Diego Vilán and WanTing Xu argue in FEDS Notes that speech measures can move earlier

Diego Vilán and WanTing Xu argue in FEDS Notes that speech measures can move earlier

Diego Vilán and WanTing Xu published this note on October ninth, two thousand twenty-six, in FEDS Notes. They compare measures of monetary-policy tone drawn from central-bank speeches with semi-structural models, which estimate policy stance from macroeconomic data. They report that the measures often move in the same direction around major policy shifts, but speech-based measures can turn earlier and vary more in the short term. The authors argue that the difference reflects what each measure captures: speeches convey policymakers’ interpretation of incoming information, balance of risks and intended policy direction, while the semi-structural measure supplies a benchmark grounded in economic data. They find that a language model and a fixed dictionary track a common cycle but differ in magnitude, especially after the Global Financial Crisis; they suggest the language model captures context and conditionality that a dictionary misses. Their conclusion is that the measures are related but not interchangeable, and may provide a richer assessment when used together. They recommend future research on language-model-based indicators as a timely complement to conventional semi-structural measures. Whether earlier shifts in speech-based measures consistently add useful information beyond the macroeconomic benchmark remains unknown.

Publication: d93aacbe-bd35-4761-16cc-036f392b93f7

attribution — pass: The reading attributes the note to Diego Vilán and WanTing Xu in FEDS Notes, which matches the byline and citation. Findings and conclusions are consistently framed as the authors' reports, arguments and recommendations.

Vilan, Diego and WanTing Xu (2026). "Assessing Monetary Policy Globally: Evidence from LLMs and Semi-structural Models," FEDS Notes.

summary — uncertain: The content is faithful, but a non-specialist lacks some essential context. The reading never says FEDS Notes is published by the Federal Reserve Board. Its opening describes 'measures of monetary-policy tone drawn from central-bank speeches' without stating that one measure is scored by an LLM (Claude Haiku 4.5), which is the note's AI connection. That link only appears indirectly later. The title 'speech measures can move earlier' is vague about what they move earlier than. Smallest repair: name the Federal Reserve Board, say the speech tone is scored by an LLM and a dictionary across the BIS speech corpus, and say 'earlier than macro-based stance estimates'.

The second is an LLM-based model that employs Claude Haiku 4.5.

date — pass: originalPublished is 2026-10-09 at 15:00 UTC, matching the source dateline of October 09, 2026. The reading states the date at day precision. It contains no hindsight or later events, and the evidence is a publisher page, so no archive disclosure is needed. The evidence lists a Last-Modified of 2026-10-16 but does not flag the page as edited after publication, and the claims rely on the note's core text.

October 09, 2026

support — pass: Each claim matches the source text. The source states the post-GFC magnitude divergence and the context/conditionality interpretation. It describes the lead as varying in timing and magnitude, and characterizes the factor patterns as descriptive. It also frames complementarity and the future-research recommendation in hedged terms. The final unknown and the question follow from that hedging rather than inventing a benefit.

although the timing and magnitude of this lead vary across countries and episodes.

media — pass: No media was selected, and no verified media candidates were supplied, so there is no media identity or relevance issue.