New Procedures for Model Selection in Feedforward Neural Networks for Time Series Forecasting

Abstract

The aim of this paper is to propose two new procedures for model selection in Neural Networks (NN) for time series forecasting. Firstly, we focused on the derivation of the asymptotic properties and asymptotic normality of NN parameters estimator. Then, we developed the model building strategies based on statistical concepts particularly statistics test based on the Wald test and the inference of R2 incremental. In this paper, we employ these new procedures in two main approaches for model building in NN, i.e. fully bottom-up or forward scheme by using the inference of R2 incremental, and the combination between forward (by using the inference of R2 incremental) and top-down or backward (by implementing Wald test). Bottom-up approach starts with an empty model, whereas top-down approach begins with a large NN model. We used simulation data as a case study. The results showed that a combination between statistical inference of R2 incremental and Wald test was an effective procedure for model selection in NN for time series forecasting.