A Novel Deep Reinforcement Algorithm with Adaptive Sampling Strategy for Continuous Portfolio Optimization

• Main Authors: Miao, Yu-Hsiang, 繆宇翔
• Other Authors: Chen, An-Pin
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/wvv2ff

碩士 === 國立交通大學 === 資訊管理研究所 === 107 === Quantitative trading finds stable and profitable trading strategies by observing historical data through statistics or mathematics methods. With the advancement of technology and the development of computing equipment, many studies that prove that deep reinforcement learning can perform well in quantitative trading without too many assumptions about the financial market, but these studies are still insufficient for the generalization of trading strategies. To strengthen the generalization ability of trading strategy, this study takes the constituents of the Dow Jones Industrial Average as the target and applies the problem of optimizing the portfolio. The goal is to construct a portfolio of five assets from the constituent stocks, and this portfolio could achieve excellent performance through our trading strategy. To optimize the setting of such problems, it is necessary for the agent to simulate and explore the possibilities. But the process of simulating all possibilities requires a lot of calculations and time. Hence, this study proposes a sampling strategy to determine which data is worth learning by observing the learning condition. By applying this strategy, the agent could learn the general trading strategy more effective within a limited period of time. In addition to the sampling strategy, we use adversarial learning during reinforcement’s learning process to enhance the model’s robustness. From the result of the experiment, we could observe that the model with our sampling strategy are better than the random learning strategy. The Sharpe ratio is increased by 6-7% and the profit value has increased by nearly 45%. The outcome of the experiment demonstrates that our proposed learning framework with the sampling strategy is conducive to obtaining reliable trading rules.