台股波峰波谷的資料探勘與預測

Data Mining and Forecasting in Peaks and Troughs of Taiwan Stock Market

方永盈(Yung-Ying Fang)

指導教授 ： 李鴻璋

淡江大學/商管學院/資訊管理學系碩士班/碩士(2016年)

https://doi.org/10.6846/TKU.2016.00131

摘要

2014年底，台灣國內上市公司總市值26.25兆元，總成交值為21.90兆元，約為該年國民生產毛額(GDP)的1.6倍及1.3倍，可看出台灣證券市場規模及影響。證券投資具有可適性資金投入、其績效具有高利潤及高風險特點。如何將風險降低，帶來更穩定的獲利，是個重大課題。
　　本文是針對加權股價指數波峰與波谷作研究。本研究分成兩部份，首先定義本論文短期與中期波峰波谷之意義，利用資料探勘技術，統計分析各常用分析指標如成交量、相對強弱指標、股價乖離率、及KD隨機指標等檢視其於加權股價指數波峰與波谷前後變化所表現出的資訊。第二部份則是利用該資料探勘中，所得出的有效的知識，以支援向量機(SVM)預測出未來波峰與波谷發生的時間，並且針對不同條件下的訓練集與測試集做深入探討。
　　本文研究結果顯示，於波峰方面，結果顯示其模型敏感度偏高、較易受市場活絡或環境因素變化所影響。本研究在透過多種技術指標組合及參數值，找尋到其中一對中期波峰中的模型對訓練集其波峰平均交叉驗證正確率達到80%以上，對測試集分類正確率皆達到70%以上，且操作正確率平均達至93.75%。波谷方面透過數據發現其較波峰容易預測，在實驗結果中，不論波谷範圍選擇短期或是中期的，皆有能對應出五年平均達到80%以上的操作正確率之特徵值組合，皆是具有一定程度可信度的數據，證實本研究所使用之特徵值所建立的波峰模型與波谷模型，對投資操作人而言具有其參考價值。

關鍵字

台灣加權股價指數 ； 資料探勘 ； 機器學習 ； 支援向量機 ； 最佳化 ； 操作正確率 ； 預測

並列摘要

The total market value of listed shares in Taiwan was 26.25 trillion and the total turnover volume was 21.90 trillion in year 2014. These was 1.6 times and 1.3 times of Taiwan GDP in that year. That indicated the profound effect of Taiwan stock market. Even with adaptive value, features of investing in stock are still high profit as well as high risk. How to reduce risk and, at the same time, produce high profit is an interesting challenge.
In this paper, we try to build a model which can predict the peaks and troughs of Taiwan stock market. First, we define peaks and troughs in short-term and medium-term period, respectively. And then, by scanning the historical data, we find the indicators like Volume, Relative Strength Index, Bias Ratio of Stock Price, Stochastic Oscillator…etc. to get information from the peaks and troughs of Taiwan stock market. After that, we use SVM along with some notable features to forecast peaks and troughs of Taiwan Stock Market. Within that evaluation, two kinds of accuracy, the general accuracy (correct or incorrect) and operation accuracy (in the range of indicating or not), is defined. Finally, we discuss parmeters and mark criterions about training sets and test sets in varities of condition.
The result shows that prediction accuracy for peaks is highly sensitive and it’s easily effected by varieties of environment. Among these, we present a model for medium-term peaks which the average Cross-Validation Accuracy for the training set is up to 80%, accuracy for the testing set is up to 70%, and average operation accuracy for test set is 93.75%. On the other side, the result shows that models for troughs more easily to predict than that for peaks. For both troughs of short-term and medium-term, they can be found a model with combined features that produce a average operation accuracy for testing set up to 80%. These promising results show that the proposed model can provide a significant value for the investors.

並列關鍵字

Taiwan Stock Exchange ； Data mining ； Machine learning ； Support Vector Machine ； Optimization ； Operation Accuracy ； Forecasting

參考文獻

[20] 李鴻璋, & 黃紹輔. (2013). 應用三重濾網下之技術指標探討台灣股市交易策略, 淡江大學資訊管理學系碩士論文。

[28] 雷祖強, 周天穎, 萬絢, 楊龍士, & 許晉嘉. (2007). 空間特徵分類器支援向量機之研究. 航測及遙測學刊, 12(2), 145-163.

[2] Ball, R., & Brown, P. (1968). An empirical evaluation of accounting income numbers. Journal of accounting research, 159-178.

[3] Beaver, W. H., McAnally, M. L., & Stinson, C. H. (1997). The information content of earnings and prices: A simultaneous equations approach. Journal of Accounting and Economics, 23(1), 53-81.

[4] Brock, W., Lakonishok, J., & LeBaron, B. (1992). Simple technical trading rules and the stochastic properties of stock returns. The Journal of Finance, 47(5), 1731-1764.