Stock Price Index Prediction Based On Nonlinear Sparse Multi-kernel Method

As a part of the capital market,the stock market is not only an important market for enterprise financing and risk diversification,but also plays a huge role in the process of adjusting income distribution.Predicting the trend of stock prices has always been a concern of business and academic circles.With the improvement of machine learning theory,it has become a new direction to use large-scale and high-dimensional data combined with machine learning technology to predict the trend of stock prices.From previous studies,the accuracy of prediction has been improved.However,the traditional nonlinear machine learning model still has many shortcomings in the practice of stock price index prediction.Aiming at these shortcomings,the nonlinear sparse multi-kernel method improves the existing nonlinear support vector machine(SVM).Firstly,in view of the complex composition of the influencing factors of the stock price index and the different influencing data are not in the same feature space,the nonlinear sparse multi-kernel method uses multikernel support vector machine to try different kernel function combinations to describe the various influencing factors of the stock and predict the stock index fluctuation.Secondly,the stock price index belongs to a time series.But multi-kernel SVM as a model of dealing cross-section data,its temporal expression ability is insufficient for financial time series forecasting.The nonlinear sparse multi-kernel method adopts the way of moving window on different period of stock price index model respectively.And with the passage of window,the model can be retrained to simulate the fluctuation trend of real stock price index more closely.Thirdly,the training speed of the model are demanding in the actual background.However,due to the large scale and low information of data,the general nonlinear models have to spend a lot of time to training.So,the nonlinear sparse multi-kernel method used the sparse strategy.XGBoost is first used to conduct sparse screening of features,while eliminating redundant features,effective features are enhanced.Then,a sparse strategy with multi-kernel weight is adopted to reduce the training time of the model and improve efficiency.In terms of empirical test,the SSE 50 index data from January 2018 to May 2020 is taken as an example to verify the effectiveness of the nonlinear sparse multi-kernel method.Firstly,the prediction results of the single kernel model and the sparse multikernel model are compared.Secondly,the performance of the original data and the sparse data in the multi-kernel model is compared.Thirdly,the influence of different sparsity degree on the prediction effect of the model is discussed.Fourthly,the prediction effect of the sparse multi-kernel model with different window step sizes is discussed.Experimental results show that the sparse multi-kernel method is better than the single kernel support vector machine in predicting the stock price index.Compared with the original data,the prediction results of the data with sparse feature processing are significantly better than the former in terms of training time while keeping the accuracy rate unchanged.Both the sparse degree of features and the moving step size of the training window will have an impact on the experimental results,which need to be adjusted according to the changes of data in practical application.