CMStatistics 2019: Start Registration
View Submission - CFE
A0568
Title: Bandwidth choice in functional cointegration Authors:  Anurag Banerjee - Durham University (United Kingdom)
Xing Wang - Durham University Business School (United Kingdom) [presenting]
Abstract: The choice of the bandwidth in nonparametric predictive regression is studied. Two methods are considered: minimization of the MSE of the predictions, and choice based on the estimation of the convergence rate of the data generating process (DGP). The bin size $k$ in piecewise local linear regression has been previously chosen by information criteria, which is an arbitrary method, while a rule of thumb $k=\sqrt{ln(T)}/T^{3/4}$ to achieve better MSE has also been proposed. We consider the bin size to be $k=T^{\alpha}$. Thus, $k$ can be chosen by estimating $\alpha$ in a similar method to the one that is used to estimate the tail index of a time series. The potential advantage is twofold. First, the method is quite general in the sense that it can be applied to linear/nonlinear, integrated/long memory case. Second, an estimated $\alpha$ provide a clear approach to develop tests on the stationarity of the residuals in functional cointegration.