When modeling temporal processes, just like in pattern recognition, selecti
ng the optimal number of inputs is a central concern. In this paper, we tak
e advantage of specific features of temporal modeling to propose a novel me
thod for extracting the inputs that attempts to yield the best predictive p
erformance. The method relies on the use of estimators of generalization er
ror to assess the predictive performance of the model. This technique is fi
rst applied to time series processing, where we perform a number of experim
ents on synthetic data, as well as a real life dataset, and compare the res
ults to a benchmark physical method. Finally, the method is extended to sys
tem identification and illustrated by the estimation of a linear FIR filter
on functional magnetic resonance imaging (fMRI) signals.