Likelihood Analysis of the <i>I</i>(2) Model (Q4376535)

A process \(X_t\) is \(I(1)\) (integrated of order 1) if its differences \(X_t-X_{t-1}= \Delta X_t\) form a stationary process. A process is \(I(2)\) if its differences form an \(I(1)\) process. The vector autoregressive model with Gaussian errors and \(k\) lags can be written as \[ \Delta^2X_t= \prod X_{t-1}- \Gamma\Delta X_{t-1}+ \sum^{k-1}_{i=1} \Psi_i\Delta^2 X_{t-i}+ \varepsilon_t, \] where \(\varepsilon_t\) are i.i.d. \(N_p(0,\Omega)\), \(\Psi_i=-\sum^{k-1}_{j=i+1} \Gamma_j\). The \(I(2)\) model is defined as a submodel of this, by two reduced rank conditions, namely: (1) \(\prod=\alpha\beta'\), where \(\alpha\) and \(\beta\) are \(p\times r\) \((r<p)\), and (2) \(\alpha'{}^\perp\Gamma\beta'{}^\perp= \xi\eta'\), where \(\xi\) and \(\eta\) are \((p-r)\times s\) \((s<p-r)\), and \(\perp\) denotes the orthogonal complement. The model describes stochastic processes with stationary second differences. A parametrization is suggested which makes likelihood inference feasible. Consistency of the maximum likelihood estimator is proved, and the asymptotic distribution of the m.l.e. is given. It is shown that the asymptotic distribution is either Gaussian, mixed Gaussian or, in some cases, even more complicated.