Research On Classification Predication Of Different Stock Returns Based On Support Vector Machine

During the last couple of decades,with the expansion of the security market and maturity of the mechanism,a growing number of institution investors and individual investors are participating in the stock market.Forecasting the return on stocks is a perpetual proposition for which the practitioners and researchers are pursuing.A lot of methods related to the classification predication on stocks are also resulting from these.Machine learning algorithms are emerging as predictive analytic techniques that can process large amounts of structured or unstructured data.It takes advantage of the parameter estimation mechanisms that avoid overfitting and the powerful adaptive learning capabilities that can be used to excavate the useful information behind the data.This article will use one of the classification algorithms to analyze the classification predication on stocks and apply on the strategy.Compared with other classification methods,Support Vector Machine algorithm takes its unique advantages as follows.Firstly,there is a higher classification performance of Support Vector Machine(SVM)under the circumstance of small sample sets.Secondly,SVM based on structural risk minimization gets the global optimal solution,avoiding falling into local extremes.Finally,by transforming the nonlinear problem into a linear problem of high-dimensional space and replacing the inner product operation in the high-dimensional space with a kernel function,SVM solves the complex computation problems artfully and overcoming the “dimension disaster” effectively.Supporting vector machine's powerful generalization ability and nonlinear mapping capability of kernel function contribute to a new method in researching classification predication of stocks.The construction of the stock selection model based on support vector is generally divided into five basic processes.Firstly,construct the feature data and corresponding target data of the learner based on selecting the sample target and sample interval.Secondly,datapreprocessing,which can ensure the accuracy and completeness of the data set used for learning and training,It performs operations such as outlier processing,missing value processing,and data normalization.Thirdly,feature dimension reduction,which can make the model more simple and effective by extracting the new and more concise dimensions from the original feature dimension or eliminating irrelevant features.Fourthly,the choice of kernel function and parameter optimization.Select the appropriate kernel functions to solve the nonlinear mapping problem in high-dimensional space and optimize the model's hyperparameter.Fifth,the evaluation,selection and verification of the model.The classification performance of the model can be measured by using the accuracy,precision and recall indicators and the confusion matrix.Cross-validation can be used to find the optimal performance classifier in the training set and test set.In this paper,14 input features that can be used for the classification model are finally selected from 26 kinds of vision factors through factor validity test and factor correlation analysis after the initial factor data is preprocessed and standardized.Take 30% as the category label for the strong and weak stocks by sorting the stock return rate.Each period of the learning training sample selects the longest historical data set within the sample interval.Firstly,support vector is used to learn feature data and target data.Both the grid search and cross-validation method are combined to auto-optimize the model hyperparameter.The learner with higher classification accuracy and stronger generalization ability was obtained.The data show that the classification model was practical under the95% confidence level.The accuracy of out-of-sample forecasting reaches 70%.Then,the model of researching on classification predication of different stock returns based on Support Vector Machine is contributed to a multi-factor stock selection method.According to the level of return on stock,the stock portfolio is constructed to form a multi-factor stock selection strategy based on support vector and the retrospective analysis shows that the multi-factor stock selection strategy based on SVM can obtain higher stable excess returns.What's more,from a statistically significant point of view,the strategic portfolio performance can be further evaluated from the perspectives of considering transaction costs,sharp probability value and threshold sample size.The evidence shows that the strategic of Sharpe ratio is significantly higher than the market benchmark is reliable if under the bettercontrol of transaction costs.Finally,SVM multi-factor selection strategy is compared with single-factor strategy,traditional multi-factor strategy,and other machine learning algorithm strategies.The evidence shows that the SVM classification model can accommodate the characteristics of variable nonlinearity and improve the combined efficiency of multiple factors.Compared with the traditional multi-factor and other predictors of machine learning algorithms,SVM multi-factors are better at evaluating strategy risks and benefits.