Monetary Policy NLP
Replicates and extends Aruoba & Drechsel (2024), who identify monetary policy shocks by scoring sentiment for 296 hand-picked economic concepts in FOMC Greenbooks. Asks: how robust is the method to the specific choice of concepts?
Can the Method Stand Up to Scrutiny?
Dictionary-based sentiment scoring of hand-picked concepts is inherently subjective. If a different researcher picked different concepts, would the identified monetary policy shocks change? This project replicates the original results exactly, then stress-tests the concept selection with LLM curation and permutation tests.
Replication + Stress Tests
Extract text from 787 Greenbooks; reproduce Table 2 (forecast error predictability) exactly and the Table 3 ridge specifications within ±0.01 (CV lambda grid variation).
DeepSeek identifies 595 concepts from 2,000 candidate n-grams and scores sentiment for all of them. Human vs. LLM curation: only 62.6% overlap (249/296).
Shuffle concept labels within each meeting (100 permutations × 2 lists) and re-estimate the shocks; compare shock correlations with the unshuffled baseline.
Sample split by Fed chair, placebo shuffle of the first principal component, and LLM-based risk signal extraction (upside/downside/balanced).
Labels Are Incidental
Label Shuffle Test · Within-Meeting Permutation
Interpretation: within any meeting, most concept scores point the same direction, so the common signal dominates regardless of which label is attached to which score. The concept set's role is dense temporal coverage of the document — the specific identity of each sampling point is secondary. The model does not overfit despite p/n ≈ 16: ridge (λ = 924) shrinks ~3,224 variables spanning ~155 effective dimensions.
Sentiment itself does add value across all Fed chairs: ΔR² of +0.16 (Volcker), +0.05 (Greenspan), +0.02 (Bernanke) over the baseline.
Deliverables include a 27-slide Beamer presentation, a 10-page discussion document with mathematical proofs, robustness-check notebooks, and all replication tables as CSV. Greenbook data is excluded from git; the repo documents the extraction and scoring pipelines.