臺灣博碩士論文加值系統

本研究以臺灣50（ETF50）「指數股票型證券投資信託基金」的股票指數作為預測目標。運用深度學習中的長短期記憶模型（Long Short-Term Memory, LSTM）進行研究，將臺灣50指數和其成分股占比最大股票之歷史資料及技術指標資料做為模型的輸入變數。藉由訓練模型預測未來的價格趨勢，並經由模擬交易確立何種交易策略獲利能力較佳。而在策略中，買賣點的價格相對重要。投資人在決策時會希望盡可能地買在低點賣在高點，因此給予投資人作為成交價參考的定價方法中，我們提出「價格區間修正法」跳脫以往常用的「預測收盤價」，進而創造更多獲利空間。此外，考量模型預測誤差會間接影響到交易策略判斷買賣點的時機，本研究也提出「誤差均值移動視窗校正法」，透過實驗找到最適合的校正天數及閥值。本研究將收集到的樣本切割成兩部份，3454筆日資料為訓練資料；384筆日資料為測試資料。經實驗發現：一、本研究提出的「價格區間修正法」獲得4%的報酬率，其結果相當接近以實際最高及最低價進行交易的理想報酬4.03%，有效達成買低賣高之目的；二、本研究加入校正策略以間接修正交易訊號，報酬率的表現由校正前的4%提升至4.55%。

This study uses Taiwan Top50 Exchange Tracker Fund (ETF50) as a forecast target. Using Long Short-Term Memory (LSTM) model in deep learning, the historical data and technical indicators of ETF50 and the largest share of its constituent stocks as input variables of the model. By training the model to predict the future price trend, and through the simulated transaction to establish which trading strategy has the best profitability. And in the strategy, the price of buying and selling points is relatively important. Investors will want to buy as low as possible and sell as high as possible when making decisions. Therefore, in the pricing method given to investors as a reference for the transaction price, we propose the "price range correction method" to get rid of the "predicted closing price" used in the past, thereby creating more profitable space. In addition, considering that the model prediction error will indirectly affect the timing of the trading strategy to determine the trading point, we also proposes the "mean error sliding window correction method", through experiments to find the most suitable correction days and threshold. Our sample data is separated into two parts, 3454 records of training data and 384 records of testing data. After the experiment, it was found that: (1) The "price range correction method" proposed in this research obtains a 4% return rate, and the result is quite close to the ideal return of 4.03% for trading at the actual highest and lowest prices, effectively achieving the purpose of buying low and selling high; (2) After adding the correction strategy, the performance of the trading signal was corrected indirectly, increasing from 4% before correction to 4.55% to achieve the best return.

口試委員會審定書 #
中文摘要 i
ABSTRACT ii
目錄 iii
圖目錄 v
表目錄 vi
第一章 緒論 1
1.1 研究動機 1
1.2 研究目的 2
1.3 論文結構 2
第二章 相關文獻 3
2.1 指數股票型基金（ETF） 3
2.2 股票市場分析 3
2.2.1 基本面分析 3
2.2.2 籌碼面分析 4
2.2.3 技術面分析 4
2.2.4 投資面向比較 7
2.3 優化模型訓練方法 8
2.3.1 資料的正規化 8
2.3.2 Dropout 9
2.4 類神經網路 10
2.4.1 循環神經網路（Recurrent Neural Networks，RNN） 10
2.4.2 長短期記憶模型（Long Short-Term Memory，LSTM） 11
2.5 國、內外相關研究 13
第三章 研究方法 16
3.1 研究流程與架構 16
3.2 資料集 17
3.2.1 輸入變數 18
3.2.2 資料前處理 18
3.3 類神經網路模型 20
3.3.1 類神經網路架構 20
3.3.2 類神經網路參數設定 20
3.3.3 模型評估方式 21
3.3.4 誤差均值移動視窗校正法 21
3.4 買賣決策規則 23
3.4.1 基本假設 23
3.4.2 交易策略 23
3.4.3 價格區間修正法 24
3.4.4 買賣決策規則評估方式 25
第四章 實驗結果與分析 26
4.1 模型訓練及評估 27
4.2 交易策略之比較 28
4.3 定價方法之比較 28
4.3.1 挑選定價方法 28
4.3.2 適當的上下界修正幅度 29
4.4 加入校正策略 31
4.4.1 適當的移動視窗天數與校正閥值 31
4.4.2 校正前後之比較 34
4.5 多頭、空頭區間之獲利表現 35
第五章 結論與未來展望 39
5.1 結論 39
5.2 未來展望 40
參考文獻 41

