spb/phd_thesis Public

1\section{Conclusion}2\label{sec:conclusion}34\subsection{Summary of Findings}56We summarize our findings under two headings: what we learn about the economics of monetary policy transmission, and what we contribute methodologically.78\textbf{Tone predicts directional returns.} Dovish tone is associated with higher equity returns that build to about 12 bps over two hours, and with lower safe-haven returns (gold: $-3.1$ bps; Treasuries: $-1.1$ bps at 15 minutes), supporting H1a--b. This is consistent with the standard transmission mechanism in \citet{bernanke2005} and \citet{gurkaynak2005}---accommodative expectations lower discount rates and shift portfolios toward risk assets---but we show that the channel operates through the \textit{qualitative} content of the statement, not only through the rate decision.910Tone effects grow stronger over time: the ES coefficient rises from 0.2 bps at $h = 5$ to 12.0 bps at $h = 120$ (H1c). \citet{rosa2013} documents continued price adjustment over hours after FOMC releases; our minute-level data show that this amplification is consistent with heterogeneous processing speeds across market participants \citep{veldkamp2011}.1112\textbf{Novelty predicts lower volatility, not higher.} Novelty is associated with \textit{reduced} implied volatility (VIX: $-5.57$ bps, $p < 0.05$; interaction $t = -5.06$), which reverses the sign we predicted under H2a. Prior work on disagreement-driven volatility \citep{patton2013} and news-based uncertainty \citep{manela2017} would predict the opposite. The difference, we believe, is that in the specific setting of FOMC statements, a departure from prior language acts as a \textit{signal of clarity}: the Fed is actively updating its message, which narrows the set of plausible policy paths. A repetitive statement, conversely, leaves open whether the Fed's views have changed but the language has not been updated. Novelty effects dissipate in under 10 minutes (H2b) and do not predict returns (H2c), confirming that this channel operates on volatility rather than on prices.1314\textbf{The composite surprise replicates known patterns.} Hawkish surprises are associated with 1--7 bps lower equity returns (H3a) and 11.2 bps higher VIX (H3b), with volatility effects concentrated in the first minutes (H3c). That our textual MPS produces results consistent with the rate-surprise literature \citep{kuttner2001,bernanke2005,gurkaynak2005,savor2014} reassures us that the NLP pipeline captures economically relevant variation.1516\textbf{Cross-asset patterns differ between returns and volatility.} Risk assets (ES, CL) respond to tone with the same sign and similar timing (H4a). Safe havens (GC, ZN, ZF) move in the opposite direction. But volatility responses vary: VIX loads on the interaction, Treasury volatility on stance alone, commodity volatility on both. This pattern is consistent with \citet{fleming1999} and \citet{balduzzi2001}, who document heterogeneous announcement effects across asset classes.1718\textbf{Volatility resolves fast; returns adjust slowly.} The VIX half-life is about 17.5 minutes; the ES tone effect continues to build for two hours after the announcement. A single-channel model cannot produce both patterns simultaneously. The fast channel (novelty $\to$ volatility) reflects the resolution of uncertainty about what the Fed will say. The slow channel (tone $\to$ returns) reflects the time needed to revise discount rates, growth expectations, and portfolio allocations across a heterogeneous investor base.1920\subsection{Methodological Contributions}2122On the measurement side, we make three contributions. First, using two architecturally different models (MiniLM, 33M parameters; BERT, 110M parameters) and comparing their outputs reduces the risk that any finding is an artifact of one particular architecture. The inter-model confidence (mean 0.837 for novelty) flags statements where the two models disagree. Second, two-stage domain adaptation (TSDAE then MNRL) yields axis separations of 0.248--0.302, well above the 0.04 minimum. Third, PCA-based reference selection replaces subjective date choices with an algorithm, removing a source of researcher discretion.2324Working at the minute level (148 events $\times$ $\pm$120 minutes) matters for two reasons: it allows causal identification from the pre-determined release time (validated by the placebo tests), and it reveals the fast-volatility / slow-return asymmetry that daily data would average away.2526\subsection{Implications for Policy and Communication Strategy}2728Three findings have implications for how the Fed drafts its statements.2930First, novelty reduces volatility. The conventional view is that central banks should change their language gradually to avoid ``surprising'' markets. Our estimates say the opposite: substantive changes in wording are associated with \textit{lower} implied volatility, presumably because a clear departure from the prior statement narrows the range of plausible interpretations. When the Fed needs to signal a regime change---a new framework, a pivot from tightening to easing---a