Forecasting Short-Term and Medium-Term Time Series: A Comparison of Artificial Neural Networks and Fuzzy Models

Abstract

The study is focused on experimental comparison of time series models of two classes, namely, artificial neural networks and fuzzy time series models. In each class, three basic models of the time series were selected to compare their predictive abilities, which were evaluated by the MAPE criterion. To the class of models using artificial neural networks, multilayer perceptron networks, as well as RNNs, containing LSTM or GRU blocks in the hidden layer, were investigated. In this study, three basic fuzzy time series models were used: the model with fuzzified time series values, the model with fuzzified first differences of time series values, and the model based on the elementary fuzzy tendency. A comparative study was conducted based on dataset of time series, which were divided into two groups relative to the length of time series. An experimental study showed that for medium-term time series on the test interval, the RNNs based on LSTM showed the smallest error on average (MAPE = 3.0013%), and for the short-term time series, the fuzzy models showed the smallest error on average (MAPE = 5.7313%), while models of the ANN class predicted the short-term time series with MAPE > 9.8% in average.