Cross validation for locally stationary processes (Q2313282): Difference between revisions

The authors deal with the adaptive estimation of a multidimensional parameter curve \(\theta_0:[0,1]\to\Theta\subset \mathbb R^p\), that is, they deal with locally stationary processes \(X_{t,n}, t=1,2,\dots,n\) parameterized by curves. Following the idea of locally stationary processes (see [the second author, ``Locally stationary processes'', in: Time series analysis: methods and applications. Amsterdam: Elsevier. 351--413 (2012; \url{doi:10.1016/b978-0-444-53858-1.00013-2})], for a fixed \(u\in [0, 1]\), \(X_{t,n}\) should locally behave like a stationary process \(\hat{X}_t(u)\). It is assumed that the time dependence of the approximation \(\hat{X}_t(u)\) is solely described by \(\theta_0\), that is, \(\hat{X}_t(u)=\tilde{X}_t(\theta_0(u))\), where \(\tilde{X}_t(\theta),\theta\in\Theta\), is a family of parametric stationary processes. As an estimator of \(\theta_0(u)\) the authors consider local likelihood (or local M-) estimators weighted by kernels, that is, \[\hat{\theta}_h(u) := \arg\min_{\theta\in\Theta}\,L_{n,h}(u,\theta),\] where \[L_{n,h}(u,\theta):=\frac{1}{n}\sum_{t=1}^{n}K_h(\frac{t}{n}-u)\ell_{t,n}(\theta), \] \(\ell\) is a given objective function, the kernel \(K:\mathbb{R}\to\mathbb{R}\) is nonnegative with \(\int K=1\), and \(h\in (0,\infty)\) is the bandwidth. They introduce a Kullback-Leibler divergence between models with parameter curves \(\hat{\theta}_h(u)\) and \(\theta_0(u)\) which can be seen as an analogue to the averaged squared error in nonparametric regression. The asymptotic optimality of the cross-validation procedure with respect to the Kullback-Leibler-type distance measure and convergence of the cross-validation bandwidth toward the deterministic asymptotic optimal bandwidth is proved. The limit distribution of the bandwidth chosen by cross validation is derived. Some processes where the main results are applicable are examined. The performance of the method for different models such as tvAR, tvARCH and tvMA is studied in simulations. A simulation study shows that the method works fairly well also in misspecified situations. For more details see [the second author et al., Bernoulli 25, No. 2, 1013--1044 (2019; Zbl 1427.60057)].