📈 economics

From Headlines to Assembly Lines: Constructing a Manufacturing Economic Policy Uncertainty Index

This paper constructs a novel monthly Manufacturing Economic Policy Uncertainty index for the United States (1985–2025) using a hybrid human-auditing and machine-learning approach, revealing that while industrial and macroeconomic uncertainty trends are similar, manufacturing-specific uncertainty uniquely correlates inversely with consumer confidence and output, offering policymakers a refined tool to assess policy impacts on the production sector.

Original authors: Simran Kahai, Aksh Dantal

Published 2026-07-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Simran Kahai, Aksh Dantal

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the economy as a giant, bustling city. In this city, everyone is constantly listening to the radio, watching the news, and reading the papers to figure out what's going on. Sometimes, the news is clear: "The sun is shining, and the roads are open." But other times, the radio is full of static, and the news anchors are arguing about whether a bridge might collapse tomorrow or if the traffic lights are about to change. This confusion is called uncertainty. When people aren't sure what the government is going to do with taxes, rules, or trade, they get nervous. They might stop building new houses, stop buying new cars, or stop lending money to friends. This nervousness is a real thing that economists study, and they call it Economic Policy Uncertainty (EPU). Think of it like a fog rolling into the city; when the fog is thick, nobody knows where to drive, so traffic slows down. For a long time, economists had a single, giant fog-meter that measured how confused the entire city was. But they realized something important: a bakery might be worried about flour prices, while a car factory is terrified about steel tariffs. The big fog-meter couldn't tell the difference between the baker's worry and the factory's panic. That's where this new research steps in, trying to build a special fog-meter just for the factories.

The researchers in this paper, Simran Kahai and Aksh Dantal, decided to build a brand-new tool called the Manufacturing Economic Policy Uncertainty Index (M_EPU). Their goal was to see if they could measure the "fog" specifically for the manufacturing sector—the places where raw materials like metal and plastic are turned into new products like cars, appliances, and machines. They knew that factories are different from other businesses. They need huge amounts of money to build machines upfront, they rely on long chains of suppliers, and they are very sensitive to government rules. So, the big question was: Does the manufacturing sector get scared by the news in a different way than the rest of the economy?

To answer this, the team created a digital detective system. They started with a massive library of newspaper articles from 1985 to 2025. First, they used a team of human readers to find a few key words that signal "manufacturing," like "plant" or "production." Then, they taught a computer algorithm (a type of machine learning) to read thousands of articles and learn which ones were actually talking about factories and which ones were just using those words by accident. The computer got really good at this, learning to spot the difference with high accuracy. Finally, they combined this new "manufacturing" list of words with the original lists of words used to measure general economic uncertainty (words about the "Economy," "Policy," and "Uncertainty"). If an article mentioned manufacturing and talked about uncertain economic policies, the computer counted it. They did this for every month for 40 years to create their new index.

What did they find? The results suggest that the manufacturing sector does indeed have its own unique personality when it comes to worry. While the new M_EPU index moves somewhat in step with the old, general EPU index (like two friends walking in the same direction), they don't always take the same steps. The paper suggests that the manufacturing index reacts much more sharply to specific events, like the tariff shocks in early 2025, than the general index does. In fact, the new index reached a peak value of 1007.73 in April 2025, while the general index was lower, suggesting factories were far more jittery than the rest of the economy during that time.

The researchers also looked at how these "fear levels" affected real-world behavior. They found that when the manufacturing uncertainty (M_EPU) went up, the actual amount of stuff factories produced went down. This makes sense: if a factory owner is scared of new rules, they might wait to buy new machines. However, the general economic uncertainty index showed a weird, opposite pattern in their data, suggesting a positive link to production, which the authors note is surprising and needs more study. They also found that regular people's confidence in the economy was more affected by the general uncertainty than by the manufacturing-specific news.

In short, this paper suggests that we can't just look at one big number to understand how the economy is feeling. The factory floor has its own specific anxieties that get lost in the noise of the general news. By building this new, specialized index, the authors provide a clearer picture of how policy uncertainty travels from the headlines all the way down to the assembly lines, potentially slowing down production in ways that a general economic report might miss. It's a reminder that while the whole city might be foggy, the factories are often in the thickest part of the mist.

Technical Summary: From Headlines to Assembly Lines: Constructing a Manufacturing Economic Policy Uncertainty Index

Problem Statement
Existing literature on Economic Policy Uncertainty (EPU), primarily established by Baker et al. (2016), relies on aggregated, macro-level indices derived from newspaper word frequencies. While effective for measuring blanket national uncertainty, these indices lack the granularity to capture differential exposure to uncertainty across specific economic sectors. The manufacturing sector, characterized by high upfront capital investment, reliance on complex supply chains, and direct regulatory involvement, faces unique dynamics that may diverge from the broader economy. The paper identifies a gap in the literature: the absence of a sector-specific EPU index capable of isolating manufacturing uncertainty from macroeconomic noise.

