Prediction Of Financial Time-Series Based On RBF Neural Network

The existence of non-linearity in financial market movements has been emphasized by various researchers and financial analysts over the last years. Taking both aspects into account, a new kind of financial analysis seems to be necessary:the non-linear analysis of integrated financial markets. Developers of neural computation area provided an interesting mathematical method for this new king of non-linear financial analysis. This method, because of its good learning ability, error correction ability and non-linear approximation ability, has played an important role in the time-series forecasting area.Prediction is one of application of time series. Financial time series data is an important part of time-series which is very closed to people's life. In this case, the research of financial time-series forecast can provide a good reference for people's venture investigation and the state regulation of the economy.This article introduced the financial time-series based on RBF neural network, and indicated that the shortcoming of RBF neural network, then on this basis an improved RBF neural network model was put forward. The improved algorithm mainly adjusted the selection of RBF center point and the radial basic width. In original nearest neighbor clustering, the first input vector into a clustering was selected as the center of the clustering, and according to it we adjusted the average of all vectors in a clustering. The choice of the radial basic width was selected by iterative optimization.In order to verity the feasibility of this new method, we applied the improved RBF neural network modes into the real time financial time-series forecasting system, and by comparing the result with the original RBF neural network modes, we found that both the predication accuracy and training efficiency were increased by the improved RBF mode.Finally, through eliminating the redundant center point of hidden layer reducing the complexity of the structure of the network and enhancing the performance of the unknown model forecast, we modulated the possible over-fitting phenomenon in the improved RBF mode. For each new center point, it will be compare dynamically with the existing center point, and determine whether there is redundancy between them. When inner product vector of these two centers approach to l,one of the center point will be deleted dynamically. Then assign this input vector into the nearest clustering. Via experiments, we found that the redundancy center point was removed successfully by eliminating the over-fitting phenomenon.