[1]Deng, L.; Yu, D. Deep Learning: Methods and Applications. Foundations and Trends in Signal Processing. 2014, 7: 3–4.
[2]Nekrasov, V. Knowledge rather than Hope: A Book for Retail Investors and Mathematical Finance Students. 2014, pages 24-26
[3]王慧婷, & 唐麗英. (2007). 以複式模擬法估計股票買賣價格之最適區間 (Doctoral dissertation).
[4]臺灣證交所–指數股票型基金（ ETF ）簡介，(https://www.twse.com.tw/zh/page/ETF/intro.html)。
[5]富邦投信–ETF基本介紹，(https://websys.fsit.com.tw/FubonETF/Academy/AcademyIntroduction.aspx)。
[6]Andrew W. Lo, Harry Mamaysky, and Jiang Wang, “Foundations of technical analysis: Computation algorithms, statistical inference, and empirical implementation”, Journal of Finance, 2000, Vol. 55, 1705-1770
[7]Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167.
[8]Overfitting. In Wikipedia, from https://en.wikipedia.org/w/index.php?title=Overfitting&oldid=964885382
[9]Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Jmlr.org. [July 26, 2015].
[10]Elman, Jeffrey L. "Finding structure in time." Cognitive science 14.2 (1990): 179-211.
[11]Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), 1735-1780.
[12]Gers, F. A., Schmidhuber, J., & Cummins, F. (1999). Learning to forget: Continual prediction with LSTM.
[13]Phi, M. (2018). "Illustrated Guide to LSTM’s and GRU’s: A step by step explanation." from https://towardsdatascience.com/illustrated-guide-to-lstms-and-gru-s-a-step-by-step-explanation-44e9eb85bf21.
[14]劉昭雨、顏士淨, 卷積神經網路在金融技術指標之應用,2017
[15]Di Persio, L., & Honchar, O. (2016). Artificial neural networks approach to the forecast of stock market price movements. International Journal of Economics and Management Systems, 1.
[16]鄭允中. (2017). 基於長短期記憶遞迴類神經網路之新台幣兌美元匯率預測模型.
[17]Qian, F. and X. Chen. Stock Prediction Based on LSTM under Different Stability. in 2019 IEEE 4th International Conference on Cloud Computing and Big Data Analysis (ICCCBDA). 2019. IEEE.
[18]Chen, K., Y. Zhou, and F. Dai. A LSTM-based method for stock returns prediction: A case study of China stock market. in 2015 IEEE International Conference on Big Data (Big Data). 2015. IEEE.
[19]黃華山; 邱一薰. 類神經網路預測臺灣 50 股價指數之研究. 國立彰化師範大學資訊管理學系研究所學位論文, 2005.
[20]Box, G. E. P. And Jenkins, G. M. (1976). Time Series Analysis, Forecasting and Control, Holden-Day, San Francisco
[21]Bollerslev, T. (1986), ”Generalized Autoregressive Conditional Heteroscedasticity,” Journal of Econometrics, 31, 307-27.
[22]Rosenblatt, F. (1962). Perceptions and the theory of brain mechanisms. Spartan books.
[23]Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1985). Learning internal representations by error propagation (No. ICS-8506). California Univ San Diego La Jolla Inst for Cognitive Science.
[24]Hecht-Nielsen, Robert. "Theory of the backpropagation neural network." Neural networks for perception. Academic Press, 1992. 65-93.
[25]Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton. "Imagenet classification with deep convolutional neural networks." Advances in neural information processing systems. 2012.
[26]Python 3.5.0，(https://www.python.org/downloads/release/python-350/)。
[27]臺灣證券交易所，(https://www.twse.com.tw/zh/)。
[28]Pedregosa, F., Varoquaux, G., Gramfort, A.,Michel, V., Thirion, B., Grisel, O., Blondel,M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D.,Brucher, M., Perrot, M. and Duchesnay, E.,Scikit-learn: Machine Learning in Python, Journal of Machine Learning Research Volume 12, 2011, pp.2825–2830
[29]Keras，https://keras.io/。
[30]Davies, P. C., "Design issues in neural network development", NEUROVEST Journal, 5, pp. 21-25, 1994
[31]Lehmann, E. L.; Casella, George (1998). Theory of Point Estimation (2nd ed.). New York: Springer.