Prediction And Analysis On The Direction Of Stock Rate Return Using Machine Learning

With the development on quantitative finance theory,quantitative investment technology has been widely applied and brought huge returns to investors.Nowadays,artificial intelligence and machine learning are popular,and great achievements have been made in many fields such as image recognition and search recommendation.Compared with the traditional time series model,the machine learning model can process and analyze massive data quickly,and often has better generalization ability,so it has a more extensive application prospect in financial data mining.In this paper,we try to apply relevant machine learning algorithms into financial data mining.Proposing a data mining method using a newly proposed model XGBoost,study the direction of stock rate return as a"pattern recognition"problem,and compare and verify the effectiveness and robustness of each algorithm.Firstly,considering that stock market is a complex system with low signal-to-noise ratio,use wavelet decomposition and threshold denoising to filter the noise in the data.By decomposing the data into sub-signals with different frequencies,the high-frequency data is cleaned and filtered to further extract the effective information.Secondly,a variety of machine learning models including extreme gradient boosting tree XGBoost,random forest,support vector machine,neural network,logistic regression and so on are introduced to study the direction of stock rate return as a classification problem.XGBoost was proposed by Chen Tianqi in 2016.It has been proved to be an efficient algorithm in many competitions,and it has been rarely used in financial data mining.Taking the components of CSI 300 index as samples,technical indicators,fundamental indicators and public opinion indicators are comprehensively considered,and the significance on each indicator is tested by Boruta algorithm.Through sliding window modeling and out-of-sample prediction,it has found that the accuracy of XGBoost algorithm is the highest,with the accuracy rate reaching 54.7%approximately,and the running speed is greatly improved.Furthermore,backtest is carried out according to the predicted signal and the strategies constructed by each algorithm can generate excess return.At the same time,a layered backtest is designed to further test the robustness of the algorithm,and verify that the algorithm has certain ability to identify the direction.Finally,considering the previous studies tend to only focus on prediction accuracy while ignoring the mechanism of the model,we further analyze the logic of the model.Defining a variable weight measurement method,and found that such as PE ratio,PB ratio is relatively important in the model.What's more,partial dependence relationship is introduced to measure the relationship between various indicators and the direction of return.It has found that PE and PB are negatively correlated with the direction of return,while ROE and weekly inflow are positively correlated.Through those analysis,to some extent,it can make up for the "black box" problem existing in machine learning and make the logic of investment strategy clearer.