Inverse Gaussian Autoregressive Models

From MaRDI portal





A first-order stationary AR process \(X_t=\rho X_{t-1}+\varepsilon_t\) is considered, where \(\varepsilon_i\) are i.i.d. inverse Gaussian r.v.'s with p.d.f. \[ (1)\qquad f(x,\lambda')=\left(\lambda'/2\pi x^3\right)^{1/2} exp\left(-\lambda'/2x\right),\quad\lambda'=\lambda(1-\sqrt{\rho})^2,\;x>0. \] In this case the p.d.f of \(X_t\) is also of the form (1) with \(\lambda'=\lambda\). Note, that this distribution has heavy tails: \(E(X_t)^\alpha=\infty\), \(\forall\alpha>1/2\). A maximum likelihood estimator for \(\beta=1/\lambda\) is evaluated in the case of known \(\rho\). If \(\rho\) is unknown the authors propose to use the \textit{P.D. Feigin} and \textit{S.I. Resnik} [Commun. Stat., Stochastic Models 8, No. 3, 479-498 (1992; Zbl 0762.62024)] estimator \(\hat\rho_n=\min_{1\leq i\leq n}(X_i/X_{i-1})\). Another estimator for \(\lambda\) is based on the empirical Laplace transform of the data. Consistency and asymptotic normality of the estimators are discussed. Simulation results are presented.











This page was built for publication: Inverse Gaussian Autoregressive Models

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4956032)