臺灣博碩士論文加值系統

本研究利用台灣經濟資料庫（Taiwan Economic Journal）的台股上市普通股（不包含F股、TDR台灣存託憑證）調整後股價日資料，將其處理成點矩陣輸入到卷積神經網絡（Convolutional Neural Network ，簡稱CNN）模型中，以預測未來5、20、60日的股價趨勢。樣本期間涵蓋1997年1月1日至2023年12月31日，共27年，並將其分為訓練期（滾動4年）和測試期（接續1年）。CNN模型輸出上漲機率數值，根據該機率將股票分成十分位數，並形成四種交易策略以評估模型的預測效果及交易策略的報酬績效。實證結果顯示，20日模型在訓練期的準確度超過90%，顯著高於5日模型。5日和20日模型在測試期的準確度均在50%至60%之間。預測未來5日的準確度最高，其次是20日和60日，而預測未來60日的精確度最高，其次是20日和5日。
在十分位數版本的交易策略方面，研究發現四種交易策略在扣除交易成本後的預測未來60日之23年平均年化報酬均呈現正報酬。其中，預測未來60日的報酬表現優於預測未來20日，預測未來20日又優於預測未來5日。台股上市股票的報酬累積並未發生在第一週，而是隨著時間增加而逐漸累積。多空策略在平均持有期間報酬、年化報酬優於買進策略。多空策略的波動度遠遠小於純買進策略的波動度；在四種交易策略中，等權重多空策略投資組合的表現較好。
若改成只買進閥值0.9以上以及同時放空閥值0.1以下的股票，所產生之交易策略，研究發現四種交易策略在扣除交易成本後的預測未來60日之23年平均年化報酬能呈現正報酬。四種交易策略的年化報酬及波動度表現各有勝負，沒有明顯優劣之分。設定閥值為0.9所產生的交易策略能有效改善報酬績效表現，但同時也需承受投組波動度增加的風險。

This study utilizes the daily adjusted stock price data of listed common stocks on the Taiwan Stock Exchange (excluding F shares and TDR) from the TEJ database. The data is processed into point matrices and input into a Convolutional Neural Network（CNN） model to predict stock price trends for the next 5, 20, and 60 days. The sample period covers from January 1, 1997, to December 31, 2023, totaling 27 years, and is divided into training periods (rolling 4 years) and testing periods (the following 1 year). The CNN model outputs the probability of stock price increases, based on which stocks are divided into deciles, forming four trading strategies to evaluate the prediction accuracy and performance of the trading strategies. The empirical results show that the 20-day model''s accuracy exceeds 90% during the training period, significantly higher than the 5-day model; the accuracy of both the 5-day and 20-day models ranges between 50% and 60% during the testing period. The prediction accuracy for the next 5 days is the highest, followed by 20 days and 60 days, while the prediction precision for the next 60 days is the highest, followed by 20 days and 5 days.
In the context of decile-based trading strategies, the study found that four types of trading strategies exhibited positive average annualized returns over 23 years, net of transaction costs, when predicting the next 60 days. Among them, the performance of predicting the next 60 days was better than predicting the next 20 days, which in turn was better than predicting the next 5 days. The cumulative returns of listed stocks in the Taiwan stock market did not occur in the first week but gradually accumulated over time. Long-short strategies outperformed buy-only strategies in terms of average holding period returns and annualized returns. The volatility of long-short strategies was much lower than that of pure buy-only strategies; among the four trading strategies, the equal-weighted long-short strategy portfolio performed better.
If the strategy was changed to only buying stocks with a threshold above 0.9 and simultaneously shorting stocks with a threshold below 0.1, the study found that the four trading strategies still exhibited positive average annualized returns over 23 years, net of transaction costs, when predicting the next 60 days. The annualized returns and volatility performance of the four trading strategies varied, with no clear winner. Setting a threshold of 0.9 for the trading strategy effectively improved performance, but it also required bearing the risk of increased portfolio volatility.

