本研究使用三種基於模型的深度強化式學習DQN、Double DQN和Dueling DQN來建構價差交易策略,本研究會選擇開發此類交易策略,主要是因為深度強化式學習的獎勵機制和建構交易策略有很好的對應性且價差交易策略能夠有效的減少市場風險。本研究採用2006/01/01至2018/11/16的台股期貨和摩台期貨進行回測,並設計隨機策略、固定策略當作基準策略,實證結果發現深度強化式學習均可以獲得比基準策略更好的表現,而整體上DQN表現勝過Double DQN和Dueling DQN。但細看可以發現,在不同的回測期間,三種深度強化式學習分別有其表現最好的時候,代表此三種模型分別學到不一樣的規則,此規則在不同的時期有不一樣的適用性。
In this paper, we implement three model-based reinforcement learning algorithms with deep learning, Deep Q-Learning Network (DQN), Double Deep Q-learning Network (Double DQN) and Dueling Deep Q-Learning Network (Dueling DQN) in pair trading strategy. In addition, deep reinforcement learning (DRL) has appealing theoretical properties which are hopefully potential since the reward mechanism in DRL with pair trading rules is able to significantly reduce the market risk. We conduct experiments in TX and TW historical data (2006/01/01-2018/11-16) and design the random strategy and fixed strategy to be the benchmark. The empirical results show that three DRL strategies can achieve better performance than the benchmark strategies overall and DQN is more desirable than Double DQN and Dueling DQN. However, during different back-testing period, we observe that they have the best performance respectively. It means that three models learn different rules separately and the rules have different applicability in different periods.