Spectral Density Based Goodness‐of‐Fit Tests for Time Series Models (Q4956071)

Let a stationary Gaussian process \(X_t\), \(t\in Z\), with spectral density \(f\) be observed at \(t=1,2,\dots,T\). A new test is proposed for testing the hypothesis \(H_0\): \(f(\lambda)=f(\lambda,\vartheta)\), where \(\vartheta\in\Theta\) is an unknown parameter. The test is based on the following property of the periodogram \(I_T(\lambda)\) of \(X_T\): \[ E(I_T(\lambda_j)/f_0(\lambda_j))=f(\lambda_j)/f_0(\lambda_j)+ O(T^{-1}\log(T)), \] where \(\lambda_j=2\pi j/T\). The test statistic \(S_{T,h}\) uses a kernel smoothed version of the relation: \[ S_{T,h}(\hat\vartheta_T)=Th^{1/2}\int_{-\pi}^{\pi} \{(Th)^{-1}\sum_{j=-N}^N K((\lambda-\lambda_j)/h))[I_T(\lambda_j)/f(\lambda_j,\hat\vartheta_T)-1]\}^2d\lambda, \] where \(K\) is a kernel, \(h\) is a bandwidth, \(\hat\vartheta_T\) is the Whittle estimator for \(\vartheta\) (some results holds for any \(\sqrt{T}\)-consistent efficient estimator, e.g. for MLEs). It is shown that under \(H_0\) \[ S_{T,h}(\hat\vartheta_T)-\mu_h\Rightarrow N(0,\tau^2), \] \[ \text{where}\quad \mu_h=h^{-1/2}\int_{-\pi}^{\pi}K^2(x)dx,\quad \text{and}\quad \tau^2=\pi^{-1}\int_{-2\pi}^{2\pi} [ \int_{-\pi}^{\pi} K(u)K(u+x)du]^2 dx. \] If \(H_0\) does not hold, then \[ T^{-1}h^{-1/2}S_{T,h}(\hat\vartheta_T)\to \int_{-\pi}^{\pi} \left(f(\lambda)/f(\lambda,\vartheta^*)-1\right)^2d\lambda, \] where \(\vartheta^*\) is the parameter of the density \(f(\lambda,\vartheta^*)\) nearest to \(f(\lambda)\) in the Whittle sense. The asymptotic confidence level and power under a set of local alternatives are considered for the test based on \(S_{T,h}\). The author notes that the convergence to the limit distributions is very slow, so a bootstrap procedure is considered. Results of simulations and a real data example are presented.