A Hybrid ARIMA–ANN Model for Forecasting Egyptian Exports

Abstract

 The present study examined the potential of a hybrid model based on the autoregressive integrated moving average (ARIMA) model and multilayer perceptron artificial neural network (ANN) for predicting Egyptian exports. Box–Jenkins methodology was applied to the annual data 1960–2016. Four ARIMA models (ARIMA(0,1,1), ARIMA(1,1,1), ARIMA(2,1,1), ARIMA(1,1,2)) were estimated and compared using the Bayesian information criterion (BIC), diagnostic residual analysis (DRA), and parameter significance. Due to the low reported BIC (47.223), the linear component was chosen as ARIMA (1,1,1) and the reported diagnostic conditions were met. Finally, the estimated ARIMA models and lagged export observations were used as inputs in the ANN model. The performance of forecasting was evaluated with mean absolute error (MAE), root mean squared error (RMSE) and mean absolute percentage error (MAPE). The lower errors presented by the hybrid model compared to time-series model only were reported as MAE decreased from 523.451 to 32.594, RMSE from 7211.342 to 58.977, and MAPE from 12.859% to 4.623%. From these results, it is concluded that the ARIMA–ANN framework is useful for the export series analyzed. The number of annual observations should be reconcilable with the number of cases reported in the training/testing case summary for the 1960-2016 period prior to journal submission the number of cases in the dataset should be 57, and there are 56 cases reported in the training/testing case summary.