spb/phd_thesis Public

1\section{Introduction}23We study how the textual content of Federal Reserve communications affects asset returns and market volatility at high frequency. The Federal Open Market Committee (FOMC)---composed of the seven members of the Board of Governors, the president of the Federal Reserve Bank of New York, and four of the remaining eleven Reserve Bank presidents on a rotating basis---meets eight times per year to set the target federal funds rate. Since 1994, the Committee has released a public statement after each meeting; since 2011, the Chair has also held post-meeting press conferences. Other communications include meeting minutes (released three weeks later) and the Summary of Economic Projections with individual rate forecasts (the ``dot plot''). Our analysis focuses exclusively on post-meeting statements, which are the first piece of information available to market participants at a precise, scheduled time. These statements affect U.S. asset prices, global capital flows, and exchange rates \citep{blinder2008,campbell2012,wongswan2009,ehrmann2011}.45Our central research question is: \textit{Do different semantic dimensions of FOMC communications---specifically policy tone and informational novelty---affect financial market returns and volatility through distinct economic channels?} If tone captures the directional policy signal (hawkish vs.\ dovish) while novelty captures how much genuinely new information the statement contains, standard asset pricing theory predicts that these dimensions should have different effects. In the standard framework, asset prices equal expected future cash flows discounted at a rate that reflects policy expectations \citep{fama1970,bernanke2005}, so the directional content of the statement---tone---should move expected cash flows and discount rates, and hence returns. Volatility, by contrast, reflects uncertainty and disagreement about how to interpret new information \citep{veldkamp2011,patton2013}, so the amount of new language---novelty---should affect information processing complexity and interpretive uncertainty, and hence volatility. We test this prediction using 1-minute price data for 7 futures contracts---the E-mini S\&P 500 (ES), futures on the Chicago Board Options Exchange Volatility Index (VIX futures, VX), the 10-year (ZN) and 5-year (ZF) Treasury notes, the U.S. Dollar Index (DX), crude oil (CL), and gold (GC)---across 148 FOMC events from 2008 to 2025, estimating panel minute-level regressions, event-level regressions, and local projection impulse response functions \citep{jorda2005}. Our identification strategy exploits the exogenous timing of announcements and is validated by pre-announcement placebo tests \citep{andersen2003micro,andersen2007real}.67To measure these dimensions, we construct two primary variables: (1) a tone measure capturing the hawkish-dovish spectrum, and (2) a novelty measure quantifying the semantic distance between consecutive FOMC communications. Concretely, we proceed in three steps. First, we convert each statement into a numerical vector (an \textit{embedding}) using two language models of different sizes---MiniLM (which produces 384-dimensional vectors) and BERT (768-dimensional vectors)---and combine their outputs; using two architecturally distinct models guards against findings that are artifacts of a single model. Second, because generic language models are not trained on central bank language, we adapt both models to the corpus of FOMC statements in two stages: an unsupervised stage based on the Transformer-based Sequential Denoising Auto-Encoder (TSDAE) of \citet{wang2021tsdae}, which teaches the models Fed-specific vocabulary, followed by a supervised contrastive learning stage on pairs of statements with known relationships. Third, we locate each statement on the hawkish-dovish axis by measuring its proximity to reference statements at the two extremes of the spectrum. Whereas previous studies select these reference statements by hand---introducing researcher degrees of freedom, since results may depend on which dates the researcher chooses---we select them algorithmically using principal component analysis (PCA), making the construction of the semantic axes fully reproducible. Novelty is then measured as the distance between the embeddings of consecutive statements.89FOMC announcements are widely seen as the most important scheduled monetary policy events in global financial markets, as they contribute substantially to shaping expectations about the future path of the U.S. economy. Because markets are forward-looking, surprising announcements are typically followed by sizable adjustments across several markets \citep{bernanke2005,blinder2008}. \citet{savor2014} show that a disproportionate share of the equity risk premium is earned on macroeconomic announcement days, with FOMC days being particularly important. \citet{brusa2015,brusa2019} find that average stock returns and Sharpe ratios on FOMC days are 20--40 times higher than on non-announcement days. \citet{lucca2012} document a systematic pre-FOMC drift in equity prices. \citet{nakamura2018} and \citet{jarocinski2020} decompose FOMC surprises into policy shocks and information shocks and show that these have different effects on asset prices.1011While earlier research focuses exclusively on the measurable policy surprise content of FOMC announcements, typically extracted from federal funds futures \citep{kuttner2001,bernanke2005}, a more recent strand of literature has begun to apply textual analysis to extract subtler information from the language itself \citep{hansen2017,shapiro2019,schmeling2019,gorodnichenko2023}. This is an important step forward, given that the early contributions identify the effect of unexpected rate changes but ignore the qualitative content of the accompanying statement---forward guidance, risk assessments, and descriptions of economic conditions---which \citet{gurkaynak2005} show can move long-term yields even when the rate decision is fully anticipated. Within this new strand of literature, however, most studies use daily data \citep{shapiro2019,schmeling2019,eklund2024}, which cannot separate the immediate market reaction to the statement from confounding information that arrives later in the day. Furthermore, the standard approach treats the statement as a one-dimensional object (hawkish vs.