Identifying stock market communities with time series analysis and spectral clustering

Abstract

Communities within the Philippine Stock Exchange (PSE) were detected via spectral clustering using time series analysis methods to form the appropriate similarity matrix. The Z metric, which combines Pearson's correlation coefficient and Mutual information, was calculated for end-of-day (EOD) stock prices of the 48 companies in the PSE that best explains the variance in price fluctuations in the initial 301-company dataset. This was used to create a similarity matrix used in spectral clustering to form 5 communities, matching the number of industries of the 48 companies. The Adjusted Rand Index (ARI) of 0.0962 tells us that these formed communities are random compared to communities based on predefined industries. Average Z-metric values also reveal that companies within these formed communities had stronger direct relationships and greater shared information with each other compared to companies within industry-based communities. Therefore, the analysis of inter-industry relationships within the stock market is important.