Stable Sentiment and Persistent Dynamics in U.S. Economic News over 45 Years

Researchers found that news about the economy has changed significantly over the last 45 years. Instead of quickly bouncing back after a news event, the "mood" of the news now stays positive or negative for much longer periods. This shift likely relates to how modern digital media and social platforms operate.

In the study of complex socio-economic systems, collective emotion is a central driver of attention and behavior. Because we cannot observe "emotion" directly, researchers typically use the tone of mass media—extracted via sentiment analysis—as a proxy. This helps them understand how public belief evolves.

Traditionally, this sentiment was treated as a reactive signal. A news shock occurs, the sentiment spikes, and then the system quickly resets. This is known as "reverting to the mean" (returning to a central average).

However, a new study from Luis E. C. Rocha analyzes 45 years of U.S. economic news. The authors report that while the average balance of positive and negative coverage has remained stable, the temporal organization has fundamentally shifted. News sentiment is no longer a series of quick, daily reactions. It has evolved into a regime of persistent, long-lived episodes.

The breakdown of reactive sentiment

The status quo in sentiment modeling assumes that news acts as a series of discrete, short-lived perturbations (disturbances). Under this assumption, a sentiment shock—such as a sudden report on inflation—should leave a temporary trace. This trace should decay exponentially over time. This is akin to dropping a stone in a pond. Ripples appear, but the water eventually returns to a calm, neutral state.

As shown in the statistical characterization of the index, the sentiment fluctuates around a stable mean.

However, its internal rhythm is changing. The authors argue that treating sentiment as a "short-memory" signal fails to capture modern reality. This signal is one that forgets its past almost immediately.

This failure is critical for modeling. If sentiment shocks leave longer traces than expected, traditional models will fail. Models assuming rapid recovery will consistently miscalculate the duration and impact of economic narratives.

Modeling the rise of persistent narratives

To understand this shift, the authors propose a "minimal endogenous-memory model." They decompose sentiment into two distinct layers: a slow-moving latent component ($x^L_t$) and a fast-moving shock component ($y_t$).

1. The Latent Component: This represents the underlying "macro-narrative." It is the slow, drifting background sentiment providing context for all news. The authors model this using fractional Gaussian noise (a mathematical tool representing processes with long-range dependence). This means the current state is heavily influenced by its distant history.
2. The Shock Component: This represents daily "noise" or immediate reactions to specific events. The authors model this as an AR(1) process (an autoregressive model where the current value depends on the previous one) with a feedback coefficient ($\phi$).

The core discovery lies in how these two components interact. The authors find that as we moved from the "pre-web" era to the "social media" era, the memory of the latent component grew stronger. Simultaneously, the "corrective feedback" (the force pushing the system back to neutral) of the shock component weakened.

In practical terms, daily news events are no longer "corrected" or neutralized by the next day's coverage. Instead, they feed into a self-sustaining loop. This keeps the news stuck in a specific emotional state.

From volatility to bimodality

The evidence for this shift is found in the changing geometry of the sentiment data. The authors use Detrended Fluctuation Analysis (DFA)—a method to quantify how much a signal depends on its past across different timescales. They report a dramatic increase in the Hurst exponent ($H$) during the social media period.

The Hurst exponent is a measure of "memory." An $H$ near 0.5 suggests random noise. An $H$ approaching 1.0 indicates a highly persistent, predictable trend.

The paper finds that at long scales (81 to 1,095 days), the Hurst exponent reaches approximately 0.94 .

This signals extremely strong long-range dependence. This isn't just about things staying the same. It is about a fundamental change in how the system responds to stress.

The authors report several complementary shifts in the data's behavior :

• Declining Volatility: Typical day-to-day swings in sentiment have become smaller. The standard deviation dropped from roughly 0.020 to 0.014.
• Increasing Bimodality: The distribution of sentiment is moving away from a "neutral" center. It now concentrates in two poles: either strongly positive or strongly negative.
• Asymmetric Bursts: The authors find that "bursts" of negative sentiment tend to last longer than positive ones. This creates a lopsided landscape of pessimism.

Limits of the media-era proxy

While the findings are striking, the authors note that this study does not provide a definitive causal map. The division of history into "pre-web," "web," and "social media" periods is descriptive. It is not a proof that social media algorithms are the sole cause of the shift.

The study focuses exclusively on U.S. economic news from 24 newspapers. It does not include broadcast media, non-U.S. sources, or raw social media data. Therefore, the researchers cannot say for certain what drives this change. It could be how journalists write, how algorithms distribute news, or how audiences react.

There is also the possibility of "semantic drift" (changes in the meaning of words over time). While the authors use a specialized lexicon to mitigate this, they cannot fully control for changes in article volume or topic composition.

The verdict: A new requirement for forecasting

If you build models for macroeconomic forecasting or financial risk monitoring, old assumptions are no longer safe. The paper suggests that news sentiment should be modeled as a persistent dynamical state. It is not a series of daily reactions that vanish overnight.

The response time of the news ecosystem has slowed down. Current sentiment values now reflect the accumulated memory of many non-recent shocks. Practitioners must account for this "lagged" reality. Treating a news-driven sentiment spike as a transient error will lead to mistakes. It will result in poor predictions regarding how long an economic narrative will dominate the market.

Figures from the paper