Research On Online Portfolio Model Based On Multi-armed Bandit

With the development of machine learning technology,quantitative investment has increasingly attracted widespread attention in academia and the financial industry.Among them,the multi-armed bandit model in the field of reinforcement learning is used in the online portfolio model because of its advantages in processing serialized data.However,the existence of problems such as the defective design of the arm and the insufficient ability to perceive the context information of the model seriously restrict the profitability and anti-risk ability of the investment model.Aiming at the problems that the cardinality constraints cannot be solved in the existing arm design methods,and the exploration cost is too high due to the excessive number of arms,this paper proposes an arm design method based on fuzzy clustering.In this method,the number of cluster centers is firstly determined by calculating the sum of squares of clustering errors,and the elbow method is used to determine the number of cluster centers;Secondly,fuzzy-C means(FCM)algorithm was used to calculate the clustering center location and membership matrix.Based on membership matrix,a portfolio satisfying cardinality constraint is constructed as an arm.To solve the problem that most of the existing online portfolio models only describe the market information from the perspective of stock price,this paper introduces factors such as investment value and market sentiment as the context information of the model to improve the perception of the model.Mean of Rank IC sequence and other indicators are used to verify the validity of the factors,and Exp4.P algorithm was used to solve the multi-arm bandit problem,and the data of effective factors is used to train the expert models.In this paper,the linear regression model is used as the expert model to predict the income of every arm,and we design the expert suggestion vector according to the prediction results.Based on the work in arm design and expert advice design,this paper proposes an online portfolio model based on context information.In order to verify the validity of the proposed model,comparative experiments with OBP and FF models were carried out on the self-selected dataset and two public datasets.The experimental results show that in the three data sets,the average annualized cumulative return of the proposed model reaches 1.4 and Sharpe Ratio reaches 0.6,which are higher than the value of the mainstream online portfolio models,indicating that the overall performance of this model is better than other similar models in terms of profitability and risk resistance.In addition,the validity of the proposed arm design method based on fuzzy clustering is verified in this paper.Experimental results show that the proposed arm design method can effectively improve the profitability of the model while controlling the risk.The online investment portfolio model based on contextual information proposed in this paper has certain advantages in profitability and anti-risk ability,and it can better adapt to changes in the market environment,and provide investors with investment decision-making reference.