Time series forecasting using Fuzzy Functional link neural network trained by improved second order Levenberg-Marquardt algorithm

Fuzzy neural networks are found to be universal approximators when all the parameters of the network are adjusted simultaneously. Therefore, they have been used extensively classification and regression problems. Normally FNN uses TSK-type fuzzy rules where the consequent part of the rule base comprises linear terms. Thus FNNS may not be able to effectively handle chaotic time series like stock or electricity price time series by providing an accurate mapping. In this paper, therefore, the consequent part of the time series comprises the output from a functional link network that provides an expanded input dimension to handle the uncertainties and chaotic fluctuation of the time series databases. Further to improve the training speed for the weights of the FNN a second order Levenberg-Marquardt (LM) algorithm is used. The performance of the Fuzzy Functional link hybrid (FLFNN) is evaluated for two financial time series data bases providing excellent prediction accuracy.