decisive rewrite of the statement may be less destabilizing than incremental edits.3132Second, tone has persistent valuation effects. Because the equity response to tone grows over 2 hours rather than being absorbed immediately, the hawkish-dovish framing of the statement has real wealth consequences: a one-standard-deviation dovish shift is associated with up to 12 bps of equity returns, which, on a market capitalization of roughly \$40 trillion, is economically meaningful.3334Third, the interaction matters. The stance$\times$novelty interaction on VIX ($t = -5.06$, persistent from 5 to 120 minutes) implies that a directionally clear statement packaged in new language reduces uncertainty more than the same message delivered in boilerplate.3536\subsection{Limitations}3738We note several limitations.3940\textit{Sample period.} Our 2008--2025 sample is dominated by extraordinary monetary policy regimes: the zero lower bound (2008--2015), quantitative easing, and the post-pandemic tightening cycle. While our sub-period stability analysis (Figure~\ref{fig:rob_subperiod}) shows that coefficient signs are generally consistent across six Fed policy regimes, we cannot rule out that the magnitude of communication effects differs in more ``normal'' policy environments. In particular, the ZLB period may overstate the importance of qualitative communication (since rate decisions were constrained, statements became the primary policy instrument), while the recent tightening cycle may understate novelty effects (since rate hikes were widely anticipated, reducing the scope for genuine communication surprises).4142\textit{Measurement error.} Our NLP-based tone and novelty measures are subject to measurement error whose properties are difficult to characterize fully. Classical measurement error would attenuate our coefficient estimates toward zero (creating an ``errors-in-variables'' bias), making our significant findings conservative. However, if measurement error is correlated with meeting characteristics (e.g., if our models systematically misclassify statements from certain Fed Chairs or policy regimes), the bias could go in either direction. The inter-model agreement diagnostics (mean confidence 0.837 for novelty, 0.562--0.806 for tone) help identify statements where measurement may be less reliable, but they do not fully resolve this concern.4344\textit{Multiple testing.} With 7 assets, 5 dependent variables, and multiple inference methods, we test a large number of coefficients. We address this through Benjamini--Hochberg FDR correction within each family of tests and by reporting multi-method robustness counts. Nevertheless, some individually significant results---particularly those with $\#\text{Sig} = 1$ in the robustness tables---should be interpreted cautiously as potentially reflecting false discovery.4546\textit{Identification threats.} Our identification strategy exploits the pre-determined timing and content of FOMC statements, validated by pre-announcement placebo tests. However, we cannot fully rule out confounding from simultaneous information releases. FOMC announcements sometimes coincide with the Summary of Economic Projections (``dot plot''), which provides quantitative rate path forecasts that may interact with our textual measures. Similarly, market expectations about the subsequent press conference (beginning 30 minutes after the statement) may influence post-announcement price dynamics within our event window. These concurrent information flows are not fully separable from the statement text itself.4748\textit{Generalizability.} We study U.S. futures markets around FOMC statements only. Whether the same tone-return and novelty-volatility patterns hold for other central banks (ECB, BOJ, BOE), other Fed communications (minutes, speeches), or other asset classes (corporate bonds, emerging-market equities) is an open question.4950\subsection{Future Directions}5152Several extensions seem natural.5354\textit{More semantic dimensions.} We measure tone and novelty; FOMC statements also vary in uncertainty language, temporal focus (forward- vs.\ backward-looking), and internal consensus (voting dissents). The PCA axis construction can accommodate additional dimensions.5556\textit{International markets.} FOMC announcements move global asset prices \citep{wongswan2009,ehrmann2011}. Applying the tone-novelty decomposition to international data would show whether both channels transmit across borders or whether one dominates.5758\textit{Structural models.} Our evidence is reduced-form. A model with heterogeneous agents and differential processing of directional versus informational content could rationalize the fast-volatility / slow-return asymmetry and generate further predictions.5960\textit{Joint analysis with rate surprises.} Our textual MPS and the rate-based surprise of \citet{kuttner2001} measure different things. Estimating both jointly would reveal whether they contain complementary information and whether the information-versus-policy decomposition of \citet{jarocinski2020,nakamura2018} maps onto our tone-versus-novelty decomposition.6162