\ dovish), ignoring other dimensions such as how much the statement departs from prior language. In this paper, we address both limitations by using 1-minute data and by decomposing statements into two distinct dimensions: directional tone and informational novelty.1213Textual analysis of financial communications has progressed from keyword counting \citep{bligh2008} and dictionary-based sentiment scoring \citep{loughran2011} to transformer-based language models that capture context and word order \citep{gentzkow2019,kenton2019,araci2019}. Within central bank communication, \citet{hayo2010,apel2012} develop systematic measures of FOMC tone and show that these predict monetary policy expectations. \citet{shapiro2019,hansen2017} use natural language processing (NLP) classifiers to place FOMC statements on the hawkish-dovish spectrum, and \citet{eklund2024} find that hawkish sentiment measures predict subsequent realizations of the consumer price index (CPI). However, this literature treats FOMC communication as essentially one-dimensional---hawkish versus dovish---and has not examined whether other dimensions of the text, such as how much the language departs from the prior statement, have independent effects on asset prices.1415We define informational novelty as the degree to which a new FOMC statement departs from the previous one, measured as the cosine distance between their sentence-transformer embeddings. It is useful to distinguish this concept from three related constructs. \textit{Tone} or \textit{sentiment} refers to the directional content of text (hawkish vs.\ dovish) and has been studied using dictionaries \citep{loughran2011} and classifiers \citep{shapiro2019}. \textit{Economic policy uncertainty} (EPU), as measured by \citet{baker2016}, is a macro-level index based on newspaper coverage, tax code provisions, and forecaster disagreement. \textit{Monetary policy surprises}, in the tradition of \citet{kuttner2001}, are the unexpected component of the rate decision itself. Our novelty measure differs from all three: it captures changes in \textit{how} information is presented between consecutive statements, independent of both directional stance and the rate decision. A statement can be highly novel yet neutral in tone (e.g., introducing new forward guidance language), or low in novelty yet accompanied by a large rate surprise.1617Why should novelty matter for asset prices? Under rational expectations \citep{muth1961,fama1970}, only genuinely new information should move prices; repetitive content should already be priced in. But theory also suggests that novel and familiar information may affect returns and volatility differently. Returns respond to revisions in expected cash flows and discount rates---that is, to directional content. Volatility may respond more to disagreement among market participants about the interpretation of new language or to the processing cost of unfamiliar information \citep{patton2013,hu2019,veldkamp2011}. Consistent with this distinction, \citet{manela2017} show that their news-based implied volatility index (NVIX) predicts future market volatility.1819Our principal finding is that tone is associated with directional asset returns while novelty is associated with volatility, consistent with two distinct transmission channels. We make two main contributions. First, we develop a decomposition framework for FOMC statements based on a dual-model ensemble (MiniLM and BERT) fine-tuned on the FOMC corpus with data-driven PCA-based reference selection. Second, we provide the first high-frequency evidence on the differential effects of tone versus novelty, using local projection impulse response functions across 7 futures contracts and 148 FOMC events. In addition, we show that these two dimensions operate through distinct channels: tone is associated with fundamental valuations (equity returns building to about 12 basis points per standard deviation within two hours), while novelty is associated with uncertainty resolution (the stance--novelty interaction on VIX futures has a $t$-statistic of $-5.06$ and persists from 5 to 120 minutes). Our results are robust to five independent inference methods (Newey--West, wild bootstrap, clustered standard errors, quantile regression, and permutation tests) and are validated through pre-announcement placebo tests.2021The remainder of this paper proceeds as follows. Section~\ref{ch3:sec:data} describes data construction and processing. Section~\ref{ch3:sec:methodology} presents our theoretical framework, propositions, and empirical methodology, including our identification strategy and econometric specifications. Section~\ref{ch3:sec:results} reports our main empirical findings. Section~\ref{ch3:sec:conclusion} concludes with implications for monetary policy transmission and central bank communication strategy.22