Abstract: In order to improve the accuracy of prediction, a combined sales prediction model is established. The historical sales data is nonlinear, timevarying time series. It consists of two parts, the linear and nonlinear. By using the ARMA model, the linear part can be predicted while the nonlinear part can be predicted by using BP_AdaBoost model. Then the two prediction results are added together. The combination model overcomes the problem of low accuracy by using ARMA model alone. What’s more, it also overcomes the problems that BP neural network model is easy to fall into local minimum. The experiments show that the combination model can improve the accuracy of sales prediction and reflect market rules more accurately and comprehensively.


Key words: ARMA model, BP neural network, AdaBoost algorithm, prediction, combination model