Prediction Of CSI300Index Based On Support Vector Machine

The stock market is an extremely complex non-linear dynamical system,due to many factors，which make it extremely dififcult to predict stock pricelfuctuation. There are many forecast models of stock price fluctuation both athome and abroad. We can roughly divided into three phases in chronologicalorder: the first phase is structural econometric model, the second is the time-series analysis and the third is intelligent prediction. Due to the nonlinearityof the stock，the traditional measurement of the structural econometric modeland time series model is diiffcult to make predictions. Therefore, this articleuses intelligent prediction methods. Neural network model is the mainrepresentative of intelligent prediction. It does not need to establish preciselogical and mathematical models of the problem to study，but to mimic theway how human being think and to construct neural network algorithm. Youcan get results as long as you directly input data, but the weakness is that theneural network model is easy to make mistakes like over-fitting and localoptima. In this paper，we combine principal component analysis withε-support vector regression model to predict the CSI300Index, which canbe a good solution to the shortcomings of neural networks.This paper first introduces the theory of principal component analysisand support vector machine theory. Second, by introducing an insensitive lossfunction, we construct a stock forecast model based on principal componentanalysis and f-support vector regression. Then，after analyzing many factorsthat affect the stock market, we begin our empirical study of the CSI300Index:we select the real data of the CSI300Index during a period of time, finish thepreliminary data processing with principal component analysis, and operatethe IJBSVM package for the CSI300Index empirical research. Throughcomparing the predictive value with the true value, we verify that the principalcomponent analysis and support vector regression model are valid andfeasible to predict the CSI300Index. Finally, through comparison with BPneural network, we verify that the principal component analysis andg-support vector regression are more accurate in prediction.