US Rates StrategyDeciphering Fedspeak

中文标题

美国利率策略 · 解读美联储言论

SMBC09/30/2026updated 10/17/2026

Citable research brief

Core conclusion · Tlines extraction

SMBC applies LLM scoring from -1 (dovish) to +1 (hawkish) to all Fed commentary since 2009. FOMC members' latest speeches generally score above 0.5, leaning hawkish, with inflation and labor market topics most frequent. Chair Warsh's July and September press conferences scored hawkish overall, though less so than his opening remarks, likely due to word hedging that avoids forward guidance. The model found no clear TU 2-year futures reaction to the tone of Warsh's answers.

Institution: SMBC
Published: 09/30/2026
Research topic: Fed Policy
Time horizon: 2023 · 2025

Fed Policy · FED

Key arguments

LLM scoring shows current FOMC commentary is generally above 0.5 and hawkish, but members are not consistent hawks or doves and shift with economic conditions.
Averaging scores to classify a speaker as hawk or dove may misidentify their leaning because some members speak rarely or focus on banking and regulation.
Chair Warsh's press conference answers score less hawkish than his opening remarks, likely reflecting word hedging to avoid forward guidance rather than an inherently dovish stance.
Warsh's pursuit of a quieter Fed and reluctance to give forward guidance could reduce the signal value of LLM sentiment models.

Key numbers

Sentiment score scale: -1 (dovish) to +1 (hawkish)
Recent FOMC speech scores: generally above 0.5
Waller score shift: -0.9 (Nov 17) to +0.2 (Sep 3)
Powell press conference scores (Jan, Mar, Apr): zero
Dictionary size: 200 words and key phrases across six topics

Main risks

LLM models are black boxes, making it hard to see how scores are derived.
Word hedging to avoid forward guidance may reduce model usefulness for analyzing the Chair's commentary.
No clear market reaction: TU 2y futures did not respond to the tone of Warsh's answers.

Corresponding source evidence

Chair Warsh’s quest for a “quieter” Fed could dampen the signal value of LLM techniques .
The scores are meant to reflect how the speaker will vote at the next FOMC meeting.
The model reads the text holistically, so while it reviews non-monetary policy topics like discount window modernization, it will score them as neutral (0.0) with low relevance.

Context: the conclusion is Tlines' automated extraction, not the institution's wording. Quoted evidence is publisher wording. Scope and date must travel with the conclusion.

Verify at the original source ↗