Research On Key Technologies Of Dynamic Fund Portfolio Management Optimization Based On Deep Reinforcement Learning

Dynamic fund portfolio management can fully disperse the risk and effectively improve the stability,so that higher profit is achieved.In existing dynamic management methods,fund portfolio are adjusted based on future price forecasting.However,it is difficult to predict the future trend of fund price,which leads to low stability of fund portfolio.A dynamic portfolio management optimization method,referred to as PPOVAR,based on deep reinforcement learning is proposed in this thesis.Based on trading strategy,the method perceives the dynamic changes of fund portfolio data in the market through deep learning network to extract features.PPO-VAR uses the value at risk(VAR)model to calculate the balance value between return and risk in all adjustment schemes.The highquality scheme is determined by the balance value.Finally,the optimal adjustment is made by using reinforcement learning to train model so as to enhance the robustness of the model.Specifically,this thesis mainly includes the following three aspects of research content:(1)The funds with both diversity and low correlation are selected to build an investment portfolio.There are many types and huge amounts of the funds in the financial market.There are certain correlations between the funds too.If a portfolio of multiple strongly correlated funds is constructed,then it will increase the systematic risk of the portfolio.While it is difficult to reduce the systematic risk with existing methods.This thesis uses the method of constructing a similarity network between the funds to measure the correlation between the funds.This thesis uses the method of network clustering in order to select funds with both diversity and low correlation to construct an investment portfolio.(2)The initial proportion of the investment portfolio is calculated.Each fund has different financial features such as return and risk in the portfolio.In order to improve the stability of the combination,the best balance between the expected return and the portfolio risk can be achieved by using the method of investing funds in different proportions.Solving the initial proportion of the investment portfolio is a nonlinear problem under the constraints of the actual transaction model,which increases the difficulty to solve the problem.This thesis proposes an artificial bee colony algorithm,called ABC-MOT,based on a hybrid orthogonal table.(3)It is essential to make a dynamic management of the portfolio.The expected return and risk of the investment portfolio will gradually lose balance due to the constant changes in the financial market,which results in a huge loss of the investment portfolio.Therefore,it is necessary to dynamically adjust the investment ratio according to changes in the financial market to maintain the best balance between investment portfolio returns and risks.Hence,the PPO-VAR method based on the trading strategy is proposed,which combines deep reinforcement learning with the value at risk model of financial portfolio theory.This thesis makes an experiment on the historical data of the fund from2013 to 2020.The data from 2013 to 2018 are training dataset,and those from 2018 to 2020 are testing dataset.The return of PPO-VAR method for dynamic portfolio management in the testing dataset is 12.92% and the maximum withdrawal is 9.67% from the experimental results.However,the profit of the control group is only 0.08% and the maximum withdrawal is as high as 18.12%,which proves that the PPO-VAR method can effectively improve the stability of the portfolio.