|摘要: ||金融市場中充斥著眾多的因素會影響投資的獲利程度，故許多探勘投資組合的最佳化方法不斷的被提出。又每個投資者有不同的考量，若只提供單一投資組合是無法滿足投資者的各種需求，所以能夠產生多個組合的群組股票投資組合(Group stock portfolio)探勘技術也接著被提出。然而，在現存方法中並沒有考量群組的多樣性。故為了改善既有方法的群組多樣性，亦即讓群組中的股票產業別可以多樣化，本論文利用群組遺傳演算法提出兩個探勘最大多樣性的群組股票投資組合方法。|
第一個方法中，為了計算每個群組的多樣性，首先在適應度函數裡加入多樣性指標(Diversity factor)來促使每個群組都盡可能有相似數量的股票產業別。接著，為了提高群組股票投資組合的報酬穩定性，根據股票現金股利設計了穩定指標(Stability factor)來保留獲利能力較好的股票並移除風險較高的股票。此外，購買單位平衡(Unit balance)和價格平衡(Price balance)亦被用來增加群組中股票價格與購買單位的相似度。演算法中結合上述的指標，設計了兩個適應度函數來評估染色體並設計合適的基因運算來產生新的染色體，包含，兩階段交配、兩階段突變和反轉運算。
第二個方法則提出了更精密的群組遺傳探勘演算法來避免高風險的股票存活於群組股票投資組合中。為了達到這樣的目標，染色體編碼除了原有的群組、股票和股票投資組合三個部分外，又新增了活躍股票部分(Active stock part)用於將股票分成活躍和非活躍兩類。接著亦修改第一個方法中的投資組合滿意度(Modified portfolio satisfaction)並設計出一般化多樣性指標(Generalized diversity factor)。結合舊有與上述兩新增指標，演算法亦利用兩個適應度函數來評估染色體的優劣並透過三階段的交配、三階段的突變和反轉運算來產生下一世代的染色體。
Since many factors may influence the returns of portfolios, lots of algorithms were proposed for deriving near-optimal stock portfolios. Since providing a stock portfolio to investor may not enough, algorithms for mining a group stock portfolio (GSP) which can generate various stock portfolios were then proposed. However, the diversity of groups, which is an important property of grouping problem, is not considered in that approach. To improve the diversity of group which means number of stock categories in a group should be as more as possible, hence this thesis proposes two approaches for mining a diverse GSP by grouping genetic algorithm.
In the first approach, to measure the diversity of groups, the diversity factor is designed to make the numbers of stock categories as similar as possible between groups, and used as a part of fitness function. Then, to increase the profit of the derived GSP, the stability factor is then designed based on cash dividends to keep good quality companies in groups and remove high risk companies from groups. The unit and price balances are also used to increase the similarity of groups. Combining them, two fitness functions are designed to evaluate chromosomes. Genetic operations, including two-phase crossover, two-phase mutation and inversion, are executed to generate new offspring.
In the second approach, a more sophisticated GGA-based approach for mining diverse GSP is proposed to avoid high risk stocks appear in a GSP. To reach the goal, it uses not only grouping, stock and stock portfolio parts but also active stock part to encode a diverse GSP into a chromosome. The active stock part divides stocks into inactive and active stocks. Then, utilizing the modified portfolio satisfaction, the generalized diversity factor and factors presented in the first approach, two fitness functions are designed to measure the quality of chromosomes. Then, three-phase crossover, three-phase mutation and inversion are executed to generate chromosomes for next population.
Finally, experiments were made on two real financial datasets to show the advantages of the proposed approaches, including the derived GSP analysis, impact of the fitness functions and the ROI of the derived GSPs.