Quantifying Narratives: Applications of Large Language Models for Fed Sentiment Drift Modeling

Central bank communication has become a central pillar of modern monetary policy. Federal Reserve wields significant influence over market expectations for various asset classes not only through policy actions but also through the way it communicates its assessment of the economy and its forward guidance. Yet, navigating the nuances of their guidance and assessment remains a challenging task even for experienced analysts. Understanding how Federal Reserve communication evolves over time and how it varies across different speakers proves to be an even harder problem. The growing literature applies textual analysis to study Federal Reserve communication. Most existing approaches, however, formulate the problem as classification at the word, sentence, or document level. While informative, such methods do not impose a coherent ordering across documents and are ill-suited for capturing cross-speaker and cross-time dynamics in policy stance. We propose a ranking-based framework that treats Federal Reserve communication as a latent ordinal object, inferred from pairwise comparisons of speeches and statements by different Fed members. Our approach combines large language models (LLMs) with learning-to-rank methods. We use LLMs to extract topic-specific content and to generate structured pairwise rankings across multiple dimensions, including labour market conditions, inflation, and interest rates forward guidance. These pairwise judgments are inherently noisy and may violate transitivity. We address that by using HodgeRank algorithm to produce a least-squares optimal global ordering scores and to quantify inconsistency in the obtained ranking. To enable out-of-sample scoring without rerunning expensive LLM inference we learn a regularized mapping from document embeddings to ordering scores using supervised dimensionality reduction and RankNet. We evaluate our measures against established empirical patterns in Federal Reserve communication from 2021 to 2025. The resulting scores capture meaningful variation across time and speakers and align closely with documented shifts in policy narratives. We also find that our forward-guidance measure contains information about movements in two-year treasury yields. Our approach provides a way to numerically quantify variability of the Central Bank's messaging through time, across speakers, and along various topics.