GPT Deciphering Fedspeak: Quantifying Dissent Among Hawks and Doves (2407.19110v1)

Abstract: Markets and policymakers around the world hang on the consequential monetary policy decisions made by the Federal Open Market Committee (FOMC). Publicly available textual documentation of their meetings provides insight into members' attitudes about the economy. We use GPT-4 to quantify dissent among members on the topic of inflation. We find that transcripts and minutes reflect the diversity of member views about the macroeconomic outlook in a way that is lost or omitted from the public statements. In fact, diverging opinions that shed light upon the committee's "true" attitudes are almost entirely omitted from the final statements. Hence, we argue that forecasting FOMC sentiment based solely on statements will not sufficiently reflect dissent among the hawks and doves.

• The paper demonstrates GPT-4's ability to quantify dissent by analyzing linguistic differences between succinct FOMC statements and comprehensive meeting transcripts.
• It employs innovative sentiment analysis methods, including sentence-level scoring and logit scaling, to detect hawkish or dovish nuances.
• Empirical results reveal that brief statements understate dissent, with 47% in statements versus 82% in transcripts, highlighting transparency gaps.

GPT Deciphering Fedspeak: Quantifying Dissent Among Hawks and Doves

The paper "GPT Deciphering Fedspeak: Quantifying Dissent Among Hawks and Doves" presents an incisive analysis of the Federal Open Market Committee (FOMC) communications. The authors deploy GPT-4, an advanced Generative Pre-trained Transformer model, to examine the linguistic differences between terse official statements and lengthier transcripts of FOMC meetings, specifically to quantify dissent in discussions on inflation.

FOMC Communication and Transparency

The FOMC is pivotal in setting monetary policy in the United States, and it communicates its policy decisions in a multi-tiered manner. Immediately after meetings, concise, stylized "teal" statements are released. Approximately three weeks later, more detailed "olive" minutes are provided, and after five years, verbatim "violet" transcripts become available. This layered dissemination strategy evolved significantly over time, shifting from a stance of strategic opacity in the 20th century to one of greater transparency in the 21st.

Data Collation and Analysis

For analysis, the authors assembled a dataset encompassing FOMC statements and transcripts from 1994 to 2016. The paper focuses on comparing the content of "violet" transcripts and "teal" statements due to their marked difference in detail and style. Initially, regular expressions were employed to preprocess the transcripts, partitioning the text by individual speakers.

Approach and Methodology

Leveraging the superior performance of GPT-4 on text quantification tasks, as evidenced by prior work, the researchers utilized this model to assess FOMC communications. Three distinct sentiment measurements were executed using "teal" statements and two using "violet" transcripts:

1. Sentence-level Sentiment Analysis: Each sentence in the "teal" statements was scored independently for hawkish or dovish sentiment.
2. Holistic Sentiment Analysis: The entire "teal" statement was ingested into GPT-4 to provide a comprehensive sentiment score.
3. Logit-scaled Sentiment Score: This approach, which emphasizes relative differences, applied logit scaling to hawkish and dovish scores to highlight extreme positions.

For the more exhaustive "violet" transcripts, sentiment was assessed at the speaker level, aggregating individual speaker sentiments to produce an overall meeting sentiment.

Empirical Findings

The analysis revealed significant divergence between "teal" statements and "violet" transcripts. While official "teal" statements tended to present a unified, neutral stance, "violet" transcripts contained a much richer tapestry of dissent among FOMC members. Specifically, 47% of "teal" statements and 82% of "violet" transcripts displayed detectable dissent, primarily centered around economic conditions and inflation projections.

The results emphasize the limitations of relying solely on the stylized "teal" documents for understanding the true diversity of opinions within the FOMC. Notably, the divergence between "teal" and "violet" sentiment scores became pronounced post-2012, with "violet" transcripts trending more neutral to hawkish, whereas "teal" statements remained dovish.

Implications and Future Directions

This paper demonstrates the potential of GPT-4 in processing complex economic texts and quantifying nuanced dissent that traditional methods might overlook. By capturing the granular sentiments expressed in FOMC meetings, this approach provides a more comprehensive understanding of the Committee's internal deliberations and can improve the predictive accuracy of economic models that rely on FOMC communication.

Future research might extend this methodology to include "olive" minutes, further refining sentiment analysis across all communication forms. Additionally, enhancing prompt engineering for GPT-4 and incorporating richer few-shot learning examples could improve model precision, mitigating the model's current tendency to produce a higher number of neutral predictions.

Conclusion

The application of GPT-4 to FOMC documents underscores significant variations between publicly disseminated statements and the comprehensive internal discussions. The paper confirms that dissent among FOMC members, often mitigated or omitted in "teal" statements, is prevalent in "violet" transcripts. This research contributes to understanding the dynamic interplay of public-facing communication and internal committee deliberations, highlighting the potential of LLMs to transform the analysis of economic texts. As these models evolve, one can anticipate their broader application to more intricate and context-sensitive economic and financial investigations.