scientific article; zbMATH DE number 1234103

From MaRDI portal
Publication:4221105





This is the thesis of the author for a Doctor degree at the University of Heidelberg under the supervision of R. Dahlhaus. The topic of this thesis is the applications of bootstrap to the inference of statistical models related to a linear real-valued stationary process. This is an interesting topic since the asymptotic distributions of most statistics in this setting are difficult to evaluate analytically and the naive bootstrap method is not consistent when applied to time series data. In the first chapter, several of bootstrap resampling methods for dependent data, which include the periodogram bootstrap, bootstrap of Fourier coefficients, wild bootstrap, and local bootstrap in the frequency domain, are reviewed. Some asymptotic results on the distributions of the functionals of the periodogram and on the convergence of their bootstrap estimates are then summarized. There is also a theorem which discusses the convergence of bootstrapped empirical spectral processes. The applications of bootstrap to some goodness-of-fit tests for a stationary linear process are presented in Chapter 2. The author first considers the testing of whether the spectral density of the process is from an unknown parametric family. The periodogram and wild bootstrap methods are used to approximate the distributions of a Kolmogrov type test [\textit{R. Dahlhaus}, Stochastic Processes Appl. 30, No. 1, 69-83 (1988; Zbl 0655.60033)] and the goodness-of-fit test based on Bartlett's Up-statistic [\textit{M. S. Bartlett}, Publ. Inst. Statist. Univ. Paris 3, 119-134 (1954; Zbl 0058.35704)], respectively. A new bootstrap method is then proposed to test whether the spectral densities of two independent processes are the same or whether their ratio is a constant. The convergence of bootstrap approximations is proved under some assumptions and the small sample properties of the proposed bootstrap tests are investigated by simulations. A data set which consists of annual indices of European wheat prices from 1500 to 1869 is analyzed using the bootstrap methods proposed. Chapter 3 deals with model selection for a linear stationary process. The nonparametric periodogram bootstrap method is used to approximate the distribution of model selection criteria based on estimates of the Kullback-Leibler information divergence measure. A new model selection criterion, which is different from ATC criterion, is derived based on this approximation. Simulations indicate this criterion works even for families that do not contain the true process. This method has been also applied to determine the model for the well-known Wolf sunspot numbers. In Chapter 4, the author applies the bootstrap to approximate the distribution of a nonparametric regression function estimated from observations with dependent errors. It is shown that the bootstrap approximation yields asymptotic correct confidence sets for the unknown regression function. A bootstrap procedure is also suggested to select the bandwidth for the regression estimate. Finally, the generalization of the periodogram and wild bootstrap methods to multivariate linear processes is briefly considered in Chapter 5. The author points out, however, there are several limitations with the generalizations proposed in this chapter: the applications of the periodogram bootstrap are limited and the practical implementation of wild bootstrap is difficult. Therefore, there is still a need for further research in this area.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4221105)