Short-term Stock Prediction Using XGBoost

Stock forecasting has always been a hot problem.Many mathematicians,statisticians and economists have done a lot of research on this problem.In short term stock forecasting,the most commonly used model is the time series model.But in this article,we will use the XGBoost model which is proposed by Dr.Chen Tianqi and which has a high attendance rate and good results in many major modeling competitions recently to make a short-term stock prediction.We get about 1.4 million rows of data from Internet using crawler.It contains share price data and financial statements data.Then we build some commonly features,and we build 3 features about pressure and support point on integer stock price basing on the experience we got in building this features.We build 80 features for in the part of data fetching.After that,for measuring the consistency of the relation between features and dependent variable,we use some techniques to apply the Kendall's W to this problem,and then propose 4 corrections for this application.Next,We give the 3 method of feature using Kendall's W and R2,and screen out 3 groups of features with them.Then we use genetic algorithm to tuning XGBoost automatically to get some parameter of XGBoost.Finally,we compare the models which use the 3 groups of features eachly.By this,we show the advantage of the corrections of Kendall's W on this problem in terms of AUC and KS.The main results and innovative points of this article are the 3 features about pressure and support point on integer stock price and using Kendall's W to feature screening and giving 4 corrections.