Innovations algorithm for periodically stationary time series (Q1613633): Difference between revisions

Let \(\{\tilde X_t\}\) be a discrete process. Define \(\mu _t=E\tilde X_t\) and \(\gamma _t(h)=E\tilde X_t\tilde X_{t+h}\). If \(\mu _t\) and \(\gamma _t(h)\) are periodic functions of \(t\) with the same period \(\nu\), then \(\{\tilde X_t\}\) is called periodically stationary. The periodic ARMA process \(\{\tilde X_t\}\), briefly PARMA\({}_{\nu }(p,q)\), is defined by \[ X_t-\sum _{j=1}^p \phi _t(j)X_{t-j}=\sigma _t \varepsilon _t-\sum _{j=1}^q \theta _t(j)\sigma _{t-j}\varepsilon _{t-j}, \] where \(X_t=\tilde X_t-\mu _t\) and \(\{\varepsilon _t\}\) are i.i.d. (0,1) variables. Two cases are considered, one with finite fourth moment of \(\varepsilon _t\) and the other with finite variance but infinite fourth moment of \(\varepsilon _t\) (under the assumption that the innovations have regularly varying probability tails RV(\(\alpha \)) with \(2<\alpha \leq 4\)). The authors adapt the innovations algorithm for estimating the parameters of the PARMA model. It is proved that the estimates are weakly consistent. If \(\varepsilon _t\) has infinite fourth moment, then the proofs require that the PARMA process is causal and invertible. The results are valid also for classical ARMA models since they are special cases of PARMA with \(\nu =1\).