Gaussian likelihood-based inference for non-invertible MA(1) processes with S\(\alpha \)S noise (Q1805794)

Let \(\{\varepsilon_t\}\) be an iid sequence of symmetric \(\alpha\)-stable random variables with characteristic function \(\exp \{-| t| ^{\alpha}\}\), \(0<\alpha<2\). The authors consider the MA(1) process \(\{Y_t,\;1 \leq t \leq T\}\) defined by \(Y_t = \varepsilon_t - \theta \varepsilon_{t-1}\), where \(| \theta| \leq 1\). Since the aim of the paper is to make inference about \(\theta \) when the true value \(\theta =\theta_0\) is 1 or near to 1, it is assumed that \(\theta_0 = 1-\gamma /T\) for some \(\gamma \geq 0\) and the parameter \(\theta\) is taken in the form \(\theta = 1-\beta/T\), where \(\beta \geq 0\). Let \({\hat{\theta}}_{\text{LM}}\) be the local maximizer of the Gaussian likelihood closest to \(\theta =1\). The authors prove that \(T(1-{\hat{\theta}}_{\text{LM}}) \overset{\text{d}} {} {\tilde{\beta}}_{\gamma}\), where \({\tilde{\beta}}_{\gamma}\) is the minimizer of a stochastic process. The limit distribution has a discrete component at 0 (so called pile-up effect) and a continuous component. It is demonstrated by a simulation study that the limit distribution of \(T(1 - {\hat{\theta}}_{\text{LM}})\) does not depend on \(\alpha\) very much and that the pile-up effect slightly increases with \(\alpha\).