EXTRACTING THE RELEVANT DELAYS IN TIME SERIES MODELLING
      =======================================================

                           Cyril Goutte


                Department of Mathematical Modeling
            Technical University of Denmark - Bygn. 321
                      DK-2800 Lyngby, Denmark


Abstract.  In this contribution,  we suggest  a convenient  way to use
generalisation  error to  extract  the  relevant  delays  from a time-
varying process,  i.e. the  delays that  lead to  the best  prediction
performance. We design a generalisation-based algorithm that takes its
inspiration from  traditional variable  selection,  and more precisely
stepwise forward  selection.  The method is  compared to other forward
selection  schemes,  as well as  to  a non-parametric  tests  aimed at
estimating  the  embedding   dimension  of  time  series.   The  final
application extends  these results to  the efficient estimation of FIR
filters on some real data.


This is the final paper, to appear in: Neural Networks for Signal
Processing VII -- Proceedings of the 1997 IEEE Workshop, Florida.