Quasi-likelihood estimation of non-invertible moving average process (Q2711684)

A moving average model with i.i.d. disturbances having PDF \(f\) is considered. For the case when the characteristic polynomial has roots inside the unit circle (non-invertibility) a (quasi) maximum-likelihood estimator of the MA coefficients is constructed. It is shown that the consistency of this estimator is intimately connected with the PDF \(f\). E.g., let \(g\) be a zero mean Gaussian density such that \(\int |x|g(x)dx=\int |x|f(x)d x\).NEWLINENEWLINENEWLINEThen the MLE estimator is consistent if for some \(0<a_1<a_2\), \(f(x)>g(x)\) for \(|x|<a_1\) and \(|x|>a_2\), and \(f(x)<g(x)\) for \(a_1<|x|<a_2\).NEWLINENEWLINENEWLINEIf \(f( x)<g(x)\) for \(|x|<a_1\) and \(|x|>a_2\) and \(f(x)>g(x)\) for \(a_1<|x|<a_2\), the estimator is inconsistent.NEWLINENEWLINENEWLINEThis result is in contrast to results on standard invertible moving average models where the quasi likelihood typically produces consistent estimates.