Methodology
The authors construct a new time-series index, the Manufacturing Economic Policy Uncertainty (M_EPU), covering January 1985 to December 2025. The methodology employs a mixed-method approach combining human auditing with machine learning to create a sector-specific lexicon.

Lexicon Construction (The M Term Set):
- Seed Lexicon: A human audit of a random sample of U.S. newspaper articles (filtered for English, U.S. publication, and the keyword "manufacturing") generated an initial seed lexicon. The three most frequently proposed keywords were "manufacturing," "production," and "plant."
- Machine Learning Training: A positive training corpus was created using the seed lexicon, while a negative corpus was generated by excluding seed keywords. The dataset underwent rigorous quality control to remove metadata and formatting errors.
- Feature Engineering: Texts were processed using Term Frequency-Inverse Document Frequency (TF-IDF) scoring. An elastic-net logistic regression (combining Lasso and Ridge regularization) was fitted to the TF-IDF matrix to identify terms predictive of manufacturing sentiment. Hyperparameters, including Minimum Document Frequency ( $f_{min}$ ) and Maximum Document Frequency ( $f_{max}$ ), were optimized via grid search to maximize the F1 metric.
- Final Audit: Terms with coefficients exceeding a threshold were subjected to a final human audit by two independent coders. The inter-rater reliability was assessed using the Kuder-Richardson Formula (KR-20). Only terms agreed upon by both auditors were included in the final M Term Set.
Index Calculation:
- An article is coded as M_EPU-reflective if it contains the term "manufacturing," at least one keyword from the new M Term Set, and one term from each of Baker et al.'s (2016) original Economy (E), Policy (P), and Uncertainty (U) Term Sets.
- The index is calculated as the ratio of M_EPU-reflective articles to the total number of published articles in a given month, rescaled to a base-period mean of 100.
Validation and Analysis:
- The constructed index was tested for stationarity using Augmented Dickey-Fuller (ADF) tests.
- Pearson and Spearman correlation coefficients were calculated between M_EPU, the macroeconomic EPU, the Consumer Confidence Index (CCI), and the Manufacturing Industrial Production Index (IPMAN).

Key Results

Model Performance: The machine learning classifier achieved an F1 score of 0.937, with an accuracy of 0.941, demonstrating high precision and recall in distinguishing manufacturing-related articles. The final human audit yielded a KR-20 score of 0.718, confirming internal consistency.
Distinctiveness of M_EPU: While M_EPU and macroeconomic EPU show a moderate positive correlation (0.702), they capture distinct sentiments. M_EPU exhibits sharper spikes during specific industrial shocks, such as the 2025 tariff implementations, compared to the broader EPU.
Correlations with Economic Indicators:
- Consumer Confidence: Both M_EPU and EPU show negative correlations with consumer confidence, though the relationship is stronger for EPU (-0.125) than for M_EPU (-0.093), suggesting consumers are more influenced by holistic uncertainty than sector-specific uncertainty.
- Industrial Production: M_EPU exhibits the theoretically expected inverse relationship with manufacturing production (coefficient: -0.169), aligning with Real Options Theory. Conversely, the macroeconomic EPU shows an unexpected positive correlation with production, a finding the authors note warrants further investigation.
2025 Shock: The index captured a significant spike in uncertainty in early 2025 (peaking at 1007.73 in April), which the authors preliminarily attribute to specific tariff actions (Section 232 tariffs on steel/aluminum and "Liberation Day" tariffs) that directly impacted manufacturing supply chains and the automotive sector.

Significance and Contributions
The paper makes three primary contributions to the academic discourse on economic uncertainty:

Methodological Innovation: It develops a novel framework for measuring sector-level uncertainty by integrating machine learning with human auditing to construct a specialized lexicon, moving beyond the "one-size-fits-all" approach of existing indices.
Empirical Evidence of Sectoral Divergence: It provides evidence that manufacturing-specific uncertainty is distinct from national measures. The M_EPU index reacts differently to economic events and maintains a theoretically consistent inverse relationship with industrial output, which the aggregate EPU fails to capture.
Preliminary Policy Relevance: The study offers contemporaneous evidence that sector-specific indices can detect unique economic dynamics, such as the impact of trade policies on manufacturing. The authors suggest that policymakers should consider M_EPU alongside macro indicators when evaluating policies affecting the industrial sector, as it may provide a more granular view of how uncertainty translates from media narratives to assembly lines.

The authors maintain a modest tone regarding causality, noting that while the correlations are robust, future work is required to establish causal links between M_EPU and specific manufacturing outcomes.

Technical Summary: From Headlines to Assembly Lines: Constructing a Manufacturing Economic Policy Uncertainty Index

More like this