基於近鄰傳播分群演算法之新型態投資組合風險分散策略

Abstract

本研究以S&P 500指數成分股探討如何使投資組合波動率標準差最小化，並以Affinity Propagation為本論文主要的分群演算法做探討。本論文有別於傳統以財報相關資訊分析，著重於證券的內在價值，而改以人工智慧的分群演算法，找尋能有效分群的特徵因子模擬相似度做為輸入，本論文將特徵因子分為四大類，分別為「時間序列共變異數」、「技術指標」、「財報相關資訊」、「雙值收益率」嘗試以不同特徵值模擬成相似度輸入做分群，試圖找出能使波動率標準差最小化的分群特徵因子。 實證研究得出實驗組皆顯著優於對照組隨機挑選，其中在單ㄧ特徵值的部份，「財報相關資訊」的效果明顯優於其它兩組，由此可知，公司財報相關資訊即使透過人工智慧分群演算法，對於投資組合的選股仍存在有效的資訊，而在「雙值收益率」組可以發現，同時採計兩個特徵值的效果優於單一特徵值，由此可知，提供較多資訊到Affinity Propagation中，能夠有效提升分群效果。

In this paper, an intelligent portfolio selection method based on Affinity Propagation clustering algorithm is proposed to solve the stable investment problem. The goal of this work is to minimize the volatility of the selected portfolio from the component stocks of S&P 500 index. Each independent stock can be viewed as a node in graph, and the similarity measurements between companies are calculated as the edge weights. Affinity Propagation clustering algorithm solve the graph theory problem by repeatedly update responsibility and availability message passing matrices. This research tried to find most representative and discriminant features to model the stock similarity. The testing features are divided into four major categories, including time-series covariance, technical indicators, previous return information, paired return value. The historical price and trading volume data is used to simulate the portfolio selection and volatility measurement. After grouping these investment targets into a small set of clusters, the selection process will choose fixed number of stocks from different clusters to form the portfolio. The experimental results show that the proposed system can effectively generate more stable portfolio by Affinity Propagation clustering algorithm with proper similarity features than average cases with similar settings.