Working Papers
Embedded Files
Inflation Narratives and Expectations (JM Paper, ECB Working Paper No 3158)
Abstract: I study how demand-supply narrative disagreement between general and specialized newspapers can explain households’ absolute gap in inflation expectations with experts. I measure inflation narratives via a Causality Extraction algorithm that can identify causal relationships between events in a text and, hence, extract the perceived triggers of inflation. I then classify inflation narratives into demand and supply narratives based on their focus on demand and supply triggers. I measure narrative disagreement between general and specialized newspapers from their attention difference on demand and supply narratives. The absolute expectation gap widens when narrative disagreement increases, especially for non-college-educated and older households. Unlike specialized newspapers’ narratives, general newspapers’ narratives incorrectly align with experts’ demand-supply views.
Presentations: ifo Dresden Workshop on Macroeconomics and International Finance 2024 · 18th Belgian Financial Research Forum 2024 · 7th Dauphine Finance PhD Workshop · 30th Annual Meeting of the German Finance Association · NETSPAR Pension Day 2024
Posters: CEPR Paris Symposium 2024 · Finalist in the European Central Bank’s Young Economist Prize 2025
Keywords: News media, Natural Language Processing, Dependency parsing, Causality extraction
What Triggers Flights to Safety? (with Lieven Baele and Frank de Jong) [SSRN]
Abstract: We apply causality extraction (CE) algorithms on more than 36,000 articles from 2 major financial newspapers (WSJ, FT) and three major newswires (DJ, MNI Market News, Reuters) over the extended period 1980-2020 to shed light on the fundamental triggers of flights to safety. Our CE algorithm searches for sentences that causally relate flight-to-safety terms to a “cause”, or candidate FTS trigger. Our method identifies nearly 3,000 unique triggers of FTS, which we subsequently allocate using a dictionary-based method to nine economically motivated primitive categories. While some primitives such as “Funding Liquidity,” “Financial Intermediaries,” or “Pandemic” are only mentioned during specific FTS spells, others such as Political, Geopolitical, Macro, and terms broadly associated with Risk Aversion, Sentiment, and Uncertainty are more generally mentioned as potential triggers of FTS. We additionally show that causal FTS articles are associated with larger market responses (compared to days with noncausal FTS articles), especially if they refer to current (instead of future) FTS events.
Presentations: 2023 ENTER Jamboree in Mannheim · 10th Asset Pricing Workshop at the University of York* · 18th Belgian Financial Research Forum 2024* · 4th Annual Bristol Financial Markets Conference* · 31st Finance Forum in Tenerife · 2024 China International Conference in Finance in Beijing*
Keywords: Empirical asset pricing, Natural Language Processing, Dependency parsing, Causality extraction
* Presented by coauthor
Work in Progress
The Time Dimension of Media Tone and Topics
Abstract: Tone and topic have traditionally served as the primary textual dimensions for analyzing the relationship between newspaper articles and stock market returns. This paper introduces a third dimension — time. Leveraging natural language processing techniques to detect time references in text, I develop measures of past, present, and future media tone à la Loughran and McDonald (2011) and attention to the “recession” topic by Bybee, Kelly, Manela, and Xiu (2024) using all front-page articles from The Wall Street Journal published between May 2000 and December 2023. I evaluate the ability of these tone and topic measures to forecast future excess returns on the S&P 500, beyond the established predictors of Goyal and Welch (2008). Both past and future media tone positively predict future excess returns, with future tone exhibiting stronger predictive power over longer horizons. While recession attention holds no predictive power, its future component negatively predicts future excess returns. Out-of-sample, my present text measures, alone and together with the past and future ones, add predictive power above and beyond the Goyal and Welch (2008) predictors at the three-month and one-year horizons. These results are not attributable to increased risk, as neither tone measure forecasts higher future realized variance. Moreover, intermediary capital constraints and risk aversion do not fully explain their observed predictability.
Keywords: Empirical Asset Pricing, Forecasting, Natural Language Processing, Dependency parsing, Economics News
Inflation Narratives and Risk Premia (with Luís Fonseca, Giulia Martorana, and Fabian Schupp)
Abstract: Theory suggests that inflation risk premia are generally positive when investors believe supply shocks will outweigh demand shocks, and negative otherwise. We measure these beliefs using demand and supply narratives derived from inflation news through Causality Extraction, a tool that identifies causal relationships between inflation and its causes. We extract Euro Area demand and supply narratives from the Financial Times and Reuters inflation news, and United States demand and supply narratives from the Wall Street Journal’s inflation news. Our primary narrative variable, NetDemand, is the difference in the number of articles attributing inflation to demand versus supply factors. Consistent with theory, for both the Euro Area and United States, inflation risk premia are inversely related to NetDemand across maturities. NetDemand’s explanatory power remains strong even after controlling for the PMI and VIX, becomes stronger as risk aversion rises in the United States, and is not subsumed by the demand and supply contributions to inflation, views of professional forecasters, and narratives obtained from LLMs.
Keywords: Demand and Supply, News Media, Natural Language Processing, Dependency parsing, Causality extraction
Page updated
Google Sites
Report abuse