PaperlayerView exact revision on arXiv

Source-linked AI summary

Mind the Shift: Decoding Monetary Policy Stance from FOMC Statements with Large Language Models

Yixuan Tang, Yi Yang

arXiv:2603.14313v1cs.CL

TL;DR

Existing approaches often classify FOMC statements in isolation, although monetary-policy communication is interpreted through shifts across meetings. DCS uses frozen LLM representations and consecutive meetings for annotation-free, delta-consistent scoring, outperforming comparison methods while producing economically relevant scores. The learned polarity remains ambiguous without post-hoc anchoring.

  • Problem

    Existing stance methods overlook sequential shifts between FOMC statements, despite markets responding to departures from prior communication.

  • Method

    DCS jointly learns absolute stance and relative inter-meeting shift scores from frozen LLM representations using a delta-consistency objective without human labels.

  • Results

    DCS consistently outperforms supervised and LLM-as-judge baselines, achieving up to 71.1% sentence-level classification accuracy and strong associations with inflation indicators and Treasury yields.

  • Takeaways & Limitations

    Frozen LLM representations encode monetary-policy information that can be recovered through relative temporal structure for macroeconomic and financial analysis.

  • Takeaways & Limitations

    Because the self-supervised objective is polarity-symmetric, hawkish and dovish directions require post-hoc anchoring before economic interpretation.

Abstract

from arXiv · show

Federal Open Market Committee (FOMC) statements are a major source of monetary-policy information, and even subtle changes in their wording can move global financial markets. A central task is therefore to measure the hawkish--dovish stance conveyed in these texts. Existing approaches typically treat stance detection as a standard classification problem, labeling each statement in isolation. However, the interpretation of monetary-policy communication is inherently relative: market reactions depend not only on the tone of a statement, but also on how that tone shifts across meetings. We introduce Delta-Consistent Scoring (DCS), an annotation-free framework that maps frozen large language model (LLM) representations to continuous stance scores by jointly modeling absolute stance and relative inter-meeting shifts. Rather than relying on manual hawkish--dovish labels, DCS uses consecutive meetings as a source of self-supervision. It learns an absolute stance score for each statement and a relative shift score between consecutive statements. A delta-consistency objective encourages changes in absolute scores to align with the relative shifts. This allows DCS to recover a temporally coherent stance trajectory without manual labels. Across four LLM backbones, DCS consistently outperforms supervised probes and LLM-as-judge baselines, achieving up to 71.1% accuracy on sentence-level hawkish--dovish classification. The resulting meeting-level scores are also economically meaningful: they correlate strongly with inflation indicators and are significantly associated with Treasury yield movements. Overall, the results suggest that LLM representations encode monetary-policy signals that can be recovered through relative temporal structure.

1 Introduction

FOMC stance is relative: markets respond to changes from prior statements, not only to isolated hawkish or dovish wording. DCS exploits this temporal structure to learn unlabeled continuous stance scores and shows strong classification and economic results.

  • FOMC statements influence expectations about future interest rates and global financial markets.
  • Existing dictionary, supervised, and LLM-as-judge methods face limitations involving context, annotation costs, subjectivity, evolving language, prompts, or reproducibility.
  • A statement can be hawkish in isolation yet represent a dovish shift when it follows a more strongly hawkish statement.
  • DCS maps frozen LLM representations to absolute stance and relative inter-meeting shift scores, using temporal ordering as self-supervision without human labels.
  • 71.1% accuracy was achieved on sentence-level hawkish-dovish classification, while meeting-level scores correlated with inflation indicators and Treasury yields.The evaluation spans LLMs from 1B to 14B parameters.
  • DCS consistently outperformed supervised and LLM-as-judge baselines across LLMs of varying scale.

2 Related Work

Research on monetary-policy text has progressed from transparent word counts to richer representations and self-supervised probing. These approaches motivate extracting stance information from pretrained language-model representations.

  • Dictionary methods score FOMC statements using predefined vocabularies and word frequencies but miss discourse-level context.
  • Machine-learning approaches use richer text representations, including unsupervised topic models and expert-annotated sentence classifications.
  • Representation probing treats pretrained LLM latent spaces as sources of semantic concepts that can be extracted with lightweight modules.

3 Method

DCS models monetary-policy stance as a continuous trajectory by separating absolute stance from relative shifts between consecutive FOMC statements. Its delta-consistency objective supplies self-supervision, while confidence regularization and post-hoc anchoring support interpretable scores.

  • 3 Method: DCS maps frozen LLM representations to continuous stance scores while constraining them with relative temporal shifts.
  • 3.1 Problem Formulation: A continuous score s_t in [0, 1] represents each statement, with values near 1 hawkish and values near 0 dovish.
  • 3.2 Representations: Absolute prompts assess individual statements, whereas relative prompts assess directional shifts from the preceding meeting.
  • 3.3 Dual-Axis Projection: A dual-axis projection maps absolute and relative representations to scalar logits for stance and inter-meeting policy movement.
  • 3.4 Delta-Consistent Objective: The delta-consistency objective aligns changes in absolute stance with predicted relative shifts between consecutive statements.The relative output is bounded with a scaled hyperbolic tangent to limit extreme shifts.
  • 3.4 Delta-Consistent Objective: The final objective combines primary delta-consistency loss with an auxiliary confidence regularizer whose weight is strengthened after temporal structure is learned.
  • 3.5 Post-Hoc Directional Anchoring: The symmetric self-supervised objective can invert hawkish-dovish polarity, so DCS uses hawkish and dovish exemplars for post-hoc directional anchoring.

