ICA/RBF-Based Prediction of Varying Trend in Real Exchange Rate

The flexibility of radial basis function neural network to handle complex patterns in the data has lead to the diffusion and implementation of such models in economics and econometrics. The financial time series such as real exchange rate are multivariate data containing many underlying factors. In this paper, ICA as one of the most popular signal decomposition technologies in recent years is introduced to excavate the potential information for better analysis of real exchange rate where ICA plays an important role of preprocessing. We address the essential difference of dimension reduction using PCA and ICA; show that these two approaches are different at the aspect of sensitivity to dimensions although they both are preprocessing methods of dynamic data, even if the accumulative contribution rate of ICA is less than that of PCA, the former still attains the same prediction results as the latter. With ICA/RBF mixed prediction model, not only we can compress the dimensions of input data greatly, but also find the factors behind to better direct prediction. In the analysis of real exchange rate, independent influence factors such as the policy influence of the enhancement of currency interest and the customer marketing period of exchange are mined to rich knowledge base for make better prediction strategy