機器學習與深度學習應用於股市交易策略之研究-以半導體與電子零組件類股為例

Application of machine learning and deep learning to stock market trading strategies - case study on semiconductor and electronic components stocks

陳楷理(KAI LI CHEN)

指導教授 ： 徐鼎欣

國立臺中科技大學/商學院/財務金融研究所碩士班/碩士(2022年)

摘要

本研究使用技術指標作為機器學習與深度學習的特徵變數，實證機器學習在交易策略上的可行性，具體技術指標包含MA、MACD、RSI、KD、MFI、WILLIAMS %R與OBV共7種技術指標，分別使用二種方式建立交易策略，方法一為運用過往文獻交易法則、方法二為機器學習與深度學習。研究資料使用2010年1月1日至2021年3月31日半導體產業與電子零組件類股共計119檔股票的日資料做歷史回測，並在回測前進行傳統技術指標策略產生做多與做空訊號漲跌分佈情形，再行分析交易策略績效後得出以下結論： 1.敘述性統計與相關係數中，半導體中游公司台積電之數據優於電子零組件公司川湖，在兩產業中找到共通點是均價高的股票技術指標變數間的相關性低，受影響程度較小 2.多空預測漲跌機率分佈情形，以看漲做多，看跌做空為例，研究發現半導體產業與電子零組件類股1日皆做空，20、60日做多會較符合預期，5日則視情況。差別在於RSI與RSIOBV兩產業變化不同，半導體產業在加入OBV後並無優化之情況，電子零組件則在加入OBV後略有優化，但程度有限。最後以MFIOBV策略對多空漲跌機率的表現最佳 3.整體的交易績效以過往文獻交易法與機器學習操作的超額報酬皆為負，無法優於買入持有，但在兩產業內皆有少數個股是有能力超越買入持有的。在加入OBV後皆有使超額報酬優化的現象，唯有半導體產業的RSI與RSIOBV並無優化的現象。

整體而言，本研究結果顯示機器學習與深度學習在股票預測上有分析能力，但並無法完全優於過往文獻交易法的策略，整體而言，運用機器與深度學習確實能超越多數過往文獻交易策略。

關鍵字

交易策略 ； 技術分析 ； 機器學習 ； 類神經網路 ； 決策樹 ； 支援向量機

並列摘要

This study utilizes technical analysis as feature variables to investigate the feasibility of machine learning and deep learning on trading strategies. The technical indicators adopted include: MA、MACD、RSI、KD、MFI、WILLIAMS %R and OBV. Two methods are adopted for the construction of trading strategies. First Method is by using the rules put forward by past literature, and the second method is by machine and deep learning.

The data gathered spans from January 1, 2010 to March 31, 2021. with a total of 119 stocks. Before back-test is conducted, first use traditional technical indicator to generate long and short signal distribution, then analyze the performance of such strategy, and the following conclusion can be drawn: 1. Descriptive statistics and correlation coefficients show that,the semiconductor midstream firm TSMC outperforms are better than that of the electronic components firm KING SLIDE, and the correlation is quite low for stocks with higher average stock technical indicator being the commonality of the two industries； 2. This study finds that in order to better fit the expectations, both the semiconductor industry and electronic components stocks need to take short position on the first trading day, and to take long position on the twentieth and sixtieth trading day. The 5th Depends on the situation. The difference is that RSI and RSIOBV are different across the two industries. After the addition of OBV into the trading strategy for the semiconductor industry no sign of optimization is observed. The result for the electronic components industry was slightly optimized after the addition of OBV, yet, there is limited degree of optimization. Finally, the MFIOBV strategy has the best performance; 3. The overall trading performance, in terms of abnormal return, is negative for past literature trading methods and machine learning, and cannot outperform buy-and-hold. However, there are some stocks in both industries are capable of outperforming buy-and-hold. Furthermore, with the addition of OBV improves the excess return, but no optimization is observed for RSI and RSIOBV of the semiconductor industry.

Overall, the results of this study show that machine learning and deep learning do posses certain analytical capabilities in stock forecasting, but not all of them can beat the strategies of the previous literature trading method. Generally, machine learning and deep learning is capable of outperforming trading methods offered by past literature.

並列關鍵字

Trading strategy ； Technical analysis ； Machine learning ； Neural network ； Decision tree ； Support vector machine

參考文獻

英文文獻

Aldin, M. M., Dehnavi, H. D. and Entezari, S. (2012). Evaluating the Employment of Technical Indicators in Predicting Stock Price Index Variations Using Artificial Neural Networks (Case Study: Tehran Stock Exchange), International Journal of Business and Management, 7(15), 24-34.

Bradrania, R., Grant, A., Westerholm, P. J. and Wu, W. (2017). Fool’s mate: What does CHESS tell us about individual investor trading performance?, Accounting Finance, 57(4), 981-1017.

Chande, T. S. (2015). A time price oscillator, Technical Analysis of Stocks and Commodities,13(9), 369-374.

Chen, Y. H. (2018). Two Essays Related to Investments: Performance Evaluation for the Stocks Screened and Stock Price Informativeness in Terms of Technical Indicators, Doctoral thesis, Tamkang University.