This research examines two neural network approaches for forecasting business and economic time series. Neural networks have been proposed as alternatives to traditional statistical techniques because they are "model free" (Kosko 1992, Sastri 1994). This study compares the results from traditional multilayer feedforward neural networks (MFNNs) trained using backpropagation and from radial basis function (RBF) networks with the results from the M-competition (Makridakis, et al. 1982). The use of the M-competition data allows for comparison of the neural network forecasts with those produced by traditional statistical methods.; Specifically, this research addresses the following major issues: (1) How does the forecasting accuracy of MFNNs compare with that of RBF networks? (2) How does the forecasting accuracy of the neural network approach compare with that of traditional statistical techniques? In addition, this study (1) examines two different ways to control overfitting in MFNNs, (2) considers several heuristics for determining the input dimension for RBF networks, (3) examines two different radial basis functions, and (4) presents new algorithms for locating centers and detecting and removing trends in data.
展开▼
