MULTISCALE NEUROFUZZY MODELS FOR FORECASTING IN TIME SERIES DATABASES

摘要

Multiscale neurofuzzy modeling combines the multiresolution property of the wavelet transform with the regression ability of neurofuzzy systems. A wavelet transform is used to decompose the time series into varying scales of resolution so that the underlying temporal structures of the original time series become more tractable; the decomposition is additive in detail and approximation. A neurofuzzy system is then trained on each of the relevant resolution scales (i.e. those scales where significant events are detected); and individual wavelet forecasts are recombined to form the overall forecast. The neurofuzzy models developed in this paper are based on Mamdani and Takagi-Sugeno-Kang approaches to the problem of fuzzy modeling based on the strategy knowledge expressed by the input-output data. Within these approaches, the proposed Neural-Fuzzy Inference System (NFIS) provides several methods that represent different alternatives in the fuzzy modeling process and how they can be integrated with the learning power of neural networks. Simulation results carried out on a forecasting problem associated with stock market, are included to demonstrate the potential of the proposed forecasting scheme.