New lattice realisation of the predictive least-squares order selection criterion

The initialisation procedure used within the implementation of the predictive least-squares (PLS) order selection criterion is studied. From this study emerges the motivation for a new growing memory covariance least-squares lattice implementation of the PLS order selection criterion (GMCPLS), which is realised in a computationally efficient lattice structure. By extensive simulation of a wide range of generated AR time-series, the true model order hit-rate is shown to be generally higher than that for a prewindowed implementation of the PLS algorithm