Bootstrap‐assisted Goodness‐of‐fit Tests in the Frequency Domain (Q4956033)

For a data generating model \(X_t=g_\vartheta(X_{t-1},X_{t-2},\dots,\varepsilon_{t-1}, \varepsilon_{t-2},\dots)+\varepsilon_t\) tests for the hypothesis \(H_0:\vartheta\in\Theta\) are considered based on the spectral distribution function \(F_\vartheta(\lambda)\) of \(X\) and its empirical counterpart \(\hat F_n(\lambda)\) (integrated periodogram). The authors use the normalized empirical process \[ Z_n(\lambda)=(n/2)^{1/2} \left(\hat F_n(\lambda)/\hat F_n(\pi) - F_{\hat\vartheta}(\lambda)/F_{\hat\vartheta}(\pi)\right) \] (\(\hat\vartheta\) being an estimator for \(\vartheta\)) to construct Kolmogorov-Smirnov, Cramér-Von Mises and Anderson-Darling type tests for \(H_0\). Bootstrap procedures are proposed to derive critical values for these tests. The consistency of these procedures is proved. Results of simulations for linear and nonlinear time-series are described including threshold autoregressive, nonlinear moving average and generalized ARCH models.