Development of a hybrid forecasting model using ARIMA and artificial neural networks(ANN)

Time series forecasting is a complex procedure especially if underlying data include linear and nonlinear components together. In this study we employ a linear method Autoregressive Moving Average (ARIMA) and a non-linear method (Artificial Neural network, ANN, approach using Multi Layer Perceptron) to capture the complex behavior of the time series data. In our research approach we first use ARIMA and ANN approaches separately on several well known data sets from the literature. Then, a hybrid methodology that combines ARIMA and ANN is tested on the same data sets. The forecasting performance of the ARIMA, ANN and hybrid approaches is presented and compared with the previous work from the literature. Different forecasting metrics are used in evaluating performance of the proposed hybrid method. A generalized hybrid model is then used on Turkish wheat yield data. It is observed that hybrid model performance is superior when it is compared to the performance of standalone ARIMA or ANN approaches on tested data set.