在學術界和金融行業中，制定反映不同經濟狀態的盈利策略已引起了廣泛關注。隨著人工智能的快速發展，近期出現許多新的投資策略，其中Q-learning是一種建立在馬可夫決策過程數學框架上的強化學習方法，可幫助投資者預測市場趨勢並提供當前情況下的最佳策略。在Q-learning中，定義狀態的重要性不言而喻，本研究提出了一種基於無監督學習和I-score的方法，用於定義執行Q-learning時的狀態，根據所定義的狀態便可為每個資產建構一個Q-learning投資策略，並可進而建立一基於多標的資產的投資組合。實證結果表明，所提出的方法產生令人滿意的投資回報。;Formulating profitable strategies for reflecting different economic states has attracted much attention in academics and financial industries. With the rapid development of artificial intelligence, new investment strategies have been created recently. Q-learning is a reinforcement learning method built upon the mathematical framework of Markov decision processes, aiding investors in predicting market trends and providing the best strategies in the current scenario. In Q-learning, the importance of defining states is self-evident. This study proposes an approach based on unsupervised learning and the I-score to define the states when conducting Q-learning. Accordingly, we construct a Q-learning investment strategy for each asset and propose an investment portfolio of multiple stocks. Empirical results demonstrate that the proposed method produces promising investment returns.
