Data-driven shrinkage of the spectral density matrix of a high-dimensional time series (Q489160)

A shrinkage estimator for the spectral density matrix of high-dimensional time series is discussed, where the number of components is large in comparison to the number of time points observed. More precisely the smoothed periodogram as nonparametric estimator for the spectral density matrix is shrunk towards a second e.g.\ diagonal matrix, which can also be chosen data-driven. Such a shrinkage estimator balances spectral fit with regularization of estimators, which is of particular importance in high-dimensional settings. Because the optimal shrinkage weight depends on the second moment structure of the shrinkage target and the smoothed periodogram matrix, these quantities need to be estimated. To this end a multivariate extension of the wild TFT-bootstrap (using a normal approximation in the frequency domain) is used as it simultaneously gives a bootstrap sample in the frequency domain (needed to get a bootstrap version of the smoothed spectral density) as well as as a corresponding sample in the time domain (needed to obtain a bootstrap version of the shrinkage target). The validity of the proposed bootstrap procedure in this context is proven. There are three approaches discussed to choose the shrinkage target: First, a multiple of the identity matrix is proposed where the factor is chosen as the mean eigenvalue of the smoothed spectral density matrix. This has been proposed by \textit{H. Böhm} and \textit{R. von Sachs} [J. Multivariate Anal. 100, No. 5, 913--935 (2009; Zbl 1157.62063)] for multivariate time series and improves the condition number of the matrix which may be of particular importance in high-dimensional settings. A second approach uses a parametrically fitted estimator of the spectral density matrix obtained from a reasonable (but misspecified) model class e.g.\ a VAR-model. This was dicussed by \textit{M. Fiecas} and \textit{H. Ombao} [Ann. Appl. Stat. 5, No. 2A, 1102--1125 (2011; Zbl 1232.62147)] but in a context where enough data was observed to get a reasonable fit for the parametric model. In this case the shrinkage can be interpreted as a nonparametric correction of a parametric estimator. In a high-dimensional context, however, a full VAR model has too many parameters to allow for a good parametric fit. For this reason, in this paper a compromise is suggested, where univariate AR models are fitted to each component of the time series and the corresponding diagonal spectral density matrix is used as a shrinkage target. In a simulation study the three shrinkage targets are compared, where it depends on the data as well as the problem at hand which one is preferable. Nevertheless, shrinkage does improve the smoothed spectral density matrix with respect to the mean integrated squared error. Finally, the methods are applied to fMRI data.