The Application Of The Combinational Forecasting Model In The Pig Price Forecasting Of Jilin Province

In China's agriculture, breeding pigs plays an important role.With respect to predicting the pig price accurately, it is not only the basis to make the right decisions for the breeding owners and the marketing managers, but also the implementation of the scientific development of China, the important means of developing agriculture .Agricultural market information is various and complex,which often makes people helpless, and difficult to make decisions on the market. In order to conduct a rational analysis to the market,it is very urgent to know how to predict the pig price accurately .However, at current China's pig price forecasting exists problems such as the wrong trend of forecasting, the low accuracy of prediction, the main research in this area remain in the statistical analysis or the qualitative level, the forecasting results are not very satisfactory, many of the best forecasting system have not yet been applied to this area, the paper is supported by the National High-tech Research and Development Program (863 Program) project, Digital agricultural knowledge grid technology research and application topic, it aims to apply some relatively mature models to the pig price forecasting and to get better precision of pig price forecasting.The paper firstly studied the meanings and characteristics of the predicting, pointed out the characteristics of pig prices forecasting, on the basis, reviewed the forecasting methods which commonly used and their characteristics,such as regression analysis forecasting model, exponential smoothing forecasting model, qualitative analysis method and BP neural network model,analyzed the advantages and disadvantages of these models, explained the rationality and feasibility of combinational forecasting model,and proposed several design ideas and modeling steps for the combinational model.When modeling, the paper used a single forecasting model for the training firstly, the model parameters were optimized through a variety of methods, the model configuration was optimal, in order to get better outcome. When conducting combined forecasts, the model was divided into two parts - factors prediction and the results prediction,in order to combine the advantages of sevral forecasting models to achieve the purpose of improving predicting accuracy. In the part of the factors predicting, single forecasting model was used to train and forecast the factors which affecting the pig price.The predicting results of this part was the input of the results prediction. The regression neural networks prediction model and GM (1,N) model was used as the model of the results prediction. After training, we got the outcome.Nine prediction models were made. three single-forecasting models are: neural network prediction model (ANN), gray system prediction model (GM (1,1)), time series triple exponential smoothing forecasting model (TS).Six combined forecasting model are as follow: ANN + ANN, GM (1,1) + ANN, TS + ANN, ANN + GM (1, N), GM (1,1) + GM (1, N), TS + GM (1, N).When finished modeling,the paper used the data of pig price, which came from Jilin province in 2000 -2008 , as well as its related factors, as input to train the prediction models. In the training process,I selected the model parameters by constantly training to get the best result.After got the result, this paper compared the results by the RMSE error evaluation criteria. Then, the best prediction model was used to forecast the pig price from July 2008 to June 2009. Finally, this paper also proposed to amend the error through prediction in order to improve the model. the error between the true values and the predicted values ,which was treated as a sequence, and modeled to predict error direction, in order to adjust the final result. Because using the same model and the input data to predict ,the longer the distance to predict ,the greater the deviation, so the error was predictability.Through the above work,the following are conclusions:(1) In the single forecasting models, the gray system model is not satisfactory in the long-term prediction, it is suitable for the short-term prediction. The time series triple exponential smoothing are better to forecast the trend of the future data than others. In neural network prediction model, when the number of samples is much larger than the number of input layer nodes, the number of hidden layer node should be designed around the number of samples, so that the model can get a better prediction accuracy.(2) In all the combinational forecasting models, the RMSE value of the GM (1,1) + ANN model is minimum, forecasting result is the best. Even in the long-range projections, it got the same effect.But the predicted results of the combinational forecasting model are not better than a single prediction model always.The ANN+ANN prediction model is the worst combinational model. The reason may be that the relevance of the two systems is greater, so when designing the combinational model, we should consider to choose two models with a low correlation. (3) The example shows that the designing of the error amendment is feasible and can improve the accuracy of prediction.(4) The pig price in Jilin Province from July 2008 to June 2009,was predicted by the combinational prediction model GM(1,1)+ANN, the results coincide with market,it proved that the model could be used as the pig price forecasting on market,and had a certain practicality.The experimental data and results of this part can be made availably to conduct comparative study for the researchers of combined prediction ,The models can be made available to the breeding owners and the marketing managers as a predictor of market conditions in order to adjusting the investment decision-making.In addition, the modeling process and the design of optimal model parameters can also be used as reference for the scholars .who do research related to the same model.The work of the error amendment provides an idea to enhance long-term forecast accuracy.However, in this paper, there are still many deficiencies and aspects for improvement, for example, the combinational forecasting models were limited to a simple series connection which can not integrate the advantages of both models fully, we should do more research on the combinational methods in the future . In addition, the paper used multi-test algorithm for the model parameter optimization,which is less rigorous and scientific. Moreover, the theory of the error amendments is not mature and needs much further study.