A Comparative Investigation of Alternative Estimation Methods in Non-Stationary Time Series: Analysis of Cotton Price

In this study, comparative analysis of the predective techniques used for non-stationary time series had been made.the method of box-jenkins for univariate analysis, multiple regression analysis and artificial neural networks techniques for multivariate analysis had been used. the monthly prices of cotton in turkey and world and also the monthly price of wheat and corn in turkey used as time series data. %10 of the data hadn’t been included in the analysis to measure the predictivite power of analysis methods. at the end of the analysis, predicted values obtained about the last 18 months period had been examined in mutual. the stationary of the variables were examined by applying unit root tests. zivot-andrews structural break unit root test which allows single structural break and lumsdaine-papell structural break unit root test which allows two structural breaks has been applied to turkey cotton prices. past values of prices turkish cotton has been applied with box-jenkins method. it has been concluded that the most appropriate level of arima (3,1,2). the dependent variable is turkey cotton prices, independent variables are world cotton prices, turkey wheat prices and turkey corn prices and lagged values of all variables for multivariate methods. it has been examined results of regression analysis. optimal network architecture (7,7,1) were found for artificial neural networks. mape, sse, mse and rmse performance criteria were used to compare the predicted values. it was determined that artificial neural networks gave more reliable results than other estimation methods.