The International Journal of Engineering and Science (IJES) || Volume || 9 || Issue || 03 || Series I || Pages || PP 36-43| 2020 || ISSN (e): 2319-1813 ISSN (p): 20-24-1805 DOI:10.9790/1813-0903013643 www.theijes.com Page 36 A Portfolio Trading System of Digital Currencies Xudong Sun 1 , Liguo Weng 1 , Min Xia 1 1 Jiangsu Key Laboratory of Big Data Analysis Technology, Nanjing University of Information Science and Technology, Nanjing, China Corresponding Author:Min Xia --------------------------------------------------------ABSTRACT--------------------------------------------------------------- Portfolio management is an effective means of investment, but it is a very difficult task to make a proper portfolio management in virtual currency market. A portfolio management system based on deep-reinforcement learning is proposed to conduct more profitable portfolios in virtual currency market.The prices of virtual currency are driven by many factors(such as investors' confidence, emotions,regulations), but the traditional methods can’t effectively deal with historical data. To solve this problem, a separable convolution is proposed. In a virtual currency market, the price rise only occurs in a flash. Experimental results in virtual currency market show that our model can get more returns. Besides, the higher Sharpe ratio indicates our method is more resistant to long-term risks. KEYWORDS:Portfolio ; Deep-reinforcement learning ; Convolution --------------------------------------------------------------------------------------------------------------------------------------- Date of Submission: 01-03-2020 Date of Acceptance: 16-03-2020 --------------------------------------------------------------------------------------------------------------------------------------- I. INTRODUCTION Portfolio management is a dynamic strategy-decision process, which determines how,when and what to invest. The core of portfolio is resource allocation, determininghow to allocate funds and time. Investors shall focus more on project selection of portfoliosas a result, it demands investors should balance their funds and assets.In order to maximize profits, investors will continue renewing asset flows.During the decision process, all projects will be evaluated, chosen and optimized respectively.Then capital will be distributed to more beneficial projects. The initial portfolio models composed of complete mathematical algorithmsstarted to emerge in 1960s. These idealized mathematical models are now called traditionalor classical methods, which included BCG, the General Electric, and Shell directional policy matrix etc[1].These methods utilized techniques such as linear programming, dynamic programming,and integer programming. Under the constraints of fixed capital, the objective of portfolio wasto optimize the expected return using existing and new asset items.Although these methods were quite attractive from the perspective of theory,these idealized methods could not be applied in practical markets.The main reasons of this phenomena were dependency on market informationand great uncertainty of market environment. Besides, through practitioners' practical use,they found it hard to comprehend and utilize these methods due to their complexity.Although portfolio management is very difficult, it plays a key rolefinancial problems, and hence a great number of portfolioapplications were introduced, for example, Fagereng[2]listed our portfolio choice in real life, Harvey [3] proposed to apply portfolio in insurance and Bertrand [4] optimized a Bayesian decision theoretic framework in portfolio. These methods realized optimalinvestment with the use of combinatorial optimization, which leads to huge computation. In additon,traditional methods were unable to effectively extract features because of their low efficiency in model optimization. In recent years, with the developmentof artificial intelligence technology [5,6], deep learningtheory has been more and more widely used in the field of datafeature extraction [7], and reinforcement learning theory isconsidered as an effective method of investment decision-making[8,9]. Reinforcement learning(RL) is a dynamic learning of an optimal policy and isused to solve decision making problems in a wide range of fields in natural sciences and engineering[10,11]. For some financial and economic issues, reinforcement learning is more suitable [9,12],like option pricing [13],and multi-period portfolio optimization [14,15],which utilized policy search to learn to trade. In handling some special issuessuch as risk management, reinforcement learning would provide better solutions in[9]and Garcia [16] proposed a new way to evaluate risk values based on deep learning.In the case of turbulent market conditions, investors are usually prone to make reckless investment behaviors [17].To relieve this dilemma, the adaptive markets hypothesis appeared in Khuntia's work [18], which may be realized by reinforcement learning.Some methods have been proposed to apply in portfolio management. The Best Stock (Best) is a benchmark widely used in portfolio selection, whose trading strategy is to invest in assets that have the best returns in the past [19].The Uniform Constant Rebalanced Portfolio(CRP) is a more challenging