Akademik Veri Yönetim Sistemi
13th International Days of Statistics and Economics, Prague, Çek Cumhuriyeti, 5 - 07 Eylül 2019, ss.231-240
Most of the fuzzy-based prediction methods have been not originally designed for prediction problem since they ignore the dependency throughout time series observations. In addition to this, fuzzy time series methods consider dependency structure of time series. Moreover, while they consider the memberships, they do not regard the non-memberships in the prediction process. In this study, a new time series prediction method which also takes into account the non-membership and hesitation values is proposed. In the proposed method, the membership, the non-membership and the hesitation values are obtained from intuitionistic fuzzy C-means. Non-linear relationships between inputs and outputs are determined by two separate Sigma-Pi neural networks (SPNNs). The membership and the non-membership values are separately used as inputs in two separate SPNNs, and crisp values are utilized as targets and outputs in both SPNNs. The outputs obtained from SPNNs are converted into final output i.e. final prediction of whole method via an approach based on hesitation margin. The optimal weights of both SPNNs are obtained by utilizing modified particle swarm optimization. To demonstrate the prediction performance of the proposed method, various real-life time series data sets have been analysed and the obtained results prove the outstanding prediction performance of the proposed method.