| 摘要: | 由於金融市場的多變性，金融資料探勘一直是個複雜且吸引人的研究主題。以最佳化技術為基礎，至今有不少最佳化方法被提出探勘不同的股票投資組合。其中，有一方法結合群組遺傳演算法(GGA)並考量投資者主觀與客觀的需求，提出一個群組股票投資組合(GSP)探勘技術。群組股票投資組合的優點在於如果投資者不喜歡某一推薦股票，可選擇同一群組內的其它股票取代。然而，該方法在探勘過程並沒有考慮股價序列。因此，為了提高同一群組中的股票相似程度，本論文考量股價序列，提出兩個可以強化群組相似度的群組股票投資組合演算法。
第一個演算法為符號為基礎的群組投資組合探勘技術，包含：SAX-GSP與ESAX-GSP。其染色體編碼包含三個部份，分別為群組，股票和股票投資組合部分。因股價序列的維度很高，所以必須先降低股價序列維度。藉由兩個知名的技術，符號化聚合近似(SAX)和延伸的符號化聚合近似(ESAX)，將每一個股價序列轉換為符號序列。由設計出序列距離指標(Series distance factor)用來評估同一群組中的符號序列相似程度，且成為適應度函數的一部分。結合了現有方法，產生兩個適應度函數用來探勘群組股票投資組合。之後，利用遺傳運算產生下一代新的染色體。
然而，當股價序列長度不同時，SAX-GSP與ESAX-GSP皆無法用來探勘群組投資組合。故第二個演算法採用重要感知點(PIP)技術來解決這個問題並命名為PIP-GSP。根據預設的重要感知點數量，所有的股價序列可利用重要感知點技術保留相同的重要感知點個數，由此便可以計算股價序列的相似度。染色體的編碼方式是和第一個演算法相同。接著，演算法使用修改過後的序列距離指標與既有的指標產生了兩個適合度函數探勘群組股票投資組合。
最後，透過真實股票資料，本論文進行一系列的實驗分析展示所提演算法的特點，包含：群組投資組合分析、群組中股價序列分析和所提方法與現有方法的投資報酬比較。
Stock portfolio mining is always an attractive research topic and a complicated problem due to the quickly economic changing. Based on optimization techniques, many algorithms were proposed to mine different portfolios. An approach was presented to derive group stock portfolio (GSP) considering investors'' objectives and subjective requests by grouping genetic algorithms. The benefit is investors can replace any stock which they do not like with substitute stocks in groups. However, that approach did not take stock price series of stocks into consideration to mine GSP. In order to increase the degree of similarity of stocks in the same group, this thesis considers stock price series of stocks and proposes two methods for enhancing the similarity of GSP.
In the first algorithm (SAX- and ESAX-GSP), three parts, including grouping, stock and stock portfolio parts, are used to encode a GSP into a chromosome. Because the dimension of the stock price series is high, it should be first reduced. By using the well-known dimension reduction techniques, named symbolic aggregate approximation (SAX) and extended symbolic aggregate approximation (ESAX), each stock price series is transformed to symbolic series. The series distance is then designed to evaluate the similarity of symbolic series in the same group, and used as a part of the fitness function. Combining the previous approach, two fitness functions are utilized to mine a GSP. Genetic operations are then executed to generate new chromosomes for next population.
However, when the lengths of stock price series are different, SAX-GSP and ESAX-GSP cannot be used to derive GSP. To solve this problem, by using perceptually important point (PIP), the second algorithm, named PIP-GSP, is proposed to find a solution. According to the given desired number of important points, all series can keep the same important points by PIP, which can then be used to calculate the similarity of series. The chromosome representation is the same as that in the first algorithm. Using the modified series distance factor, the two fitness functions are also used to derive a GSP.
At last, experiments were conducted on a real dataset to show the merits of the proposed approaches, including the analysis of the derived GSPs, the analysis of stock price series in groups and comparing the proposed approaches with the previous approach in terms of ROI. |
