Training an ETF Trading Agent with Deep Q-Learning
| Autor: | HONG, SHAO-YAN, 洪紹晏 |
|---|---|
| Rok vydání: | 2019 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 107 AlphaGo recently received a lot of attention after beating professional Go players. Thanks to AlphaGo, deep reinforcement learning is now well known to the academia. Consequently, the deep Q-learning, one of the reforcement learning methods, has been applied to more and more problems. As reinforcement learning trains the agent to maximize accumulated rewards, the agent may act as a successful trader (speculator). In this thesis, we study how to apply reinforcement learning to train an agent for trading ETFs. To use the reinforcement learning, we need to determine states, actions, and rewards in the problem. In our setting, the agent takes the prices as states to make decisions, such as buy, sell or no action, and use portfolio returns as rewards. In this work, the agent performs one action per five days based on the closing prices (and others) of the prior five days of the ETF to be traded. The agent is trained with 4 years of trading records, and is tested on the interval of one year. The performance of the agent is compared with the buy-and-hold strategy. The experimental results show that the proposed trading agent using double deep Q-learning can beat the buy-and-hold in the bear market. Therefore, this approach is suitable for diverse market situations. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|