Case study — 2026
Voice of Monetary Policy
classifying the Fed's hawkish/dovish stance from how policymakers say it, not just what they say
- Python
- LibROSA
- FFmpeg
- PyTorch/TensorFlow
- BERT (Transformers)
- Bootstrap OLS
- DGX H200
replicates Gorodnichenko, Pham & Talavera (2023, AER) · 792 segments / 181 FOMC meetings, 2011–2019
01 — The problem
Can you read the Fed's stance from how they say it?
Most research on FOMC press conferences treats the transcript as the whole signal — what the Chair said, run through text sentiment models. But the base paper this project replicates argues that's only half the story: tension, confidence, and hedging often show up in vocal tone before they show up in careful, sanitized language. The question: does voice carry information about monetary policy stance that text alone misses — and does that gap show up in how markets actually move?
02 — The system
Two modalities, fused into one stance signal.
Each press conference is split into a preprocessed audio track and an aligned transcript, run through two independent classifiers, then fused into a combined voice-tone + text-sentiment reading that feeds an econometric model of market response.
- 01792 segments / 181 meetings
FOMC press conference videoOfficial Federal Reserve recordings
- 02FFmpeg
Preprocessing16kHz audio extraction, VAD silence removal (-40dB), DTW alignment to transcript
runs in parallel
03LibROSAMel-spectrogram features64 mel bands, 300Hz–8kHz, 25ms Hamming window, 10ms hop
- 0478% accuracy
SER classifierMLP on spectrogram statistics → positive/negative vocal valence
- 05Transformers
BERT tokenization + encodingWordPiece, 512-token sequences over Statement, Remarks, and Q&A segments
- 06F1 0.82
BERT classifierFine-tuned on FOMC language → hawkish/dovish text sentiment
- 07
Multimodal fusionVoice tone (emo) + weighted text sentiment (S_QASR), early + late fusion
- 0822 instruments · 0–15 day horizon
Bootstrap OLS regression2000 replications, BCa confidence intervals at 90%
- 09
Hawkish / dovish stanceValidated against cumulative abnormal returns
$ note — The fusion step controls for Swanson monetary-policy shocks and the Wu–Xia shadow rate, so voice and text are tested for information that survives alongside known policy-shock signals — not against a blank baseline.
03 — Results
What the numbers say.
“Voice tone stayed statistically significant even after controlling for text sentiment and policy shocks.”
792/181
segments / FOMC meetings, 2011–2019
78%
SER accuracy — vocal valence classification
0.82
BERT F1 — hawkish/dovish classification
Incremental, not redundant
Voice-tone coefficients remained statistically significant after controlling for text sentiment and Swanson policy shocks — the two modalities carry complementary information, not the same signal twice.
Q&A carries the signal
Unscripted Q&A segments produced the most informative vocal signals — likely because they reveal states that formal, prepared statements sanitize away.
Small-n econometrics
Bootstrap OLS with 2000 replications and BCa confidence intervals compensates for the small-sample problem inherent to meeting-level FOMC data — only ~181 events across a decade.
04 — Honest limitations
What the model doesn't capture.
− US Fed only
+ The framework is validated on FOMC communications alone — whether it generalizes to the ECB, BoE, or BoJ, with different communication cultures, is untested.
− No visual modality
+ Voice and text only — facial expression and gesture aren't captured. A CNN/ViT visual branch is scoped as future work, not part of the current pipeline.
− Binary sentiment
+ Both classifiers force a positive/negative, hawkish/dovish split, collapsing dimensional emotion — anger and nervousness look identical to the model.
− ~181 meetings
+ A decade of data is still a small-n problem; longer-horizon return estimates carry high variance despite the bootstrap correction.
Want the details?