A new hybrid recurrent artificial neural network for time series forecasting

The forecasting problems can be effectively solved by using artificial neural networks. Classical forecasting methods are not sufficient to forecast nonlinear and complex time series structures such as a stock exchange time series. In this study, a new hybrid recurrent artificial neural network is proposed for nonlinear time series forecasting. The proposed network is a combination of simple exponential smoothing and the single multiplicative neuron model. The combination weights are also weights in the proposed network, and they are automatically estimated in the training processes. The training of the proposed network is achieved by using a particle swarm optimization-based training algorithm. The training algorithm uses restarting and early stopping strategies to prevent overfitting problems. The proposed network is applied to S&P500, Dow Jones stock exchange data sets, minimum temperature data and wind speed data. The performance of the proposed method is superior to two popular deep artificial neural networks and a high-order artificial neural network.