TL;DR
This paper establishes a bidirectional causal link between financial market shocks and semantic shifts in news media, demonstrating that markets and narratives exist in a mutually reinforcing feedback loop. By leveraging LLM embeddings and Granger causality tests, the researchers reveal that news outlets' reactivity and predictive power are significantly shaped by their partisan orientation, especially during "black swan" events like COVID-19.
What Problem Does This Paper Address?
Economists have long debated the "chicken and egg" relationship between market events and the narratives that describe them. Do news stories drive investor behavior and market shifts, or do market fluctuations force journalists to change their tune? This is a problem of two-way endogeneity, making it notoriously difficult to model.
Traditional approaches often rely on simplistic text features like sentiment scores or word counts, which fail to capture the nuanced "semantic shocks" occurring in public discourse. Furthermore, while we know politics colors our worldview, there has been little empirical evidence on how partisan leanings (Left, Center, Right) specifically influence the speed and nature of how news outlets react to—or even precede—financial market shocks.
Methods
The authors propose a novel framework to quantify "textual shocks" and link them to market data:
- Textual Representation: They use a transformer-based LLM (
all-mpnet-base-v2) to generate high-dimensional embeddings of news articles. By focusing on articles containing economic keywords (e.g., inflation, labor), they calculate a daily average embedding for each news outlet. - Semantic Shifts: A "News Feature" is defined as the cosine distance between an outlet's embeddings on consecutive days. This represents a sudden shift in how an outlet is talking about the economy.
- Market Shocks: Using a Structural Vector Autoregression (SVAR) framework, the authors decompose S&P 500 and bond yield data into four orthogonal shocks: Growth, Monetary, Common Premium (inflation-related), and Hedging Premium.
- Causal Inference: They apply Granger causality tests to determine if past values of market shocks predict future semantic shifts, and vice versa. They compare standard Linear Regression against non-linear Kernel Ridge Regression (KRR) to capture complex relationships.
Datasets
- News Dataset: A massive corpus spanning 2000–2022, featuring articles from over 40 major outlets (e.g., The New York Times, Fox News, WSJ). Outlets are categorized as Left, Center, or Right based on AllSides.com labels.
- Market Data: Daily asset-price-derived indicators capturing the four distinct types of economic news mentioned above.
Key Insights and Ideas
- Markets Lead, News Follows: The most dominant relationship is reactive. Market shocks have a significantly greater influence on changes in news text than the other way around. This suggests markets are generally efficient at pricing in information before narratives fully shift.
- The Partisan Reactivity Gap: Left-wing and Centrist outlets are significantly more reactive to market shocks than Right-wing outlets. This difference is most pronounced at short time horizons (1-3 days).
- Delayed Predictive Power: While news is mostly reactive, it does contain predictive signals at longer lags (5-10 days). Interestingly, Right-wing outlets show a unique predictive effect on markets specifically at lags of 5-7 days.
- Inflation Feedback Loops: Shocks related to inflation (Common Premium) are the most likely to trigger feedback loops, where market moves drive news shifts that then further drive market moves. The authors suggest this is because inflation affects everyone, making it a "viral" narrative topic.
- COVID-19 as a Turning Point: During the early stages of the pandemic, traditional market indicators failed to predict future market behavior. However, text-based signals became highly predictive. This highlights the value of language data as an exogenous variable during unexpected global crises.
Key Results
- Non-linearity Matters: The non-linear KRR model identified significantly more causal pairs than the linear model, suggesting the relationship between narratives and markets is complex and non-proportional.
- Quantitative Links: 19.5% of tested pairs showed market shocks causing semantic shifts in news, while 6.34% showed news shifts preceding market shocks.
- Partisan Specifics: Left-wing outlets were particularly predictive of inflation-related shocks ($\omega_{cp}$), with a 10.7% significance rate, compared to lower rates for Centrist and Right-wing outlets.
- Feedback Prevalence: Feedback loops were found in 4.22% of inflation-related shock pairs, compared to almost none for growth shocks.
Limitations and Future Work
The authors acknowledge several constraints:
- Language & Culture: The study is limited to English-language news, which may not generalize to other linguistic or cultural contexts.
- Granularity: The analysis uses daily data; moving to intraday data could reveal even faster transmission of shocks.
- Static Labels: Partisan labels are treated as fixed, though media bias can be fluid over two decades.
- Future Directions: The researchers suggest applying this framework to multilingual contexts and using richer, domain-adapted language models to detect even subtler shifts in economic discourse.
Summary updated