Comparing Crobex Network Structures Under Different Correlation Filtering Methods

Financial markets are complex systems, and one way to better understand them is to examine how different stocks are connected. This study examines how various methods for filtering correlations between stock returns can influence the structure of financial networks. For the Croatian stock market, three popular filtering approaches are compared: p-value thresholding based on partial correlations, False Discovery Rate (FDR) correction, and shrinkage estimation. Two very different time periods are examined: the 2008 financial crisis and a calmer market phase in 2023. To avoid noise from overall market movements, the common market component is removed and the focus is on residual, or idiosyncratic, returns. This helps reveal more specific connections, less driven by broad trends. The results show that filtering choices matter a great deal: the p-value method does the best job at highlighting meaningful changes between the two periods, showing stronger clustering and structure during the crisis. Shrinkage filtering, on the other hand, often produces dense but harder-to-interpret networks, whereas FDR tends to be overly cautious, especially in smaller markets. These insights are valuable for researchers and practitioners trying to use network analysis to track market behaviour, stress, or structural change over time.