| Title |
Trading Bot Implementation and Performance Comparison Using DQN and DDQN |
| Authors |
김민태(Min Tae Kim) ; 김병욱(Byung Wook Kim) |
| DOI |
https://doi.org/10.5370/KIEE.2021.70.1.158 |
| Keywords |
Trading bot; Deep Q-Network; Double Deep Q-Network; Reinforcement learning |
| Abstract |
This study addresses a reinforcement learning(RL)-based trading bot that can automatically trade stocks using stock trading data and various indicators. To construct the training data, we used daily basic stock price and technical auxiliary indicator, and input them as state variables to conduct trading actions. By comparing the holding assets and stock prices, a Buy agent that predicts the value of buy/hold behavior or a Sell agent that predicts the value of sell/hold behavior is selected for trading action. For the performance analysis, the returns according to the stock price trend during the training and test period are shown for Deep Q-Network (DQN) and Double DQN (DDQN). Experiments showed that both algorithms for test data showed an increase in assets over the long term and follows the trend of the stock price. While DQN reacts sensitively to changes in stock prices, DDQN reacts relatively monotonically, and thus DDQN can guarantee the increase in assets and stable trading. |