Predicting stock movements based on financial news with segmentation

Abstract

With the development of machine learning technologies, predicting stock movements by analyzing news articles has been studied actively. Most of the existing studies utilize only the datasets of target companies, and some studies use datasets of the relevant companies in the Global Industry Classification Standard (GICS) sectors. However, we show that GICS has a limitation in finding relevance regarding stock prediction because heterogeneity exists in the GICS sectors. To solve this limitation, we suggest a methodology that reflects heterogeneity and searches for homogeneous groups of companies which have high relevance. Stock price movements are predicted using the K-means clustering and multiple kernel learning technique which integrates information from the target company and its homogeneous cluster. We experiment using three-year data from the Republic of Korea and compare the results of the proposed method with those of existing methods. The results show that the proposed method shows higher predictability than existing methods in the majority of cases. The results also imply that the necessity of cluster analysis depends on the heterogeneity in the sector, and it is essential to perform cluster analysis with a larger number of clusters as the heterogeneity increases.