Using LLM Sentiment Scores to Analyze FOMC Transcripts

Summary

The document introduces general and finance-tuned language models, then describes using natural language processing to turn financial text into sentiment measures. It outlines a workflow for collecting and preprocessing Federal Open Market Committee transcripts, scoring segments with FinBERT, and tracking sentiment over time. A trading example uses rolling sentiment scores and positive and negative thresholds to define long and short entries and exits during a meeting.

The article also discusses model training and adaptation, including pretraining and human feedback, and mentions applications to news, speeches, and reports. Its evidence is illustrative: it describes minute-level transcript scores and a threshold-based strategy, but the supplied text omits much of the trading rules and does not report performance or validate the signals against market prices. Sentiment scores are model outputs rather than direct measures of future returns, and results may depend on the text, model, and chosen thresholds.

Key ideas

• LLMs can extract sentiment from financial text such as news, speeches, and meeting transcripts.
• Finance-tuned language models may better capture domain-specific wording than general-purpose models.
• The described workflow preprocesses FOMC transcript segments and assigns sentiment scores with FinBERT.
• Rolling transcript sentiment and thresholds can be used to form trading signals, though the available strategy description is incomplete.
• Sentiment signals need validation against market prices before they can support claims about trading performance.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.