Prediction Of Chinese Stock Market Returns Via Machine Learning

For a long time,people have thought that stock prices are unpredictable.The random walk hypothesis and the efficient market hypothesis point out that if the market is efficient,then economic gains cannot be obtained in this market.Because China's stock market is not a strong efficient market,the stock price is somewhat predictable.Forecasting stock market price trends is a chalenging task,and improving forecast performance will bring high returns.In the past,scholars mostly used statistical and econometric models for the research of stock price prediction.When the number of predictive variables is close to the observed count or the predictive variables are highly correlated,the traditional forecasting method will be interrupted.Gu et al.(2019)found that when using machine learning,valid variables can be selected from a large number of existing variables for stock return prediction.This means that machine learning can compensate for the lack of capacity of econometric methods when dealing with a large number of predictors accumulated over the past 20 years,and can combine existing predictors while evaluating the predictive increment brought by newly proposed predictors.In China,machine learning is still an emerging field,and there are fewer institutions that actual y implement machine learning to predict stock returns.Due to the significant differences between the Chinese and American legal systems,the structure of listed companies,the degree of development of the capital market,and the structure of investors,the suitability of machine learning algorithms for China's stock market remains to be proved.This paper studies whether the applicability of machine learning algorithms in China's stock market can improve the predictive power of the domestic stock market.In addition,with the rise of emerging machine learning algorithms,this paper presents several classic machine learning methods(simple linear model,principal component regression model,penalty linear model,enhanced regression tree,random forest,and neural network model)in Gu et al.(2019),and added emerging machine learning algorithms-LSTM,Xgboost,and Adaboost,to explore whether emerging machine learning methods can outperform classic machine learning methods and further improve prediction capabilities.This paper finds that compared with simple linear models,machine learning methods can significantly improve the forecasting performance of China's stock market returns,and the emerging machine learning methods has better performance than the traditional machine learning methods.The results show that,Adaboost has the best prediction performance and is significantly better than all other models.The neural network algorithm may not perform as well as Adaboost and Xgboost due to the training sample size,parameter settings,and even slightly worse than the random forest model,but it is still better than the linear model.Among them,NN2 has the best prediction performance,second only to Xgboost,NN5 has the worst performance,which shows that when using neural network algorithms,shalow learning is better than deep learning,and The LSTM is better than NN4,NN5 and the network linear model,but worse than other models.The improvement of the linear model(dimension reduction and the addition of penalty terms)did not significantly improve the prediction performance of the simple linear model.The prediction performance of the non-linear model shows that the prediction performance of the non-linear function is significantly better than that of the linear model,and it is more effective in predicting stock returns.This paper also identifies the most important predictors,which are divided into four categories according to Mc Lean and Pontiff(2016).The first category of event factors: Change in Asset Turnover,Post Earnings Drifts,Change in Profit Margins,and R&D change.The second category is market factors,9-month momentum,Total Volatility,and 52-Week High.The third type of valuation factors: Enterprise Component of B/P,Advertising/MV,Enterprise Multiple,Advertising/MV,and the market value ratio of cash flows;the fourth type of factors is the fundamental factor,Age,Accruals,and Earnings Consistency.The linear model and tree algorithm produced very similar rankings of the most informative stock return forecasting indicators,but the importance of variables in neural network algorithms is no rule.In this paper,because of the "blind spots" in the internal logic of the neural network,it is not possible to perform a similar variable importance analysis in the neural network.The research of this paper has certain practical significance and theoretical contribution.This article attempts to apply the emerging machine learning algorithm(e.g.,LSTM,Xgboost and Adaboost)to the research of the return of stock market,compare the ability between emerging machine learning and classic machine learning algorithms in China's stock market prediction and explore the advancement of emerging machine learning algorithms.the work of this paper is helpful to promote crossover study among artificial intelligence,machine learning and financial discipline,and improve prediction accuracy of the stock market in academia and the industry.