# A Century of Inflation Narratives | Hoover Institution Source: https://www.youtube.com/watch?v=pbt8EBFYmN0 Recap page: https://rapidrecap.app/video/pbt8EBFYmN0 Generated: 2026-03-12T07:33:10.929+00:00 --- ## Quick Overview Alexandra Zante's team defines an inflation narrative strictly as an explanation of the causes and/or effects of inflation, distinguishing it from mere description or news, and they operationalize this definition using trained large language models (LLMs) on a century of US newspaper text (1923-2025) to track the evolution and predictive power of these narratives. **Key Points:** - The research team defines an inflation narrative as a causal explanation of events, specifically focusing on the causes and/or effects of inflation, differentiating it from literary stories or simple descriptions. - The methodology involves training a Llama 3.1 LLM on 2,000 manually annotated sentences from historical (1960-1980) and contemporary (2010-2023) periods to classify narratives in 4.2 million sentences from US newspapers (1923-2025). - Descriptive analysis shows that pre-1980 narratives emphasized fiscal causes, whereas post-Volcker, supply and monetary causes gained prominence in media discussions. - Newspaper narratives headquartered in Democratic-leaning areas focus on monetary causes and savings effects, while Republican-leaning areas focus on fiscal causes and cost of living effects. - Inflation narratives show strong predictive power for household inflation expectations, especially for lower-income households, even when controlling for past realized inflation and news volume. - Federal Reserve public communications tend to attenuate media narratives on average, although monetary narratives pass through most strongly; however, media narratives directly predict household expectations slightly more strongly than direct Fed communications. - The researchers specifically track causes like fiscal, monetary, and supply shocks, and effects like interest rates, cost of living, and savings impacts, noting that 'rates' discussions pick up significantly post-1980. **Context:** Alexandra Zante presented early-stage, highly interdisciplinary work involving computer scientists, computational linguists, and economists to analyze a century of inflation narratives found in US newspaper archives spanning 1923 to 2025. The core motivation stems from growing literature, including the work of Robert Shiller, suggesting narratives influence beliefs and policy, requiring a precise definition to operationalize textual data analysis. ## Detailed Analysis The research defines an inflation narrative precisely as an explanation of the causes and/or effects of inflation, contrasting this with the vaguer 'story' concept used by Shiller, and grounding it in the literary concept of temporal accounts of events formalized mathematically as causal explanations (A causes B causes C). The team uses a top-down approach, pre-specifying an ontology of causes (e.g., demand, supply, fiscal, monetary, expectations) and effects (e.g., reduced purchasing power, cost of living, interest rates). To extract these narratives from 4.2 million sentences in the ProQuest database, they trained a Llama 3.1 LLM after manually annotating 2,000 sentences, achieving about 80% accuracy on the test set for narrative detection and classification. Descriptive results indicate a shift in media focus from fiscal causes pre-1980 to monetary and supply causes afterward, and a spatial split where Democratic-leaning papers emphasize monetary causes while Republican-leaning papers focus on fiscal causes. Furthermore, the narratives exhibited predictive power for household inflation expectations, particularly among lower-income groups. Analysis of Federal Reserve communications showed that while the Fed focuses on monetary causes, the media tends to attenuate Fed narratives, though monetary framing passes through strongly. The discussion acknowledged econometric challenges, including non-random measurement error from the LLM classification and endogeneity between media and Fed communications, noting that while correlation is established, causal inference requires further structural modeling. ### Research Project Overview - Highly interdisciplinary team including PhD students and faculty from computer science, computational linguistics, and economics - Work is early stage, soliciting feedback on methodology and approach - Focus is on operationalizing the concept of a narrative distinct from mere description. ### Defining Inflation Narratives - Narrative defined as a causal explanation of events, focusing on 'what causes what and what are the consequences' - Strict definition: 'an explanation of the causes and/or effects of inflation' - Distinction made between narratives and tautological statements like 'inflation causes the cost of living to go up'. ### Methodology - Extracted narratives from US newspapers (1923-2025) using LLMs (Llama 3.1) trained on human-annotated data (2,000 sentences) - The process is a classification exercise to detect narrative presence and assign it to pre-specified causal categories (ontology) - Human annotators agreed on narrative presence 90% of the time, and on category 68% of the time. ### Evolution of Media Narratives - Pre-1980 media emphasized fiscal causes of inflation - Post-Volcker era saw prominence shift to supply and monetary causes - Democratic-leaning newspaper areas focus on monetary causes/savings effects; Republican-leaning areas focus on fiscal causes/cost of living effects. ### Predictive Power - Inflation narrative shares strongly predict household inflation expectations, controlling for past inflation and news volume - Predictive power is stronger for lower-income households than higher-income households. ### Federal Reserve Narratives vs. Media - Fed communications focus on monetary causes, precedence, and uncertainty - Media attenuates Fed narratives on average, but monetary framing passes through most strongly - Media narratives directly predict household expectations slightly more strongly than direct Fed communications in this observational data.