Download
Abstract
This paper investigates practical issues associated with the use of the local estimator in forecasting models subject to parameter instability. We propose a bandwidth selection procedure for out-of-sample forecasting, derived by minimizing the conditional expected end-of-sample loss, and show that it is asymptotically optimal. We further discuss the implications on the choice of kernel functions and derive the optimal kernel. Theoretical properties are assessed through an extensive Monte Carlo study and three empirical applications: bond return predictability, yield curve forecasting, and real-time inflation forecasting, which demonstrate the superior performance of the local estimator with the proposed optimal bandwidth selection.