Research Of Portfolio Based On The Deep Reinforcement Learning And Improved Mean-Variance Model

Since Shanghai and Shenzhen Stock Exchanges set up in 1990 and 1991,the size of China's securities market continues to expand.The equity issuance system has been maturing,the supervision system has been strengthened and the capital market has achieved rapid development.Although China's securities market is developing rapidly,but for the majority of investors,it is often difficult to grasp the stock band of Chinese market.From 2013 to 2014,the Growth Enterprise Market Index undergone the bull market,while the Shanghai Stock Index teetered on 2000 points.In 2015,it experienced a huge bull market,but then there were many "thousand shares fall".The market experienced the structural bull market of Blue-chip share,while Growth Enterprise Market Index kept falling.Thousands of investors have become "cut leeks" among the turns of stock wheel movement in Chinese stock market.From the perspective of investors,they often pursue risk maximization while avoiding risks.In order to get certain profits in the market,portfolio management plays a more and more important role.As early as 1952,the famous American economist Markowitz proposed the theory of portfolio management,and established the Mean-Variance model.However,this model solves the problem of single period static portfolio selection.Furthermore,the model only considers the proportion of venture capital investment.And the stock market is changing rapidly.Investors need to adjust the allocation of risk assets and adjust the proportion of risk-free assets according to market conditions.The main contents of the thesis are as follows:(1)For the covariance matrix calculation in the mean variance model,the correlation coefficient is calculated by the static calculation of historical data,and the DCC-GARCH model is used to calculate the time variant correlation of each return sequence.Time varying correlation can get the fluctuation relation of all yield series more quickly and reduce the time lag effect of the traditional mean-variance model.(2)The variance calculation in mean variance model is improved.This thesis argues that measuring volatility is only one aspect of closing price changes.Next,this thesis presents a comprehensive calculation method of variance,which is weighted by the volatility rate of opening price,closing price,the highest price and the lowest price.(3)A deep reinforcement learning model is created to adjust the market timing and mean variance model parameters.The deep reinforcement learning model uses current status information to dynamically select investment positions for the next one month.The position can be divided into three categories,full in stock,half in stock and none in stock.Moreover,we use depth prediction to predict the next position information and dynamically adjust the mean-variance model of each period.This thesis uses the method of monthly dynamic switching when conducts the investment,and uses the data of the ten industry indexes from April 2016 to April 2017 to carry out an empirical test.The model gains 10.68%of the investment income,which is higher than the earnings of the 7.28%of the Shanghai stock index and the mean-variance model of 3.59%.In terms of risk control,the sharp ratio in this thesis is 1.95,which is much higher than the 0.58 of Shanghai stock index and 0.34 of the mean variance model.The model embodies certain practical value.