Bootstrap based nonparametric curve and confidence band estimates for spectral densities

We consider the problem of global bandwidth optimisation and confidence interval estimation for spectral density estimates obtained by applying a nonparametric curve estimator to the periodogram. The use of a local quadratic regression smoother is examined as a possible way to reduce the bias inherent in classical kernel spectral density estimators, which are simply local mean regression smoothers. It is found that while quadratic smoothers are much less sensitive to a poor choice of bandwidth, they do not always outperform mean smoothers.