Research On Quantitative Investment Strategy Based On Stacking Integrated Learning Model

The stock market is a complex system,and its changes are affected by a large number of factors.Many scholars have gone forward and tried to establish various models to explain the changing stock market.In recent years,the accumulation of data and the geometric level of computing power have made artificial intelligence methods shine in many fields(such as image and abnormal detection),and even surpass humans.The application of artificial intelligence in the field of stock investment is the focus of current research.The research object of this thesis is the monthly data of the constituent stocks of the CSI500 Index from 2012 to July 2018.First,a combination of 70 factors including valuation factors,leverage factors etc.was established with financial knowledge.Secondly,using weighted least squares regression to initially select 43 effective factors,and then select 29 factors as the training features of the later model through correlation de-redundancy;then after a series of data preprocessing and time-series cross-validation tuning parameters,Rolling training and backtesting on 5 machine learning models using 6-month and 24-month data respectively;and then stacking the selected two base models to construct the final Stacking integrated learning stock selection model;Finally a series of optimizations have been made to the Stacking strategy,which has greatly improved the performance of the strategy.It is found that the two most suitable base models for fusion are a random forest model trained on 6 months of data and a support vector machine model trained on 24 months of data.Secondly,after performing Stacking fusion,it was found that the backtesting performance of the Stacking model was indeed better,and the model fusion was effective,achieving an annualized income of 32.85%,and a Sharpe ratio of 0.967.Compared with the base model,the annualized returns increased by 5.2% and 1.7%,respectively,and the Sharpe ratio increased by 0.186 and 0.062,respectively.Finally,after the number of stocks purchased,weight optimization and maximum retracement,and ATR volatility stop loss,the strategy performance has improved significantly,and it is more reference in real investment.This thesis makes a new attempt on the feasibility of applying machine learning algorithms and model fusion methods to the stock investment market,and designs a complete set of quantitative investment strategies.This strategy has achieved relatively obvious and stable excess returns during the backtest period.It has certain application value in the A-share market and has certain inspiration for investors.