Full metadata record
DC FieldValueLanguage
dc.description.abstract人工智慧技術應用於股票市場已經是必然的趨勢,從許多的文獻以及實務上的應用可以發現傳統統計分析方法無法正確掌握到市場持續不斷變動的特性。本論文研究嘗試從籌碼面的角度切進,探討是否可由籌碼面因子來判斷未來股價之趨勢。 首先,本文採用相關分析(Correlation Analysis)方法,來探討選定之籌碼因素與未來股價的相關程度,實驗發現所選擇的因子雖然在實務界或是過去文獻常被利用來當作選股的先行指標,但大部分卻與未來股價趨勢有著很低的關聯性,甚或毫無相關。進而由這些因素來作股價之迴歸預測分析,根據簡單迴歸分析(Simple Regression Analysis)得到的結果,由這些因子過去的歷史資料並無法得到很好的預測價值。然而,這些因子確實有其價值所在,是否因為傳統統計方法無法瞭解整個環境的變動去作動態學習,導致其靜態地以線性方法作分析,而得到不佳的效果,則是以下的探討。 基本面或是技術面之股市分析已被廣泛地討論,相對地,籌碼面分析較少被研究,然而在台灣市場愈來愈開放、資金愈來愈充裕的情況下,法人對於台灣市場的影響力亦愈來愈重;再者,類神經網路(Neural Network)、模糊邏輯(Fuzzy Logic)、基因演算法(Genetic Algorithms)等人工智慧技術對於挖掘股票知識的研究已有相當充分的貢獻;而學習分類元系統(Learning Classifier System)擁有整合與外界的環境作互動以去學習和基因演算法的功能,為一個以規則為基準(Rule-Based)且在變動不確定的環境下運作良好的系統;因此,本文從學習分類元系統的角度切進,以更貼近人類生活的學習方式去瞭解股票市場脈動,進而計算所得的累積報酬,期望建構一個價值高、可以獲取高額利潤的市場機制。 本研究的目的在於利用學習分類元系統之適應動態環境學習的特性,將股票市場中之籌碼面資料予以模擬,由於統計分析有時無法解釋出影響環境的真正因素,而落入統計陷阱(Statistical Pitfalls),造成預測結果的不準確性,大大影響了投資人的決策;而分類元系統可以將環境中複雜因素考量進去以利於學習之行為模式,此可作為提供決策者做正確抉擇的依據。zh_TW
dc.description.abstractArtificial intelligence technique in stock market fluctuation evaluation has been documented in other papers and applied in many application domains. Based on previous surveys, to accurately handle the uncertainties and variations of the environment is a significant task. The thesis tries to probe into the judgment of future price trend on the basis of the institutional factors. Firstly, the traditional statistical tests are performed to discover whether there exists a relationship between the experimental indicators or not. The correlation analysis is performed on the five indicators as the condition part and future price trend. However, the experiment indicates low or non correlation between the stock price and the institutional factors. On the other hand, using institutional indicators as the simulated factors on the stock market has been proved to be quite persuasive in many documents. Thus, the simple regression analysis is used to perform for the prediction of the future price trend. However, according to the results, the future price trend cannot be accurately predicted based on the historical data. Though the consequences are unsatisfactory, the selected institutional indicators are certainly valuable for the dynamic stock market. The poor outcome has to be understood and discussed for discovering the statistical pitfalls. Therefore, the thesis attempts to apply learning classifier system (LCS), which is intended as a rule-based framework that integrates the concept of genetic algorithms, to learn and interact with the stock market environment. There are a large number of elements affecting the stock environment, selecting the most significant ones is capable of making the best investment strategies and improving the system performance. As mentioned above, the thesis focuses on the institutional indicators for modeling the behaviors in such complex environment to help investors obtaining optimal and satisfactory profits. Surveys on the selected institutional indicators, such as buy/sell of qualified foreign institutional investors (QFII), buy/sell of securities investment trusts, balance of margin purchasing, balance of short selling, and trading volume, show that market prices can be directly influenced by the above-mentioned indexes. The promising results demonstrate that, by implementing the LCS model, the rules that are discovered can be utilized to make investment strategies with progressive benefits. The statistical pitfalls might be occurred due to the incapability of understanding and modeling the uncertainties in such situation. However, the learning classifier system is capable of taking the complicated factors into account for discovering the unknown behaviors and learning the inward knowledge of the environment.en_US
dc.subjectLearning Classifier Systemen_US
dc.subjectInstitutional Analysisen_US
dc.subjectStock Marketen_US
dc.titleApplying Learning Classifier System to Stock Market Based on Institutional Analysisen_US
Appears in Collections:Thesis

Files in This Item:

  1. 452301.pdf