Gated Recurrent Unit Network-based Fuzzy Time Series Forecasting Model

Abstract

Time series forecasting and prediction are utilized in various industries, such as e-commerce, stock markets, wind power, and energy demand forecasting. An accurate forecast in these applications is an essential and challenging task because of the complexity and uncertainty of time series. Nowadays, deep learning methods are popular in time series forecasting and show better performance than classical methods. However, in the literature, only some studies use deep learning methods in fuzzy time series (FTS) forecasting. In this study, we propose a novel FTS forecasting model based upon the hybridization of Recurrent Neural Networks with FTS to deal with the complexity and uncertainty of these series. The proposed model utilizes Gated Recurrent Unit (GRU) to make predictions using a combination of membership values and past values from original time series data as model input and produce real forecast value. Moreover, the proposed model can handle first-order fuzzy relations and high-order ones. In experiments, we have compared our model results with state-of-art methods by using two real-world datasets; The Taiwan Stock Exchange Capitalization Weighted Stock Index (TAIEX) and Nikkei Stock Average. The results indicate that our model outperforms or performs similarly to other methods. The proposed model is validated using the Covid-19 active case dataset and BIST100 Index dataset and performs better than Long Short-term Memory (LSTM) networks.