibly permanent effects, 2) medium-run forces and 3) short-run forces giving rise to transitory effects. These forces drive the low-, medium- and high-frequency component, respectively, composing the time series of the variables used in the model. We explicitly model and estimate such cause-and-effect relationships. The analysis of the shocks and the frequency components provides a timely and comprehensive overview of the nature of the movements in the yields. Furthermore, using the forecast of the frequency components to forecast the yields enhances forecast accuracy, also at long prediction horizons. To perform the frequency decompositions, to identify the forces governing the evolution of the model variables, and to perform the out-of-sample forecasts we use a dynamic filter whose embedded feedback control corrects for model uncertainty.