The Research Of Time Series Prediction Based On Support Vector Machine

Time series is a series in accordance with the time order. In practical problems, the system are generally nonlinear, and the time series are almost non-stationary. It is often difficult to create the model by the mechanism of the system, we can use the input andoutput data to predict the unknown data.This paper introduces a variety of modeling and forecasting principles about time series, through analysis and comparison, we found statistical method, gray forecasts and neural networks have some advantages, but can not meet the requirements of complex time series, so support vector machine for time series is put forward in the thesis. Taking shanghai stock index time series prediction as an example, the simulation results show that the SVM is an effective method of time series forecasting.The parameters of SVM have great influence on the learning ability and generalization ability, so far there is no unified theory to guide this field, people always use large numbers of experiments and selecte the satisfied solution, but this method was time consuming and the parameters are not optimal.Cross certification,experience formula and genetic algorithms are also inadequate. This paper introduced the PSO algorithm that is simple and effective, but easy to a local minimum.So the QPSO algorithm which has stronger global search capability is used, but the convergence is not satisfiyed.A improved QPSO algorithm is put forward and the sunspot time series are take as an example for the simulation research. The results show that the improved QPSO has stronger global search, faster convergence, and the forecast error is smaller. That is an effective method of optimization of the parameters of support vector machines for time series prediction.