Construction of the emotional vulnerability factor based on network and large language models

With the evolution of empirical asset pricing theory and the increasing complexity of markets, the exploration of innovation - driven alternative factors has attracted growing attention. This study integrates network structure analysis methods into textual analysis to innovatively expand the concept of financial system vulnerability, introducing the notion of emotional vulnerability. This study employed the Moka massive mixed embedding (M3E) model to embed textual sentiment data to construct individual stock sentiment networks. The Ricci curvature of these networks was calculated to quantify emotional vulnerability factors, which were effectively demonstrated to account for excess weekly returns through rigorous validation using portfolio permutation tests, factor redundancy checks, and time series regression analyses. These factors maintain significant explanatory power even when adjusted for the Fama–French five factors. Emotional vulnerability factors enhanced asset risk assessment accuracy, improved market trend predictions, and increased investment decision-making precision and efficiency.