ABSTRACT
As described in Part II, several approaches have been proposed to select portfolios from financial markets. The pattern matching-based approach, which is intuitive in nature, can achieve best performance at the present time. However, one key challenge to this approach is to effectively locate a set of trading days whose price relative vectors are similar to the coming one. As detailed in Section 6.1, existing strategies often adopt Euclidean distance to measure the similarity between two preceding market windows. Euclidean distance can somehow measure the similarity; however, it simply considers the neighborhood of the latest market windows and ignores the linear or nonlinear relationship between two market windows, which is important for price relative estimation. In this chapter, we propose to exploit similar patterns via a correlation coefficient, which effectively measures the linear relationship, and further propose a novel pattern matching-based online portfolio selection algorithm “CORrelation-driven Nonparametric learning” (CORN) (Li et al. 2011a). The proposed CORN algorithm can better locate a similarity set, and thus can output portfolios that are more effective than existing pattern matching-based strategies. Moreover, we also proved CORN’s universal consistency,∗ which is a nice property for the pattern matching-based algorithms. Further, in Part IV, we will extensively evaluate the algorithm on several real stock markets, where the encouraging results show that the proposed algorithm can easily beat both market index and best stock substantially (without or with small transaction costs) and also surpass a variety of the state-of-the-art techniques significantly.
