A Two-Stage Plug-In Bandwidth Selection and Its Implementation in Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation

The performance of a kernel HAC estimator depends on the accuracy of the estimation of the normalized curvature, an unknown quantity in the optimal bandwidth represented as the spectral density and its derivative. This paper proposes to estimate it with a general class of kernels. The AMSE of the kernel estimator and the AMSE-optimal bandwidth are derived. It is shown that the optimal bandwidth for the kernel estimator should grow at a much slower rate than the one for the HAC estimator with the same kernel. A solve-the-equation implementation method is also proposed. Finite sample performances are assessed through simulations.