Research

Running an NLP Model During the Shutdown Data Void

By Giovanni Del Ben, Andrea Ma & William Liang

Pity the Federal Reserve (Fed) and its data-dependent mantra. Although participants can make some general assumptions about economic conditions—inflation remains above its target and isn’t falling while cracks continue to appear in the labor market—investors and the Fed are now starved for data on both sides of its mandate thanks to the partial shutdown of the federal government. In an environment where the Fed is under heightened scrutiny and investors remain fixated on each policy-related nuance, what are investors to do?

One place that we’ve recently returned to is a model that aims to measure the economy’s “exhaust,” or the information emitted in the normal course of daily life. Although this information was not meant to be an economic statistic per se, it can provide pertinent and timely economic signals. To that end, we’ve applied natural language processing (NLP) techniques to large volumes of media text to measure how much attention “unemployment” and “inflation” receive. With that in mind, let’s look at the Fed’s mandate and show how this modern dataset can shed some light on what has become a data void.

As early as the 1940s, policymakers debated how monetary policy should support broader economic goals, with the emphasis shifting over time between employment and inflation control. These debates have consistently acknowledged the inherent tensions in pursuing multiple goals, especially when the objectives conflict. The current economic backdrop puts the Fed in a difficult position: tightening monetary policy to curb inflation poses risks to the job market, while policy easing to support jobs may exacerbate inflationary pressures. Figure 1 shows how U.S. unemployment and inflation have evolved recently.

 

Figure 1: Unemployment Low, but Trending Higher, While Inflation Remains Uncomfortably Above Target

zoom_in
Source: Bloomberg. Data as of Aug 31, 2025.
close
Source: Bloomberg. Data as of Aug 31, 2025.

In the context of this tension amidst a data void, our suite of supervised NLP-based indicators dynamically extracts topics from a wide range of high-quality news sources.1 Unlike traditional NLP analyses that focus on sentiment, our models aim to quantify the attention paid to a particular topic over time.

The methodology employs the transformer architecture, which captures complex relationships between words by representing them as dense numerical vectors, and uses attention mechanisms to assign context-dependent weights to different words in a sequence.2 Leveraging this capability, we scan thousands of articles daily to extract underlying semantic topics.

 

Figure 2: Prominence in News Flow: Unemployment and Inflation

zoom_in
Source: PGIM, Bloomberg Professional Services. Data as of Oct 9, 2025. * The Oct 2025 data point reflects a partial month from Oct 1, 2025 through Oct 9, 2025.
close
Source: PGIM, Bloomberg Professional Services. Data as of Oct 9, 2025. * The Oct 2025 data point reflects a partial month from Oct 1, 2025 through Oct 9, 2025.

The scatter plot in Figure 2 shows the attentiveness scores for both unemployment and inflation from our in-house NLP models.3 Higher scores reflect greater prominence of the topic in publicly available news sources.

It’s evident that from July to September 2025, attention to both topics intensified, with unemployment drawing relatively more focus than inflation. One plausible explanation is that the recent uptick in inflation may be transitory, driven by one-off price effects such as tariff-related increases, rather than persistent pressures. Since the U.S. government shutdown at the beginning of October, while attention scores are less elevated than in September, they remain relatively prominent for both unemployment and inflation. Furthermore, the attention score to unemployment observed in the early October data is among the highest since 2000.

We also note how the attention to unemployment and inflation evolved in earlier periods. As the unemployment rate rose in 2024 from 3.9% in April to 4.2% in July, attention shifted to unemployment over inflation. In contrast, the highest attention scores for inflation appeared in the post-COVID summer of 2022, coinciding with a period when both Core CPI and Core PCE were elevated in the 5%-6.6% range.

In an era where partial shutdowns of the U.S. government are increasingly frequent occurrences, investors are left to scour for alternative data sources. Our NLP-based attentiveness scores offer a timely read on which side of the dual mandate is commanding the conversation, providing context around expectations for market pricing and/or potential policy adjustments.

 

1 Articles supplied via Bloomberg Professional Services; underlying sources include Bloomberg News, The New York Times, other major newspapers, and select web scrapes. https://www.bloomberg.com/professional

2 Vaswani et al. (2017). “Attention Is All You Need.” https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf

3 Each point represents the monthly average of daily scores; coverage spans Jan 2000–Sep 2025. (Internal calculation and methodology to be detailed in forthcoming paper.)

References

Andrea Ma