應用資料探勘技術及總體經濟變數以預測台灣加權指數

Apply Data Mining Techniques With Macroeconomics Variables in TAIEX

魏楚蕎(Chu-Ciao Wei)

指導教授 ： 周宗南

朝陽科技大學/管理學院/財務金融系/碩士(2015年)

摘要

經濟過程不斷在熱絡與低迷中循環，而股市也在這之循環中不斷起伏，在台灣，股市的熱絡程度一直有增無減，不管以基本面、技術面、籌碼面來預測的投資者都大有人在，隨著科技的日新月異，投資人取得資訊的管道變得更多更容易。但對於這龐大的資料數據中如何篩選出有助於判斷股市預測的訊息在近幾年來成為一個值得探討的議題，因此希望藉由資料探勘能夠挖掘出有助於我們判斷股市預測的訊息。本研究將觀測期間區分為全部觀測期間、下跌期間、上漲期間、盤整期間並依照有無特徵萃取進行最近鄰近法、倒傳遞類神經、線性迴歸、支援向量機四種資料探勘技術以及灰預測來進行預測。實驗結果顯示，以預測月平均指數來看，以特徵萃取後的資料在盤整期間，以線性迴歸預測可得到較小平方絕對誤差177.693，以未特徵萃取的資料在上漲期間以支援向量機預測可得到較小相對絕對誤差 21.07%，及同樣於上漲期間由未特徵萃取的資料以倒傳遞類神經預測可得到較小相對平方根誤差 21.46%。灰預測部分，在預測月平均指數或預測月底指數時，盤整期間的平方絕對誤差與均方根誤差誤差皆較小，全部期間來預測時則以相對絕對誤差與相對平方根誤差較小，因此當股市處於盤整期間時，使用總體經濟變數來預測台灣股票加權指數較為適合。

關鍵字

特徵萃取 ； 主成份分析 ； 灰預測 ； 資料探勘

並列摘要

Economy experiences boom-and-bust cycle, and stock prices, in the meantime, rise and decline in this business cycle. There has been steadily growing enthusiasm towards stock market trading in Taiwan. Investors can base their prediction of stock prices on the fundamental analysis, the technical analysis, or the central tendency of the stock holders; there is no shortage of followers of any of the above analyses. With the rapid development of technology, investors have more and easier access to information. However, how to select from such huge data useful information so as to decide and judge market prediction has recently become an issue worth exploring. Therefore, I hope to find out information useful for deciding and judging market prediction by using data mining techniques. This paper classifies the observation periods as follows: the whole observation period, the period of declining stock prices, the period of rising stock prices, and the correction period. Depending on the feature extraction, this paper also applies gray prediction, and applies four kinds of data mining techniques to predict, including the methods of nearest neighbor, back-propagation neural network, linear regression, and support vector machine. The results are as follows: to predict the monthly average index, using linear regression to analyze data with feature extraction during the correction period, we get a smaller squared absolute error of 177.693; using support vector machine to analyze data without feature extraction during the period of rising stock prices, we get a smaller relative absolute error of 21.07%; and using back-propagation neural network to analyze data without feature extraction during the period of rising stock prices, we get a smaller root mean squared error of 21.46%. As for grey prediction, whether predicting the monthly average index or month-end index, both the mean-absolute error and root mean squared error are smaller during the correction period, whereas the relative absolute error and root relative squared error are smaller in the whole observation period. In conclusion, during the correction period of stock market, using macroeconomic indicators to predict Taiwan’s weighted security index is more appropriate.

並列關鍵字

Grey Prediction ； Feature Extraction ； Principal Component Analysis ； Data Mining

參考文獻

2.余尚武、盧瑞山、林盈成(2009)，應用動量權重及灰預測傅立葉殘差修正模型於台灣股票市場電子股之投資組合策略，電子商務學報，第11卷(1期)，頁259-286。

4.陳啟斌、連文仁、李昆遠、 裴蕾(2004)，股票投資組合風險衡量模型精確度之評估，立德學報，第1卷(第2期)，頁34-48。

13.鄭美幸、詹志明(2002)，灰色理論與時間序列模型在匯率預測績效上之比較，台灣金融財務季刊，第3卷(第2期)，頁95-104。

1.李仁森(2014)，以灰預測方法應用於本國銀行有關BaselIII、資本適足率、第一類資本比率與普通股權益比率之研究—以某金融控股公司之子銀行為例，修平學報，(第29期)，頁27-49。

2.Fama, Eugene F.,(1970), "Multiperiod consumption-investment decisions, " American Economic Review ,vol.60,pp.163-174.