New methods for training multi-layer perceptron network for forecasting problems are presented. The first method exploits spectral characteristics of time series to get faster learning and improved prediction accuracy. A neural network scheme for real time implementation of this method is also presented. The second method suggests the use of two new weight initialization schemes which give very fast convergence besides giving better prediction. The foreign exchange time series is used to illustrate the efficacy of the proposed methods.
展开▼
