Bootstrap based inference for sparse high-dimensional time series models (Q2040070)

In this paper new estimators and hypothesis tests based on bootstrap procedures, applied to sparse (several zero coefficients) high-dimensional time series, are proposed. The first step of the proposed method consists of transforming the model with dispersed coefficients into an equivalent one, but whose coefficients are not dispersed (``de-sparsified estimator of VAR parameters''). Next, based on the limiting distribution of the previous estimator and by means of a bootstrap procedure, various inferences (including hypothesis testing) are made about the VAR\((d)\) model. An example of a hypothesis testing problem in this context may be testing whether a subset of the parameters is zero. In discussing this problem, the authors propose a statistic that avoids the problem of inverting a high-dimensional matrix. Under some assumptions, results on the asymptotic behavior of the proposed estimator are shown. By a simulation study, the performance of the proposed bootstrap method is analyzed for some VAR models. An important part of this study has been implemented using R package routines. Also, the results of the proposed method on an example with real data are displayed. The authors of this paper believe that although there are several real life phenomena that could be well analyzed with a model of series with sparse coefficients, it is an underdeveloped area, possibly due to the difficulties for a rigorous study of its asymptotic theory. Unfortunately the notation in several places in the text is confusing. Also, it is regrettable that some notations and examples of interest to understand this paper are in supplementary material that must be purchased separately from this paper.