SVR Model And Its Predict For Economic Data

Support Vector Machine is a new and highly efficient machine learning method developed from the basis of statistical theory. It was firstly proposed by the professor Vapnik in the 1990 s. In recent years, the researches on SVM had made rapid development, including theoretical research, algorithm implementation research and application research. Particularly in the field of application research, SVM is widely used to solve many of the practical application problems. SVM is divided into Support Vector Regression and Support Vector Classification. SVR has good performance on data prediction, and the generalization ability of SVR model is strong,even in long range, the accuracy of predicted values is still high. Thus, SVR is widely used to predict data for a lot of practical problems. However, in the economic research field, the researches that adopted SVR to predict economic data is still lack, this thesis will apply SVR to predict three economic data, including electricity consumption prediction, commercial house price prediction and stock price prediction. This thesis mainly studied and discussed in the following areas:(1) For electricity consumption prediction, the multi-variable SVR model and two-variable SVR model are both established and compared. The experiments show that multi-variable SVR model has higher prediction accuracy than two-variable SVR model. This thesis also compares SVR with Autoregressive Integrated Moving Average Model, it is concluded that SVR has better performance than ARIMA when they are both applied to predict electricity consumption.(2) Because of the highly nonlinear character of stock price, a single prediction model is difficult to describe the overall character of stock price. So this thesis proposes a combination prediction model of ARIMA-SVR. Firstly, mallat algorithm of binary orthogonal wavelet is used to wavelet decompose and reconstruct stock closing price, the low frequency and high frequency information has have been obtained. Then the ARIMA model of the high frequency information and the SVR mode of the low-frequency information are established to train and predict data,respectively. Finally, the prediction results of combination model are got by integrating the results of two models. The experimental results show that ARIMA-SVR model is better than the single SVR model.(3) Multi-scale wavelet kernel is constructed as SVR kernel function by combining wavelet analysis knowledge with kernel function knowledge. Firstly,MWK kernel function expression is constructed through Morlet wavelet andmulti-scale kernel function; then adopting MWK kernel function, SVR model is used to predict the stock closing price. The experiments show that the SVR model adopting multi-scale kernel function has better performance than the single SVR model.