A time series prediction model using constructive neural network

Abstract

In time series forecasting, the artificial neural networks (NN) such as the popular multilayer perceptron (MLP) may be used to handle both linearity and nonlinearity underlying the data generating process, but finding a right network size such as the number of hidden layers and/or hidden units is always a troublesome and time-consuming task. This paper presents a time series prediction model that is based on the use of one-hidden-layer (OHL) constructive neural networks (CNN). The CNN training begins with an initial OHL NN that only has one hidden unit. New hidden unit is added one at a time to the existing network according to the complexity of the data being modeled, which makes the CNN more capable than the fixed-size NN. A modified quick-prop algorithm is used to perform the input-side training of the CNN hidden units. The CNN-based model is applied to three types of real-world data sets to demonstrate its superiority over the AR and the fixed-size NN models.