Research On Quantitative Strategy Of Machine Learning In China's A-share Market

The development of quantitative finance in foreign countries has matured over the past decades.Quantitative investment has become one of the mainstream investment methods and research contents.The quantitative investment method has also been widely used in China's A-share market.During recent years,under the wave of artificial intelligence,major financial institutions have continuously applied technical methods in the fields of machine learning and data mining to quantitative transactions,and new technical methods have gradually been applied in securities analysis.In this context,this paper chooses a variety of machine learning classification algorithms that are different from traditional linear regression,and regards the stock selection problem as a classification problem.Combining asset pricing theory with the multifactor quantitative model under machine learning algorithm.This paper selects a total of 35 factor indicators to build a factor library in six categories,and selects the monthly data of all listed companies in the A-share market from May 2004 to December 2018 as a sample.At the end of each month,the factors in the accounting factor pool are used as the sample.The characteristic data is used to calculate the stock excess return of the next natural month as the label related data.Using stochastic gradient descent,support vector machine,naive Bayes,random forest,neural network to learn the feature and label data on the training set and parameter tuning,using the parameter-adjusted model to construct multi-factor quantification on the test set The strategy compares the strategy results with the linear multi-factor quantization strategy and gives an analysis of the multi-factor quantization strategy under the machine learning algorithm based on the results.By analyzing the performance of industry-neutral strategies based on each model,it is found that each machine learning model can obtain positive annualized excess returns in the retracement interval,but the retracement performance of each model is inferior to the HS300 index,in which the random forest model The prediction performance is optimal,and the stochastic gradient decline model has the worst prediction performance.The comparison with the traditional linear regression model shows that the overall performance of the random forest,neural network and naive Bayesian model is better than the linear regression model,but the retracement performance of the neural network model is inferior to the linear regression model.This paper argues that machine learning models that rely more on input data are more sensitive to changes in future markets.When faced with unknown market information,machine learning models may fail,and prediction results may deviate significantly from market results.