Feedforward and Recurrent Neural Networks for Time Series Forecasting: Comparative Study

This study aims at examining and comparing the ability of ANNs variances, including MLP, RBFNN, ELMAN and JORDAN in forecasting monthly data of four different time series patterns. As well as to criticize the concept asserting that MLP is a "universal approximator". The tackled patterns include time series with seasonality, trend, combined seasonality and trend, and non-constant variance. Thereafter, based on statistical metrics, the results of this study showed that ELMAN network outperforms the remaining models in dealing with all types of time series patterns. Therefore, we can confirm that ELMAN network is more suited to time series forecasting than feedforward networks, thanks to its embedding memory between its input and hidden layers.