4 Experiments

The experiments evaluate DCS against dictionary, supervised, prompting, and probing baselines, testing both sentence-level classification and the economic relevance of meeting-level scores. DCS performs strongly on expert-aligned stance classification and shows associations with inflation indicators and Treasury yields.

  • Evaluation dimensions: DCS is evaluated on sentence-level expert annotations and meeting-level associations with inflation indicators and Treasury yields.The evaluation uses FOMC statements, a sentence-level hawkish–dovish benchmark, and macroeconomic and market data.
  • Baselines: The baselines span dictionary scoring, an expert-trained FOMC-RoBERTa model, LLM-as-judge variants, and supervised linear probes on frozen LLM representations.The dictionary score is based on the difference between hawkish and dovish word counts, while the linear probe uses labeled training data.
  • Sentence-level stance classification: 71.1% accuracy is DCS’s best sentence-level result, achieved with Qwen3-4B, while DeepSeek-R1-14B achieves the best macro-F1 of 0.74.Across all four backbones, DCS outperforms the corresponding LLM Judge, Logit-Based Judge, and Linear Probe baselines; its accuracy gain over Linear Probe ranges from about 1 to nearly 10 percentage points.
  • Macroeconomic validation: 0.6237 is DCS’s highest Spearman correlation with inflation changes, obtained with DeepSeek-R1-14B.Across backbones, DCS’s Spearman correlations range from 0.5382 to 0.6237 for CPI and from 0.4546 to 0.5530 for PPI, consistently exceeding same-backbone judge and probe baselines.
  • Market validation: β = 2.12 is the strongest reported DCS association with Treasury yields, at the 2-year maturity on DeepSeek-R1-14B.DCS remains significant at the 10- and 20-year maturities and is the only method reported to show consistently significant associations across maturities.

5 FOMC Stance Trajectory and Robustness Across Policy Regimes

DCS stance scores are evaluated across four Federal Reserve policy regimes, including the near-zero-rate period where standard rate signals weakened. DCS remains positively related to inflation indicators across regimes, supporting robustness to regime change through relative inter-meeting shifts.

  • Policy-regime partitioning: Four evaluation periods span conventional policy, near-zero rates, normalization, and the pandemic-and-tightening cycle.The periods are P1 (2003–2008), P2 (2008–2015), P3 (2015–2020), and P4 (2020–2025).
  • Inflation robustness: DCS is the only method with positive correlations with both CPI and PPI across all four periods under Pearson and Spearman metrics.The supervised Linear Probe remains near zero in Periods 1–3 and becomes unstable in Period 4.
  • Baseline comparison: Prompt-based baselines are less stable across regimes, with Logits Based Judge correlations turning negative in Period 2 and LLM Judge correlations near zero or negative.Logits Based Judge performs strongly in Period 1 but loses positive correlations during the near-zero-rate period.
  • Interpretation: Modeling relative inter-meeting shifts improves robustness to regime change compared with relying on fixed labels or isolated stance signals.The reported interpretation is that temporal structure keeps DCS stable across the policy environments in the sample.

6 Ablation Study

Ablations identify delta-consistency as DCS’s primary learning signal, while confidence regularization calibrates the absolute stance scale and dual-axis projections separate stance dimensions. Final-layer representations provide the strongest inflation correlations.

  • Losses and architecture: Removing Ldelta lowers accuracy to 0.47 and CPI correlation from 0.62 to 0.38, making delta-consistency the primary learning signal.The ablation reports the largest performance drop for this removal.
  • Losses and architecture: Removing Lconf preserves sentence-level F1 at 0.72 but reduces CPI correlation to 0.25, indicating its role in calibrating absolute stance.The regularizer affects economic calibration more than the reported sentence-level F1.
  • Losses and architecture: Replacing the dual-axis projection with a shared representation lowers all four metrics, supporting separate representations for absolute and relative signals.The reported design separates the two stance-related signals rather than combining them in one projection.
  • Layer selection: The final layer yields the strongest inflation correlations, with ρ = 0.62 for CPI and ρ = 0.55 for PPI.It outperforms all intermediate layers by a clear margin.

7 Conclusion

The paper concludes that DCS can recover continuous monetary-policy stance from frozen LLM representations by aligning absolute stance with relative shifts between consecutive FOMC meetings. Across four LLMs, it outperforms the evaluated baselines and produces scores associated with inflation and Treasury yields.

