Description

Conditional Value-at-Risk (CoVaR) is a widely used metric for quantifying systemic financial risk, capturing the expected loss of one asset conditional on another experiencing significant distress. We propose a novel Transformer-based methodology that integrates financial news articles directly with time-series market data into a unified modeling framework. Unlike approaches that utilize predefined sentiment scores, our method incorporates raw text embeddings extracted from a large language model (LLM), enabling the system to explore the impact of news content on systemic risk. The prediction error bound suggests that it learns even on smaller datasets. We demonstrate our method using US market returns and Reuters financial news from 2006 to 2013, covering both the 2007–2009 financial crisis and the 2011 debt ceiling crisis. Our out-of-sample analysis highlights substantial improvements in CoVaR forecasts when incorporating news data, underscoring the value of textual information in systemic risk modeling.