Applying neural networks in time series forecasting

Neural networks were first applied in time series forecasting in the 1970's. In the early stages this application was known as an adaptive filter. Due to the simplicity of the single layer network leaning mechanisms, adaptive filter failed. Recently, the computing abilities and learning mechanisms of networks are much greater than those of the first wave of neural networks. Therefore, the purpose of this study is to re-examine the possibility of applying neural networks in time series forecasting. Empirical comparisons are made between forecasts from the neural networks and those from traditional time series methods: the exponential smoothing model and the Box-Jenkins ARIMA procedure. The empirical results show that neural networks perform as effectively as the two traditional methods.