Chapter 3· Manuscript, 2026
Returns and Volatility Around FOMC Announcements: A High-Frequency Analysis of Policy Tone and Novelty
Fed Tone Moves Prices, Novelty Moves Volatility
When the Fed speaks, what moves markets — what it says or how new it is? Tone drives directional returns; novelty drives volatility.
Simon-Pierre Boucher — contact@spboucher.ai
+12 bps
Equity gain within two hours after a 1σ dovish tone shift
t = −5.06
Stance × novelty interaction on VIX, persisting 5–120 minutes
6 of 7
Contracts where stance moves realized volatility (p < 0.01)
148
FOMC events with 1-minute futures data, 2008–2025
217
FOMC statement releases scored for tone and novelty, 2000–2025
7
Futures contracts studied: ES, VX, ZN, ZF, DX, CL, GC
Abstract
This essay asks what moves markets when the Federal Reserve speaks: the policy tone of the statement or its informational novelty. FOMC statements (217 releases, 2000–2025) are decomposed into hawkish/dovish tone and novelty using a dual-model NLP ensemble (MiniLM and BERT with TSDAE+MNRL fine-tuning and PCA-based reference selection), then linked to 1-minute futures data across seven contracts around 148 FOMC events from 2008 to 2025.
The two dimensions play distinct roles. Tone predicts directional returns: a one-standard-deviation dovish shift generates equity gains building to about +12 basis points within two hours. Novelty predicts volatility: the stance-by-novelty interaction on the VIX persists from 5 to 120 minutes (t = -5.06), and stance moves realized volatility in six of seven contracts at the 1% level, while pre-announcement placebo tests are null.
Key Results
Separating policy tone from novelty
Decomposes FOMC statements into hawkish/dovish tone and informational novelty, showing the two dimensions have distinct, complementary effects on returns and volatility.
Dual-model NLP ensemble for Fed text
Combines MiniLM and BERT with TSDAE+MNRL fine-tuning and PCA-based reference selection to score 217 FOMC statements from 2000 to 2025.
Tone predicts directional returns
A one-standard-deviation dovish shift generates equity gains that build to roughly +12 basis points within two hours of the announcement.
Novelty predicts volatility dynamics
The stance-by-novelty interaction on the VIX persists from 5 to 120 minutes after the release, with a t-statistic of -5.06.
Broad cross-market volatility evidence
Policy stance moves realized volatility in six of seven futures contracts at the 1% significance level, spanning equities, rates, currencies, and commodities.
Placebo-validated announcement effects
Pre-announcement placebo tests are null, confirming the estimated tone and novelty effects are driven by the FOMC announcements themselves.
Data
FOMC statements, 2000–2025
217 statement releases scored for hawkish/dovish policy tone and informational novelty via the dual-model NLP ensemble.
1-minute futures data around 148 FOMC events
Minute-level prices from 2008 to 2025 for seven contracts: ES, VX, ZN, ZF, DX, CL, and GC.
Methodology
NLP ensemble (MiniLM + BERT)
A dual-model text ensemble with TSDAE+MNRL fine-tuning and PCA-based reference selection extracts tone and novelty from FOMC statements.
Event regressions and minute-level panels
High-frequency event-window regressions and minute-level panel specifications link tone and novelty to returns and realized volatility across contracts.
Jordà local projections
Local projections trace the dynamic response of returns and volatility to tone and novelty shocks over the post-announcement window.
Placebo tests and five inference methods
Pre-announcement placebo windows and five alternative inference methods validate that the estimated effects are announcement-driven and statistically robust.
Reproducibility
- Full LaTeX source in PHD_chapitre3_theses_20260731/ is organized as chapitre3.tex plus sections/, tables/ (24 tables), 19 figures, a mathematical proofs appendix, and master.bib, with the compiled chapitre3.pdf included.
- The folder is a frozen July 31, 2026 snapshot of the March 6, 2026 manuscript; thesis adaptations live separately in these-ulaval/chapitre3/.
- In the assembled thesis, the chapter's proofs and extras appear as appendices A–B and its labels are prefixed ch3:, with references merged into the 271-key consolidated bibliography.