Artificial Neural Network Time Series Modeling forRevenue Forecasting

Abstract

The objective of this study is to investigate the effect of applying different number of input nodes, activation functions and pre-processing techniques on the performance of backpropagation (BP) network in time series revenue forecasting. In this study, several preprocessing techniques are presented to remove the non-stationary in the time series and their effect on artificial neural network (ANN) model learning and forecast performance are analyzed. Trial and error approach is used to find the sufficient number of input nodes as well as their corresponding number of hidden nodes which obtain using Kolmogorov theorem. This study compares the used of logarithmic function and new proposed ANN model which combines sigmoid function in hidden layer and logarithmic function in output layer, with the standard sigmoid function as the activation function in the nodes. A cross-validation experiment is employed to improve the generalization ability of ANN model. From the empirical findings,it shows that an ANN model which consists of small number of input nodes and smaller corresponding network structure produces accurate forecast result although it suffers from slow convergence. Sigmoid activation function decreases the complexity of ANN and generates fastest convergence and good forecast ability in most cases in this study. This study also shows that the forecasting performance of ANN model can considerably improve by selecting an appropriate pre-processing technique.