- 學位論文
傳統BI訊號結合機器學習應用於交易策略之研究－以台灣上市股票為例
Traditional BI signals combined with machine learning for trading strategy: A study based on Taiwan listed stocks.
摘要
科技的快速發展帶來了計算能力的巨大飛躍，為各個行業提供了更便利的問題解決方式。在金融領域中，預測股票價格趨勢以獲取超額報酬對投資人具有極大吸引力。本研究結合傳統的技術分析指標（BI訊號）和機器學習演算法，將兩種指標作為特徵變數進行研究。過去的研究通常將這兩種指標分開應用於預測，本文則將二者結合起來，並實證其在交易策略中的可行性。本研究的回測時間範圍為2016年1月1日至2022年12月31日，使用的資料為台灣上市股票的月度數據。研究中運用了決策樹（Decision Tree）、支持向量機（SVM）、多層感知器神經網絡（MLP）等演算法來訓練模型並進行預測。特徵變數包括技術面、籌碼面和基本面等共計100種變數。最終，我們以累積報酬率和準確率作為評估指標，比較了傳統BI訊號結合機器學習演算法與單獨使用傳統BI訊號在預測能力上的差異。根據交易策略評估的結果，我們得出以下結論：1.在本研究中使用的三種演算法中，33種投資策略在累積報酬率方面均優於同期大盤加權指數；2.在大部分情況下，移動平均線（MA）、布林通道（BBANDS）和投信連續3日買超等策略表現稍遜，但在大多數時間內優於同期加權指數；3.傳統BI訊號結合機器學習演算法中的決策樹在預測結果上表現最為優異；4. 本研究中的股價淨值比策略在各個衡量評估結果上均表現最佳。總體而言，傳統BI訊號結合機器學習演算法在股價淨值比策略上對股票長期趨勢具有更佳的預測能力，然而對於其他策略仍存在改進的空間。
並列摘要
The rapid development of technology has brought about a tremendous leap in computational power, providing more convenient ways to solve problems across various industries. In the financial sector, predicting stock price trends to gain excess returns is highly appealing to investors. This study combines traditional technical analysis indicators (BI signals) with machine learning algorithms, using both indicators as feature variables for analysis. Previous research has typically applied these two indicators separately for prediction, whereas this paper combines them and empirically tests their feasibility in trading strategies. The backtesting period for this study ranges from January 1, 2016, to December 31, 2022, using monthly data from Taiwan's listed stocks. Decision tree, support vector machine (SVM), and multi-layer perceptron neural network (MLP) algorithms are employed to train models and make predictions. The feature variables include technical factors, chip factors, fundamental factors, and a total of 100 variables. Finally, cumulative return rate and accuracy are used as evaluation metrics to compare the predictive capabilities of combining traditional BI signals with machine learning algorithms versus using traditional BI signals alone. Based on the results of the trading strategy evaluation, the following conclusions are drawn:1.Among the three algorithms used in this study, 33 investment strategies outperformed the benchmark Taiwan Weighted Index in terms of cumulative return rate. 2.In most cases, strategies such as moving averages (MA), Bollinger Bands (B.BANDS), and three consecutive days of net buying by institutional investors performed slightly inferior, but generally better than the benchmark index. 3.The decision tree algorithm in the combination of traditional BI signals and machine learning algorithms exhibited the best performance in terms of prediction results. 4.The price-to-book ratio strategy employed in this study performed the best across various evaluation measures. Overall, the combination of traditional BI signals with machine learning algorithms demonstrated better predictive capabilities for long-term stock trends, specifically in the price-to-book ratio strategy. However, there is still room for improvement for other strategies.