View article

Prequential model selection is a data-driven methodology for selecting between rival models on the basis of their predictive ability where the predictive ability of a model is measured by its accumulated prediction error on a given set of observations. Given i.i.d. observations, we propose a regression estimator-based on neural networks-that selects the number of "hidden units" using prequential model selection, and establish a rate of convergence for the statistical risk of the proposed estimator.