論文審定書 i
誌謝 ii
中文摘要 iii
Abstract iv
第一章 緒論 1
第一節 研究背景 1
第二節 研究動機 2
第三節 研究目的 3
第四節 研究流程 4
第二章 文獻回顧 6
第一節 技術分析相關文獻 6
第二節 卷積神經網絡模型相關文獻 7
第三節 文獻綜述與本研究之貢獻 9
第三章 研究方法 11
第一節 資料來源 11
第二節 研究模型 12
第三節 建立圖像（創建點矩陣） 16
第四節 訓練CNN模型 19
第五節 股票池篩選 23
第六節 交易策略設定 25
第七節 模型及交易策略評估指標 26
第四章 實證結果 29
第一節 模型評估指標 29
第二節 交易策略評估指標（未扣除交易成本） 32
第三節 交易策略評估指標（扣除交易成本後） 38
第四節 交易策略波動度及年化報酬 44
第五節 閥值版交易策略報酬表現（扣除交易成本） 50
第六節 閥值版交易策略波動度及年化報酬 58
第五章 結論與建議 66
第一節 研究結論 66
第二節 未來改進方向及建議 67
參考文獻 69

英文文獻
1.Jiang, J., Kelly, B., & Xiu, D. (2023). (Re‐) Imag (in) ing Price Trends. The Journal of Finance, 78(6), 3193-3249.
2.Taylor, M. P., & Allen, H. (1992). The use of technical analysis in the foreign exchange market. Journal of international Money and Finance, 11(3), 304-314.
3.Bessembinder, H., & Chan, K. (1995). The profitability of technical trading rules in the Asian stock markets. Pacific-basin finance journal, 3(2-3), 257-284.
4.Hsu, P. H., Hsu, Y. C., & Kuan, C. M. (2010). Testing the predictive ability of technical analysis using a new stepwise test without data snooping bias. Journal of Empirical Finance, 17(3), 471-484.
5.Park, C. H., & Irwin, S. H. (2007). What do we know about the profitability of technical analysis?. Journal of Economic surveys, 21(4), 786-826.
6.Lee, M. C., Chang, J. W., Hung, J. C., & Chen, B. L. (2021). Exploring the effectiveness of deep neural networks with technical analysis applied to stock market prediction. Computer Science and Information Systems, 18(2), 401-418.
7.Gudelek, M. U., Boluk, S. A., & Ozbayoglu, A. M. (2017, November). A deep learning based stock trading model with 2-D CNN trend detection. In 2017 IEEE symposium series on computational intelligence (SSCI) (pp. 1-8). IEEE.
8.Chen, J. F., Chen, W. L., Huang, C. P., Huang, S. H., & Chen, A. P. (2016, November). Financial time-series data analysis using deep convolutional neural networks. In 2016 7th International conference on cloud computing and big data (CCBD) (pp. 87-92). IEEE.
9.Chen, Y., Fang, R., Liang, T., Sha, Z., Li, S., Yi, Y., ... & Song, H. (2021). Stock Price Forecast Based on CNN‐BiLSTM‐ECA Model. Scientific Programming, 2021(1), 2446543.
10.Peng, Y., Albuquerque, P. H. M., Kimura, H., & Saavedra, C. A. P. B. (2021). Feature selection and deep neural networks for stock price direction forecasting using technical analysis indicators. Machine Learning with Applications, 5, 100060.
11.Selvin, S., Vinayakumar, R., Gopalakrishnan, E. A., Menon, V. K., & Soman, K. P. (2017, September). Stock price prediction using LSTM, RNN and CNN-sliding window model. In 2017 international conference on advances in computing, communications and informatics (icacci) (pp. 1643-1647). IEEE.
12.Maas, A. L., Hannun, A. Y., & Ng, A. Y. (2013, June). Rectifier nonlinearities improve neural network acoustic models. In Proc. icml (Vol. 30, No. 1, p. 3).
13.Gu, S., Kelly, B., & Xiu, D. (2020). Empirical asset pricing via machine learning. The Review of Financial Studies, 33(5), 2223-2273.
14.Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
15.Ioffe, S., & Szegedy, C. (2015, June). Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning (pp. 448-456). pmlr.
16.Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1), 1929-1958.

書籍雜誌
1. 溫政堯 (譯)。施威銘研究室 (監修) (2021)。自學機器學習：上Kaggle接軌世界，成為資料科學家。旗標出版社。(チーム・カルポ, 2020)
2.François Chollet with J. J. Allaire. (2018). Deep Learning with R. Shelter Island, NY: Manning Publications Co.