Uncovering financial network structures and connectedness from textual risk disclosures

Abstract Risk assessment is crucial for financial institutions, especially financial holding companies (FHCs), due to the inherent organizational complexity and systemic risks embedded in their interconnected structures. We propose a general and flexible disclosure-to-network framework that integrates topic modeling and network analysis to construct a risk-similarity financial network from firms’ textual risk disclosures. The key idea is to convert unstructured narratives into firm-year risk representations and define inter-firm links based on similarity in disclosed risk exposures, yielding connectedness measures that are comparable across institutions and trackable over time. We employ Sentence Latent Dirichlet Allocation as a robust topic modeling approach on risk disclosure text in Chinese FHC annual reports from 2013 to 2020. We document a sustained rise in interconnectedness with spikes around the 2015–2016 market turmoil and later regulatory tightening. Banks are the most connected entities; subsidiaries are more interconnected than parents; and cross-sector subsidiary linkages are stronger within the same holding group. Higher connectedness is associated with lower profitability and higher bankruptcy risk, highlighting implications for systemic-risk monitoring.