A new criterion of NN structure selection for financial forecasting
A. Perrone, Gianfranco Basti · 2003
For the evaluation and the selection of the optimal neural net (NN) structure complexity, as a function of the minimization either of the approximation error or of the generalization error, we discuss briefly the minimum description length (MDL) method. Because of the theoretical and practical limitations of this criterion-overall for stochastic time series previsions-we shortly introduce our new dynamic sampling window (DSW) method for the optimal NN structure definition for financial forecasting.