Interpreting Fedspeak with Confidence: A LLM-Based Uncertainty-Aware Framework Guided by Monetary Policy Transmission Paths

In one sentence

Summary

The paper addresses the challenge of interpreting "Fedspeak," the nuanced language used by the U.S. Federal Open Market Committee (FOMC) to signal monetary policy stances. Traditional sentiment analysis models struggle with the contextual ambiguity of central bank communications, while existing LLM approaches often lack interpretability and reliability metrics. The authors position their work to bridge this gap by combining domain-specific economic reasoning with uncertainty-aware decoding, aiming to enhance both classification accuracy and the trustworthiness of model predictions for financial forecasting and algorithmic trading. The proposed method augments Fedspeak texts by extracting financial entity relations and reasoning over monetary policy transmission paths using structured templates, emulating expert economic analysis. This augmented data is used to fine-tune LLMs, specifically Qwen-3-14B. A key innovation is the dynamic uncertainty decoding module, which quantifies Perceptual Uncertainty (PU) by decomposing it into Cognitive Risk (CR) and Environmental Ambiguity (EA) based on token logits. The model adaptively selects aggressive or conservative decoding strategies based on PU thresholds. Experiments are conducted on the Trillion Dollar Words FOMC dataset, covering meeting minutes, press conferences, and speeches from 1996 to 2022, evaluating performance using Macro-F1 and Weighted-F1 metrics against various zero-shot and fine-tuned baselines. Results show the framework achieves state-of-the-art performance, outperforming the best baseline by 6.6% in Macro-F1. The ablation study confirms that transmission path reasoning provides the largest performance gain, while uncertainty quantification significantly improves reliability. Statistical analysis reveals a strong positive correlation between PU and error rates, with low-PU predictions achieving high accuracy (0.7822 Weighted-F1) compared to high-PU predictions (0.4372). Limitations include reliance on hand-crafted templates, limited generalizability to non-US central banks, and difficulties with implicit statements and contextual confusion. The work offers a robust tool for policy stance analysis with built-in confidence calibration.

Research problem

The paper tackles the difficulty of automatically parsing and interpreting "Fedspeak," the stylized and ambiguous language used by the U.S. Federal Reserve to signal monetary policy stances. It addresses the limitations of traditional sentiment analysis models, which lack contextual nuance, and existing LLM approaches, which often lack interpretability and reliability metrics. The core problem is to develop a framework that not only classifies policy stances (hawkish, dovish, neutral) with high accuracy but also provides a measure of confidence to identify unreliable predictions.

Core contributions

• A domain-specific reasoning framework that augments Fedspeak texts with monetary policy transmission path relations to emulate expert economic analysis.
• A dynamic uncertainty decoding module that quantifies Perceptual Uncertainty (PU) by decomposing it into Cognitive Risk and Environmental Ambiguity.
• An adaptive inference strategy that switches between aggressive and conservative decoding based on PU thresholds to improve prediction reliability.
• State-of-the-art performance on the FOMC policy stance analysis task, significantly outperforming existing zero-shot and fine-tuned baselines.
• Empirical validation that PU serves as a reliable diagnostic signal, with a strong positive correlation between high PU and prediction error rates.

Method

• Data Augmentation: Extract financial entity relations and map them to monetary policy transmission paths using structured templates.
• Model Fine-tuning: Fine-tune LLMs (e.g., Qwen-3-14B) on the augmented FOMC dataset using LoRA.
• Uncertainty Quantification: Compute Perceptual Uncertainty (PU) as the product of Cognitive Risk (CR) and Environmental Ambiguity (EA) derived from token logits.
• Dynamic Decoding: Use PU thresholds to switch between aggressive (greedy) and conservative (sampling/neutral) decoding strategies.
• Evaluation: Compare against zero-shot and fine-tuned baselines using Macro-F1 and Weighted-F1.
• Statistical Validation: Use T-tests and logistic regression to validate the predictive power of PU on error rates.

Results & findings

• The proposed method achieves a Macro-F1 of 0.7327 and Weighted-F1 of 0.7426 on the combined FOMC dataset.
• This represents a 6.6% and 6.2% improvement over the strongest baseline (GPT-4.1 zero-shot).
• Performance gains are most significant for meeting minutes (7.4% Macro-F1 improvement) and speeches.
• Ablation shows that removing transmission path reasoning causes the largest performance drop, highlighting its importance.
• Predictions with low Perceptual Uncertainty (PU) achieve a Weighted-F1 of 0.7822, while high-PU predictions drop to 0.4372.
• Statistical tests confirm a significant positive correlation between PU and error rates (p < 0.001).
• The model struggles with implicit statements and contextual confusion, often misinterpreting third-party views or indirect signals.
• Fine-tuning benefits smaller models more than large zero-shot models like GPT-4.1, but the proposed method still outperforms them.

Data & experiments

• Dataset: Trillion Dollar Words FOMC dataset containing meeting minutes, press conference transcripts, and speeches from January 1996 to October 2022.
• Baselines: Comparison against zero-shot models (GPT-4.1, Gemini-2.5-pro, etc.) and fine-tuned models (FinBERT, Qwen-3, GLM-4, etc.).
• Metrics: Macro-F1 and Weighted-F1 scores for classification accuracy.
• Ablation Studies: Analysis of the impact of transmission path reasoning, entity relationships, and uncertainty decoding on performance.
• Uncertainty Analysis: Statistical tests (T-test, Mann-Whitney U, logistic regression) validating the correlation between PU and prediction errors.
• Case Studies: Qualitative analysis of model performance on explicit, contextual, and implicit sentence types.

Limitations & caveats

• Relies on hand-crafted templates for data augmentation, which may limit scalability.
• Limited generalizability to non-US central banks (e.g., BoE, ECB) due to data availability.
• Struggles with implicit statements and contextual confusion, leading to incorrect predictions.
• The 'refuse to answer' option was not implemented in main experiments to ensure fair comparison.
• Runtime and strategy constraints due to the complexity of the uncertainty decoding module.
• Performance on press conference transcripts is lower than on other text types due to dynamic context dependencies.

Curator notes

Discovery status: specific_domain_round3_promoted. Source links: FinDKG: Dynamic Knowledge Graphs with Large Language Models for Detecting Global Trends in Financial Markets