scientific article; zbMATH DE number 1337245

Publication date 15 September 1999

autoregression hypothesis testing maximum likelihood probability of large deviations

Parametric hypothesis testing (62F03) Time series, auto-correlation, regression, etc. in statistics (GARCH) (62M10)

For the AR(1) process \(\xi_i=\theta\xi_{i-1}+w_i\), \(i=1,\dots,n\), where \(w_i\sim{\mathcal N}(0,1)\), the simple hypothesis \(H^n: \theta=\theta_0\) is tested against the simple alternative \(\widetilde H^n: \theta=\widetilde\theta\not=\theta_0\). The Neyman-Pearson test \( \delta_n=I(\Lambda_n>d_n)+q_n I(\Lambda_n=d_n) \) is constructed, where \(\alpha_n\in(0,1)\), and \( \Lambda_n = (\widetilde\theta -\widetilde\theta_0)\sum_{i=1}^n\xi_{i-1}w_i -{1\over 2}(\widetilde\theta-\theta_0)^2\sum_{i=1}^n\xi_{i-1}^2 \) is the log likelihood ratio for \(H^n\) against \(\widetilde H^n\), \(d_n\) and \(q_n\) are parameters chosen from the condition \(E_{\theta_0}\delta_n=\alpha_n\). The probability of the second type error is \(\beta_n=E_{\widetilde\theta}(1-\delta_n)\). Let \(b(\varepsilon)=-\theta_0-\varepsilon(\widetilde\theta-\theta_0)\), \(c(\varepsilon)=1+\theta_0^2+\varepsilon(\widetilde\theta^2-\theta^2_0)\), \( \kappa(\varepsilon)=(1/2)\ln(0.5(c(\varepsilon)+ (c^2(\varepsilon)-4b^2(\varepsilon))^{1/2})) \), \(\gamma_0=\kappa'(0)\), and \(\varepsilon(\gamma)\) is the unique root of the equation \(\kappa'(\varepsilon)=\gamma\), \(I(\gamma)=\gamma\varepsilon(\gamma)- \kappa(\varepsilon(\gamma))\). The following results are proved. If \( \theta_0 <1\), then \[ \lim_{n\to\infty}n^{-1}\ln \Pr{}_{\theta_0}(n^{-1}\Lambda_n>\gamma)= \lim_{n\to\infty}n^{-1}\ln \Pr{}_{\theta_0}(n^{-1}\Lambda_n\geq\gamma)=-I(\gamma) \] for all \(\gamma\in(\gamma_0,\gamma_{+})\;\) (if \( \widetilde\theta <1\), then \(\gamma_{+}=+\infty\)); \[ \lim_{n\to\infty}n^{-1}\ln \Pr{}_{\theta_0}(n^{-1}\Lambda_n<\gamma)= \lim_{n\to\infty}n^{-1}\ln \Pr{}_{\theta_0}(n^{-1}\Lambda_n\leq\gamma)=-I(\gamma) \] for all \(\gamma\in(-\infty,\gamma_0)\). If \( \theta_0 <1\), \( \widetilde\theta <1\), then \( \lim_{n\to\infty}n^{-1}\ln \Pr{}_{\widetilde\theta}(n^{-1}\Lambda_n\leq\gamma)=-I(\gamma)+\gamma \) for any \(\gamma<\kappa'(1)\). With the help of these results the authors describe the relation between \(\lim_{n\to\infty}n^{-1}\ln\alpha_n\) and \(\lim_{n\to\infty}n^{-1}\ln\beta_n\).

Cited in

(8)

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 Q4258919)