使用增強式學習法建立臺灣股價指數期貨當沖交易策略

Abstract

增強式學習法具有與環境互動及延遲報酬兩大特色，適合應用在決策控制系統的問題上，因此本研究採用增強式學習法來建立臺灣股價指數期貨的當沖交易策略。在系統設計上，我們嘗試了三種不同的狀態定義方式、採用Q-learning及SARSA兩種不同的演算法，另外也針對停損、停利點的設置進行討論。 為檢測其可用性，我們採用2004年1月1日至2008年6月30日之臺灣股價指數期貨歷史資料進行學習訓練及績效檢測。

Learning from interacting with environment and delayed reward are the two most important features of reinforcement learning. Because of these two characteristics, reinforcement learning is suitable for control problems. This thesis adopts reinforcement learning to establish several Taiwan stock index future intra-day trading strategies. We design three different definitions of state and use Q-learning and SARSA to implement reinforcement learning. In addition, we discuss the effect of setting maximum acceptable loss and minimum acceptable profit. To verify the usability of our strategies, we use real historical data for back testing and then examine the performance of the trading strategies.