A0586
Title: Weighted kernel estimators for forecasting under breaks
Authors: Sze Him Isaac Leung - The Chinese University of Hong Kong (Hong Kong) [presenting]
Abstract: Predicting future observations in non-parametric regression models that are subject to a structural break at an unknown time point is studied. A weighted kernel estimator is proposed to estimate the post-break regression function and forecast future observations, where the weights are location- and time-dependent. It is shown that putting some weights to pre-break observations can improve the estimate of the post-break regression function in terms of the mean squared forecast error (MSFE). This is related to the bias-variance trade-off induced by including pre-break observations. The MSFE optimal weights and bandwidth are estimated in a two-step approach, and the properties are examined. Simulation studies demonstrate that the proposed weighted kernel estimator has a smaller MSFE than post-break methods under various non-parametric regression functions.