  • Conclusion: DCS maps frozen LLM representations to continuous stance scores without annotation by aligning absolute scores with relative shifts between consecutive meetings.The framework uses two prompt views: one for individual-statement stance and one for directional change between consecutive statements.
  • Conclusion: Across four LLMs, DCS consistently outperforms supervised and prompt-based baselines on sentence-level classification and yields scores correlated with inflation indicators and Treasury yields.The conclusion also reports that ablations confirm contributions from each component, with delta-consistency as the primary learning signal.
  • Implications: Temporal self-supervision provides an effective way to extract economically meaningful monetary-policy information from frozen LLM representations.The paper frames the resulting measure as useful for macroeconomic analysis and financial decision-making.
  • Prompt views: The absolute prompt represents an individual statement’s tighter-or-looser stance, while the relative prompt represents the directional shift between consecutive statements.The prompts use fixed templates with statement text substituted into their placeholders.

B Annual Statement Counts

This material documents the corpus, filtering, anchoring, and model settings used to analyze FOMC statements. The corpus spans 2003–2025, and the pipeline filters statements to retain policy-relevant content before applying DCS.

  • Annual statement counts: Table 5 reports the annual distribution of FOMC statements in the corpus from 2003 to 2025.The corpus includes official post-meeting statements spanning January 2003 to December 2025.
  • Directional anchoring: Post-hoc hawkish and dovish anchors resolve the polarity ambiguity of the learned stance axis after training converges.The anchors are not used for gradient-based training; both learned axes are sign-flipped when their polarity is reversed.
  • Statement filtering: A rule-based sentence filter removes boilerplate and procedural language while retaining policy-relevant statement content.The filter uses indicator and directional-movement terms to identify sentences likely to convey stance.
  • Model configuration: DCS uses four frozen LLM backbones with final-layer hidden states at the last-token position and a dual-axis projection module.The standard confidence-regularizer schedule uses 100 warm-up epochs followed by a 100-epoch linear ramp.
  • Model configuration: Table 9 defines τ as the delta-consistency-loss temperature and λ as the auxiliary confidence-regularizer weight.These are the model-specific hyperparameters summarized for DCS experiments.

F Regression Analysis of Inflation

The regression analysis tests whether meeting-level DCS stance scores are associated with year-over-year inflation changes. Using Qwen3-4B for the OLS analysis, higher scores show positive, highly significant associations with both CPI and PPI changes.

  • Regression Setup: OLS regressions relate year-over-year CPI and PPI changes to meeting-level DCS stance scores using Qwen3-4B.Standard errors are reported in parentheses, with *** indicating p < 0.01.
  • CPI Results: 3.18 percentage points: a one-unit stance-score increase is associated with higher year-over-year CPI change (p < 0.0001).The CPI model explains 25.2% of variation (R2 = 0.2521, N = 199).
  • PPI Results: 12.77 percentage points: a one-unit stance-score increase is associated with higher year-over-year PPI change (p < 0.0001).The PPI model explains 23.0% of variation (R2 = 0.2303, N = 174).
  • Interpretation: Higher DCS stance scores tend to align with stronger inflationary conditions, consistent with the correlation analysis.The reported relationships concern associations between extracted stance scores and observed inflation changes.

H Ablation: Loss Components

The ablation study separates the contributions of DCS’s delta-consistency loss, confidence regularizer, and dual-axis representation. The delta-consistency loss provides the primary temporal learning signal, while the confidence regularizer helps calibrate absolute scores and separate representations improve overall validity.

  • Study Design: The ablation removes one DCS component at a time and evaluates sentence-level classification alongside meeting-level inflation correlations.The study uses DeepSeek-R1-14B and compares the full model with three variants.
  • Delta-Consistency Loss: Accuracy falls to 0.47 and CPI correlation to 0.38 without Ldelta, versus 0.62 for the full-model CPI correlation.The confidence regularizer alone pushes scores toward extremes but lacks temporal directionality.
  • Confidence Regularizer: Without Lconf, sentence-level F1 remains 0.72, but CPI correlation drops from 0.62 to 0.25.Delta-consistency alone can rank statements but struggles to calibrate the absolute score scale.
  • Representation Design: The single-axis variant remains reasonable but falls below the full model on all four metrics.It uses the absolute-prompt representation for both stance and shift estimates instead of separate representations.

I Layer Selection Analysis

Layer selection shows that DCS performance varies substantially across DeepSeek-R1-14B layers. The highlighted final layer L48 produces the strongest CPI and PPI correlations, clearly exceeding intermediate-layer results.

  • Layer Comparison: The final layer L48 achieves the highest correlations with both CPI and PPI indicators.Figure 4 compares DCS stance-score Spearman correlations across layers.
  • Best Layer: ρ = 0.62 for CPI and ρ = 0.55 for PPI: L48 outperforms all intermediate layers by a clear margin.The second-best layer is L26, with ρ = 0.52 for CPI and ρ = 0.49 for PPI.
  • Layer Variation: Performance varies substantially across layers, with many early and late-middle layers yielding weak correlations.This variation supports selecting the final layer rather than assuming all hidden layers encode the signal equally well.
Loading 2603.14313v1…