Distribution and neural network based fuzzy time series models

Fuzzy time series approaches generally have three steps as fuzzification, identification of fuzzy relation and defuzzification. In this study, a new approach and a new method have proposed for forecasting univariate first order neural network-based fuzzy time series (NNBFTS). Firstly, distribution based length approach has used instead of getting a constant length of interval in the step of defining length of interval in partition of universe of discourse. Convenience of operation has provided by creating a new algorithm in the step of fuzzification. Also, weighted indices have firstly used in this step of proposed method. Adjustment of all degrees of membership has provided in the step of identification of fuzzy relation. Not only Multilayer Perceptron (MLP) but also various ANNs such as Generalized Regression Neural Networks (GRNN) and Radial Basis Function Neural Networks (RBFNN) have applied for improving forecasting performance. Various hidden layer and nodes are used for getting the best result instead of being number of nodes which is the sum of inputs and outputs and one hidden layer in NNBFTS forecasting methods for these architectures of ANNs. Proposed method and approach is compared with various NNBFTS or without NNBFTS forecasting methods proposed in the literature by using a data set of enrollment for the University of Alabama which is a well-known and mostly used in the literature and also a big data set of Istanbul Stock Exchange (ISE) national-100 index during 2006-2010 years. The results show that the new proposed method outperforms other methods proposed in the literature.