Title: Data-adaptive estimation of time-varying spectral densities

Abstract: This paper introduces a data-adaptive approach for spectral density estimation of non-stationary processes. Estimation of time-dependent spectra commonly proceeds by means of local kernel smoothing. The performance of these non-parametric estimators depends however crucially on the smoothing bandwidths that need to be specified in both time and frequency direction. The objective of this paper is to construct local spectral density estimates where the respective smoothing kernels are iteratively adapted to the data at hand. The main idea, inspired by the concept of propagation-separation developed in Polzehl and Spokoiny (2006), is to describe the largest local vicinity of every design point in the time-frequency plane over which smoothing is justified by the data. Our method circumvents the problem of optimal bandwidth selection in the strict sense without imposing additional assumptions. The procedure permits full flexibility for the degree of smoothing and automatically adjusts for structural breaks in the time-dependent spectrum.