Application Of Time Series Analysis Methods In Agricultural Economic Forecasting

It is always difficult to predict time series data because of its notable timing characteristics, complicated nonlinearity and tough dynamic continuity. Whether the prediction results are good or bad depends on the data analysis methods and the prediction models. A proper prediction model is the premise and foundation of accurate forecasting. Order determination, distributor selection and training sample selection are still the key but difficult points of time series analysis which have significant impact on the forecasting results. But they always attach themselves on the actual prediction model. Agricultural economic time series is the classic but complicated time series data. Hence, in this paper, we chose the two most popular time series prediction methods Back-propagation Neural Network (BPNN) and Support Vector Machine Regression as the forecasting models. Both BPNN and SVR have the excellent fitting ability to nonlinear systems. Then, aiming at the improvement of prediction methods and time series analysis technologies, we proposed two time series prediction methods with high forecasting ability. One is an improved BPNN method named as REMCC-BPNN which can optimize the training net of BPNN according to the minimum correlation coefficient of the absolute value of the N nearest neighbor training samples’fitting relative error and the N training samples’ time order. Another is an improved SVR method named as GS-RSR-SVR which determines the order of time series data based on the geo-statistics semi-variogram function range and selects the training samples according to the minimum correlation coefficient of fitting absolute relative error of training sets of different rejected sizes and the samples’ time order.1) REMCC-BPNNThe prediction results of traditional BPNN are always not good enough due to the defects of complicated operating process, difficult to determine the parameters, easy to fall into local minimum. REMCC-BPNN determines the training net of BPNN according to the minimum correlation coefficient of the absolute value of the N nearest neighbor training samples’ fitting relative error and the N training samples’ time order. Namely, the parameters of BPNN are determined as an entirety. The principle is simple and the process is easy to operate. Furthermore, several real-world datasets, the grain yield from1985to2011in China, the index of gross agricultural output value from1952to1980in China and the index of gross agricultural output value from1978to2008in China, were used to test the effectiveness of REMCC-BPNN. The results showed that the prediction accuracy of REMCC-BPNN are better than that of several frequently-used prediction models for time series, such as BPNN, SVR, ARIMA and CAR. The REMCC-BPNN has high prediction capability, good stability and strong generalization ability. It has an extensive application prospect in the fields of time series forecasting.2) GS-RSR-SVRTime series has the characteristics of aftereffect. That is, dependent variable yt is not only affected by independent variables at the moment of t, but also relative to dependent variable and independent variables at the moment of t-1, t-2and even much earlier. Therefore, it is necessary to take former dependent variable and independent variables into account as independent variables for a certain test sample. However, there are some defects in most existing order determination methods, such as much too complicated procedure, horrible time consuming, inadequate final order and so on. GS-RSR-SVR complements the process of order determination rapidly and adequately based on the geo-statistics semi-variogram function range.Due to the empirical selection procedure of original distributors and the process of order determination, there is a lot of overlapping information existing in the independent variables. It is necessary to eliminate those overlapping information to promote the prediction accuracy. GS-RSR-SVR screens the independent variables by leave-one-out method based on minimum mean square error (MSE) to remain as little as possible independent variables to cover all the information impacting on the dependent variable.The whole historical samples or samples selected by fixed scroll window method can not clearly reflect the strong timing relationship of time series. GS-RSR-SVR rejects the oldest samples uninterruptedly according to the minimum correlation coefficient of fitting absolute relative error of training sets of different rejected sizes and sample time order, which can simultaneously reject the useless old samples effectively and keep the strong timing characteristics of time series.Furthermore, several real-world datasets, the grain yield from1985to2011in China, the index of gross agricultural output value from1952to1980in China and the index of gross agricultural output value from1978to2008in China, were used to test the effectiveness of GS-RSR-SVR. The results showed that GS-RSR-SVR has higher prediction precision and more stable prediction ability than MLR, ARIMA, CAR, BPNN, SVR and SVR-CAR, even REMCC-BPNN. GS-RSR-SVR has an extensive application prospect in the fields of time series forecasting.