A FinBERT-Driven Framework for Quantifying Federal Reserve Communication
This paper develops an Automated Monetary Policy Stance Index (MPSI) leveraging fine-tuned FinBERT to quantify the Federal Reserve's communication stance from FOMC statements, meeting minutes, and Chair speeches. The hybrid framework integrates contextual NLP (30%) with domain-specific keyword detection (70%) to produce a continuous score bounded between −100 (maximally dovish) and +100 (maximally hawkish).
Validated against 284 Fed documents spanning 2020–2026, MPSI achieves a Pearson correlation of r = 0.689 (p < 0.001) with the Federal Funds Rate. Cross-correlation analysis reveals MPSI leads FFR by 2–3 months (r = 0.853 at lag +3), demonstrating strong forward-looking signal value for monetary policy forecasting. MPSI outperforms all baselines including Loughran-McDonald (r = 0.482) and VADER (r = 0.413).
Domain-specific lexicon of 35+ hawkish terms (tighten, hike, restrictive, inflation risk) and 35+ dovish terms (accommodate, ease, support, labor market slack) developed for Fed communication. Weighted net keyword score scaled to [−100, +100].
ProsusAI/FinBERT processes 512-token document windows to classify hawkish/neutral/dovish probability. Handles negation and conditionality that keyword-only approaches miss. MPSI₍FinBERT₎ = P(hawkish) − P(dovish) × 100.
MPSI = 0.70 × MPSI_keyword + 0.30 × MPSI_FinBERT
| Pearson r | 0.689*** |
| R-squared | 0.475 |
| 95% CI | [0.581, 0.779] |
| Regression slope β₁ | 0.102 |
| Intercept β₀ | 3.320 |
| Granger Causality (F) | 8.42, p = 0.001 |
| Breusch-Pagan | p = 0.14 ✓ |
| Shapiro-Wilk (W) | 0.976 ✓ |
| Durbin-Watson | 1.89 ✓ |
| Directional Accuracy | 80% |
| Document Type | N | Mean MPSI | Coverage |
|---|---|---|---|
| FOMC Statements | 96 | -3.2 | 2020–2026 |
| FOMC Minutes | 48 | -12.8 | 2020–2026 |
| Chair Speeches | 38 | -9.2 | 2020–2025 |
| Economic Outlook | 102 | -1.8 | 2022–2026 |
| Total | 284 | -7.9 | 74 months |
Dompreh, F. (2026). Automated Monetary Policy Stance Index (MPSI): A FinBERT-Driven Framework for Quantifying Federal Reserve Communication. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.6859459View on SSRN