Portfolio Construction Method Based On Deep Reinforcement Learning Research And Application

In the financial market,there is an increasingly urgent need to build intelligent and efficient investment portfolios,which can help investors predict future market trends,avoid investment risks,and increase investment returns.Traditional investment portfolio construction methods based on data analysis often use simple statistical models,which are difficult to discover market operating rules and perform poorly when processing large amounts of data.However,deep reinforcement learning algorithms have powerful data processing and analysis capabilities,and can adaptively adjust strategies by learning from data,handle nonlinear problems and large-scale data,extract effective information from massive financial data and deal with complex and ever-changing market environments,and provide scientific advice for investment decisions.This article proposes a portfolio construction method based on deep reinforcement learning and designs and implements the corresponding portfolio construction assistant software system.The main content includes the following three aspects:(1)A method for financial factor indicator selection based on multi-objective optimization is proposedTo address the issue of information redundancy among a large number of factors in highdimensional financial data,a multi-objective financial factor selection method is proposed.The multiple objectives refer to predicting asset return levels and asset price directions,which are the two main tasks in the portfolio construction process.Predicting asset return levels can provide an estimate of investment returns,making it easier to select factors that contribute more to the construction of high-yield portfolios.Predicting price directions can help make timely trading decisions when constructing portfolios.As a result,we use these two tasks as target tasks to screen for factors with high correlation with asset returns,which can help construct high-yield investment portfolios.The method primarily consists of two stages.In the first stage,the information gain of each factor with respect to the multi-target task labels is ranked,and factors with low rankings are filtered out.In the second stage,the SFFS algorithm based on multi-objective loss optimization is used to further select the factors that will be used to construct the investment portfolio from the first stage results.Comparative experiments show that our method achieves better predictive performance.(2)A portfolio construction method based on deep reinforcement learning is proposedUsing the Deep Deterministic Policy Gradient(DDPG)algorithm based on deep reinforcement learning,a Parallel Portfolio Feature Extraction Network(PPFNet)is designed as the policy network for portfolio construction.In order to achieve the goal of constructing a high-earning portfolio,the DDPG algorithm updates the parameters of the policy network PPFNet based on the returns corresponding to the decisions,optimizes the reward function with the objective of portfolio returns,and ensures that the objective function is maximized.To address the issues that financial asset prices are non-stationary and that assets have interdependence,Non-stationary Transformers are introduced in PPFNet to solve the representation problem caused by the non-stationarity of financial time series sequences;a Graph Convolutional Network(GCN)is used to extract interdependence features between assets to avoid high-risk situations in the investment portfolio.Finally,the decision is made by integrating the features obtained from these two parts.Experiment’s results show that PPFNet achieves the optimal returns and stability compared to other mainstream portfolio construction methods.(3)A portfolio construction assistant software system is designed and implementedIn order to solve the problem that the model is difficult to be applied by investors,this paper designs and develops a portfolio construction system based on the proposed model,which is developed based on Django framework and mainly contains user management module,data management module,factor management module,portfolio construction module and portfolio backtesting analysis module.Investors can manage and optimize their portfolios based on this system,and analyze portfolio performance through backtesting function.