Abstract: It has been shown in recent economic and statistical studies that combining forecasts may produce more accurate forecasts than individual ones.However,the linear combination forecasting method is known to have the limitation for composite modeling and forecasting of foreign exchange rates.This paper presents a new nonlinear composite forecasting method for exchange rate modeling and forecasting based on fuzzy neural network.Furthermore,the corresponding composite learning algorithm including Kohonen self organizing future map,supervise competitive learning and back propagation are used to learn the connection weights of fuzzy neural network and partitions of fuzzy subsets.It has been shown by the composite modeling and forecasting results about the exchange rate time series of the British pound,the French franc,the Swiss franc and the Japanese yen against U.S. dollar that the method has reinforcement learning properties and mapping capabilities.With respect to composite modeling and forecasting of nonlinear system which has some uncertainties,the method is available.

Key words: exchange rates, combining forecasts, information sets, fuzzy neural network