A similarity measurement for time series and its application to the stock market. (15th November 2021)

• Record Type: Journal Article
• Authors: Zhao, Feng; Gao, Yating; Li, Xinning; An, Zhiyong; Ge, Shiyu; Zhang, Caiming
• Is Part Of: Expert systems with applications. Volume 182(2021)
• Publication Date: 2021-11-15
• Subjects: Similarity measurements -- Personalization -- Multi-perspective -- Stock prediction; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2021.115217
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Reflecting the personalization of stock time series by weighting the time series. Utilizing dynamic time warping to cope with time shifts and warpings. Embedding Canberra distance for eliminating the impact of singularities. Abstract: The stock market is a very important financial market, and the prediction of the stock has always been of great interest to many investors. Nowadays, many methods for predicting stocks have been developed and one of the most commonly adopted strategies is to seek similar stocks through historical data to make predictions. The key to this strategy is the construction of a reasonable similarity measurement. In this paper, for accurately describing the similarity between a pair of time series, a novel similarity measurement is proposed, which is named as the dynamic multi-perspective personalized similarity measurement (DMPSM). Specifically, the segmented stock series are weighted according to the principle that the closer to current data, the more weight will be given. Then, Canberra distance is embedded into the dynamic time warping (DTW) to measure the similarity between any pair of time series. By this way, the DMPSM can not only reflect the personalization of stock time series, but also eliminate the impact of singularities and apply to one-to-many matching. To validate the efficiency of DMPSM, experiments utilized 285 stocks from the Shanghai Stock Exchange and the results demonstrated the superiority of the proposed approach over similarity measurements, including Euclidean distance, Canberra distance and DTW.