Each day, news coverage traces how economic actors are perceived to be connected, offering a real-time map of interconnected vulnerabilities. This paper asks whether the structural position of financial institutions within the network of reported causal relationships encodes uncertaincy and systemic risk information. We introduce the Conditional Expected Centrality (CoEC), defined as the expected change in narrative centrality of financial institutions in daily news networks conditional on market-wide distress. The measure is estimated a DCC-GARCH framework on daily directed causal networks extracted by the Causality Link platform from over 8,000 international publishers from 2015 to 2025. At the financial system level, principal components combining coverage intensity and topological position centralities yield positive and statistically significant out-of-sample R2 values over horizons of up to 15 trading days for prediction of the upper tail of macroeconomic uncertainty measures. At the bank level, eigenvector centrality-based CoEC leads market-based marginal expected shortfall. Across both levels, we show that it is the structural reorganisation of the narrative network, not only the volume of coverage that carries forward-looking information, pointing to network-based media monitoring as a tractable input for macroprudential surveillance.
This work leverages a unique database that identifies written causal statements between economic indicators and events with daily precision, from a total of 128 million documents processed since 2014, covering over 40,000 firms across 187 exchanges worldwide. This database offers a detailed view of causal structures within written media content, allowing a detailed analysis of its impact on the world’s economy. Our study models these interactions as an evolving dynamic network, represented by daily weighted directed graphs. We use and adapt the preferential attachment (PA) model to estimate the probability that new nodes connect to existing ones. PA is an appealing edge-generating mechanism for capturing the growth dynamics of evolving networks and for modelling the tail-heaviness of the distributions of node degrees and strengths. Model parameters are estimated using both parametric and nonparametric methods, which are then used to efficiently generate future day-ahead weighted and directed PA networks. The predictive accuracy of the proposed approach is assessed by examining how well the generated networks reproduce the daily characteristics and empirical linking probabilities observed in the data. This methodology enhances our understanding of media’s influence on the global economy and evaluates the model’s suitability for complex dynamic network analysis.