A robust algorithm in sequentially selecting subset time series system using neural networks (Q2744941)

The model under consideration is a subset vector rational distributed lag (VRDL) model whose predictor is of the form \( z(t)=-\sum_{i=1}^p h_i^{'}(I_s)y(t+1-i)\), where \(h_i^{'}\), \(i=1,\dots,p\), are parameter matrices and \(h_i^{'}(I_s)=0\) if \(i \in I_s \), i.e. \(I_s\) specifies the integers between 1 and \(p-1\) that correspond to excluded entries. A VRDL model can serve as an infinite moving-average representation of rational vector AR, ARMA and ARMAX models. In the present paper a numerically robust lattice-ladder learning algorithm is developed for the sequential selection of the best specification of subset VRDL models using a neural network approach. The proposed algorithm, which computes concurrently both the a priori and a posteriori residuals in the recursion cycle, possesses better numerical accuracy and is less sensitive to roundoff errors than direct matrix inversion counterparts. Two examples, which demonstrate the usefulness of the proposed algorithm, are discussed.