This paper presents a new mixture of experts neural network architecture for the prediction of the US Dollar Swiss Franc exchange rate. This architecture achieves improved prediction results on noisy and non-stationary data. In contrast to previous efforts the current system was designed with a particular emphasis on solving the problems of local overfitting & underfitting caused by non-stationarity and noise in the data. The cascade correlation constructive neural network training algorithm was used for the fast training of near optimal complexity global & local experts. The Kohonen Self Organizing Map was used to find regions of the data on which to train local experts. Improved results were obtained by using a combination of the outputs of the global & local experts.
展开▼
