‫استخدام نماذج بوكس-جنكنز‬ ‫ونماذج الشبكات العصبية‬ ‫الاصطناعية للتنبؤ في السلاسل‬ ‫الزمنية للقطاع الزراعي السوداني‬

Abstract

Abstract This study tackled the use of Box-Jenkins (ARMA) and Artificial Neural Networks (ANNs) Models in Economic time series forecasting. The application was on Sudanese Agricultural Sector represented in the time series of the productivity of wheat, Sorghum, groundnut and sesame during the period from 1954 to 2005. The aim of this study is to bring out the relationship between the methods used in time series prediction and the degree of accuracy of predictions obtained in one hand, and what effects of the time series variations that happen, and the randomness and non-linearity of data, have on the performance of these methods. What distinguishes this study from previous studies is that, it approaches the field of "time series prediction" from various angles as shown below: - The study dealt with type of data as a major factor in determining the method used in the prediction. - The effect of different variations, particularly the random variation on the results of the prediction models. - The effect of non- stabilization of variance on the accuracy of the prediction obtained from the model used. - The application of ANNs model on agricultural time series. Most of the economic ANNs applications were confined on the financial time series. The most significant findings of this study are: 1- The models: -Box-Jenkins models: wheat series ARIMA(0,1,1) , Sorghum natural logarithm models ARIMA(0,1,1), ground nut ARIMA(0,0,0) and sesame ARIMA(3,1,0). -Neural Networks models: these models were built by using Multi Layer Perceptrons (MLP), whose architecture consists of three layers (input layer, hidden layer, and output layer). The logistic function was used as an activation function in the hidden layer and the output layer, and the quick propagation algorithm was used for training these networks. 2- The degree of variations and particularly the random variations has direct effects on the results obtained through both of the studied methods. The greater the variations in time series, the less efficient ARMA models in comparison to ANNs models. 3- The higher non-linearity in the time series data, the lower efficient the models of ARMA are, in the prediction. 4- Both of the studied methods have a problem dealing with the time series that have non-stabilization of variance. However, the models of ANNs are preferable ARMA models. 5 5- The ANNs models are clearly influenced by the amount of data available (time series length). The larger and sufficient the amount of data to show all the variations in series, the higher learning will be attained in the network and this increases the efficiency of ANNs in prediction. 6- The longer the duration of the prediction in future, the more accurate the ANNs results than ARMA findings. That is clear from the prediction results obtained from these models. According to these findings, we recommend the following: 1- It is preferable using ARMA models in time series that are less complicated. The more complicated the series gets, the more preferable using the ANNs. 2- To raise the efficiency of ARMA and ANNs models in predicting the time series, we should remove the variations influence from the data of the time series before applying these methods. 3- It is preferable using the models of ANNs to the models of ARMA on the data with noisy and non-stabilization of variance. 4- If the time series is not long enough to show all the variations clearly, it is preferable using ARMA models to the ANNs models