Application Of Stock Price Short-term Prediction Based On The Improved Support Vector Machine

The stock markets produce large amounts of transaction data every day, which often implies a lot of useful information, but is not easily discovered by people. As people's dependence on the stock historical transaction data increasing, from the original transaction data mining useful information, and accurately predict the future trend of the stock, to help investors, securities institutions and stock investment professionals for scientific investment decisions, reduce investment risk is of great practical significance.Support vector machine(SVM) technique is a new machine learning method currently developed. It in dealing with nonlinear model identification and small sample learning has good performance, it can also dealing with common problems such as classification problems, regression problems and distribution estimation problems, but it has some inherent drawbacks to be solved in the practical application. The paper focuses on SVM kernel parameter selection and optimization problem, based on the studies of genetic algorithm and particle swarm algorithm optimizing SVM kernel parameters, proposing the improved GA-SVM stock regression prediction model and hybrid PSO-SVM stock regression prediction model, and simulate them by Shanghai Stock Index sample data. Experiment results show that the improved algorithm can obtain better parameter optimization results and predictive effect. Accordingly, the innovation of the paper mainly includes the following several parts:In selecting regression prediction model's kernel function, the paper tests the commonly used four kinds of kernel function, and select from the minimum mean square error(MSE) of the kernel function as the experimental kernel function.In optimizing support vector machine kernel parameter, on the one hand, the paper imports loss function which has significant impact on SVM into the genetic algorithm, establishes(, , ) GA-SVM parameter optimization model based on improved genetic algorithm; on the other hand, the paper lead compressibility factor, random inertia weight, second-order oscillation theory, and the mechanism of natural selection of genetic algorithm in standard particle swarm optimization, proposes hybrid PSO-SVM parameter optimization model.In the part of selecting experiment's stock indicator, the paper according to the degree of impact on the stock's closing price the next day, selecting six experimental indexes ranking- the stock closing price, highest price, lowest price, opening price, trading volume and turnover as the research indicators and analysis data, while the front of the closing price as input variable, other indicators as output variable involved in the model analysis validation.Finally, the paper makes simulation experiment between two kinds of established parameter optimization models and(, )GA-SVM model, GS-SVM model, standard PSO-SVM model three groups of model by organized simple data. The experiment results show that the improved(, , )GA-SVM model and hybrid PSO-SVM model can gain small deviation range and higher prediction precision in the SVM parameter optimization and stock price prediction compared with original model.