Comparative Study On The Prediction Methods Of Time Series

Time series forecasting is applied extensively in many fields, it uses historical sample to build a statistical model. By this model, we can explain the data regularity, so as to control and forecast the future value. The research about how to model and pre-dict for stationary time series, especially linear model, is long-standing. The common methods are AR model, MA model, ARMA model, least squares, ridge regression, las-so regression and so on, and they all have achieved good results by their advantages in time series. But in reality, there are many time series which data are high-dimensional, mass, non-stationary and nonlinear, so the above methods are no longer suitable for use. And then, statisticians have proposed artificial neural network and support vec-tor machine method. Their nonlinear prediction and practical characteristics are more prominent than those of least squares, ridge regression, lasso regression. Thus the two methods are becoming increasingly popular.This paper focuses on the basic theory of least squares, ridge regression, lasso regression and support vector machine, and makes a comparison between them, and demonstrates the differences by examples. In final, we can achieve that the ridge re-gression method is improved on the basis of least squares, especially when handling the problem of multicollinearity. Lasso regression is much better than ridge regression in terms of coefficient of compression. The support vector machine method can deal with the problem of nonlinearity and high-dimensional pattern recognition, and has strong generalization and learning ability, so it shows good performance in actual application.