Do Risk-Related Words Predict Financial Market Responses? Evidence from Federal Reserve Press Conferences

Abstract

Federal Reserve press conferences convey policy path information and uncertainty beyond formal decisions. This study tests whether transcript-derived risk language predicts the magnitude of financial market responses after policy surprises and event characteristics enter the model. The dataset contains 93 official press conference transcripts from April 2011 to June 2026, with 90 scheduled events in the primary sample. A Q&A lexical risk index combines standardised frequencies of negative, uncertainty, and weak modal terms from the Loughran–McDonald dictionary. The outcome is an equal weight composite of absolute S&P 500 returns, two- and ten-year Treasury yield changes, and US dollar returns during a 70 min press conference window. OLS models use HC3 standard errors, asset-specific regressions, influence analysis, permutation testing, and leave-one-out cross-validation. Q&A lexical risk does not predict larger responses. The full-model coefficient equals −0.063 (p = 0.444), incremental R² equals 0.0034, and prediction error rises by 0.60% after adding the index. Policy surprise magnitude remains the strongest predictor. The 90-event sample limits precision for small effects, with an approximate 80% minimum detectable effect of 0.230 response index units. The estimates describe conditional association and incremental predictive content. They do not test acoustic delivery, realised intraday volatility, or a causal communication effect.

1. Introduction

Central bank communication has become an operating instrument of monetary policy. Policy decisions affect asset prices not only through unexpected changes in the current policy rate but also through information about the future policy path, the central bank’s assessment of the economy, and the uncertainty surrounding both. The distinction is economically important because an unchanged rate decision can still generate substantial repricing when the accompanying message alters expectations. Gürkaynak et al. (2005) show that Federal Open Market Committee (FOMC) announcements contain separate target and policy path factors, with the path component accounting for much of the explainable movement in longer-term yields. Swanson (2021) further distinguishes conventional policy, forward guidance, and asset purchase news. Communication is therefore part of the monetary policy shock rather than a neutral explanation delivered after the decision (Blinder et al., 2008).

Most empirical research initially treated communication as text. Statements, minutes, speeches, and press reports have been analysed for sentiment, policy stance, uncertainty, novelty, and disagreement. Hansen and McMahon (2016) distinguish forward guidance language from information about economic conditions. Ehrmann and Talmi (2020) find that material changes in central bank wording are associated with greater market volatility, while semantically similar statements are processed more smoothly. Jarociński and Karadi (2020) demonstrate that central bank announcements can combine monetary policy shocks with information shocks about the economic outlook. Cieslak and Schrimpf (2019) similarly show that non-monetary news is particularly prevalent in contextual forms of communication. Such findings make clear that market reactions cannot be attributed to tone or delivery unless policy news, economic information, and textual content are represented separately.

The post-meeting press conference creates a particularly informative setting. Prepared remarks remain tightly controlled, but the question-and-answer session requires the Chair to respond in real time, interpret incoming questions, acknowledge uncertainty, and reconcile the Committee’s position with rapidly changing conditions. Acosta et al. (2025) show that press conferences contain economically important news independent of the associated policy statements. Narain and Sangani (2026) document frequent reversals between statement window- and press conference-related market movements, together with substantially higher press conference-related equity volatility during the Powell era. Byun et al. (2026) find that statements provide comparatively coarse policy signals, while press conferences communicate more graduated information that is related to future policy rates and intraday asset price changes. Live interaction can therefore refine, qualify, or partially reverse the written message.

Live communication contains lexical, acoustic, and visual information. Gorodnichenko et al. (2023) report market relations for acoustic vocal emotion, while Marchal (2020) reports relations for visual behaviour. These studies motivate a strict separation between transcript-derived lexical measures and non-verbal measures. The present paper studies the lexical channel only. Acoustic and visual results therefore provide neighbouring evidence, not proxies for the variable estimated here.

Risk-related wording also has an ambiguous economic meaning. A response might reveal a new concern, acknowledge a known concern, qualify earlier guidance, or resolve uncertainty through clarification. A conference-level dictionary count therefore mixes communicative functions with different expected market effects. The empirical question in the present study asks whether the net lexical signal retains any association with cross-asset response magnitude after policy news and event characteristics enter the model.

Existing research leaves a narrower lexical question open. Prior studies document monetary policy surprises, semantic novelty, policy stance, uncertainty language, acoustic emotion, and visual behaviour. Less evidence tests whether a transparent risk-word measure adds event-level predictive content after policy surprise controls across several asset classes. Conference aggregation also raises a measurement issue because risk revelation and risk resolution might occur within the same Q&A session.

The study asks whether risk-related words in FOMC press conferences predict the magnitude of financial market responses after conventional monetary policy surprises and observed event characteristics enter the specification. The design compares prepared remarks with Q&A speech, separates lexical content from policy news, and evaluates in-sample fit together with out-of-sample prediction. The analysis does not treat transcript words as a measure of the Chair’s voice.

The dataset contains 93 official FOMC press conference transcripts from April 2011 to June 2026. The prespecified estimation sample contains 90 scheduled conferences. Chair responses are separated into opening remarks and Q&A speech. A transparent lexical risk index combines standardised frequencies of negative, uncertainty, and weak modal terms from the Loughran–McDonald financial dictionary (Loughran & McDonald, 2011). The principal dependent variable is an equal weight index of absolute S&P 500 returns, two-year and ten-year Treasury yield changes, and US dollar returns during the 70 min press conference window. Policy surprise magnitude, Chair word count, Summary of Economic Projections meetings, Chair tenure, the 2020 period, and calendar trend enter as controls. Predictive comparisons, influence diagnostics, alternative outcome constructions, and asset-specific models supplement HC3 inference.

Results from the lexical benchmark do not support the predicted positive relationship. The full-model coefficient on Q&A risk language is −0.063, with an HC3 standard error of 0.082 and a two-sided p-value of 0.444. Its 95% confidence interval ranges from −0.226 to 0.100. Adding the risk index increases in-sample R² by only 0.0034 and produces a partial R² of 0.0072. Leave-one-out cross-validation worsens prediction error by 0.60% relative to the control model. No asset-specific coefficient is statistically significant, and the conclusion remains unchanged under Newey–West inference, winsorisation, alternative composite measures, rank-based estimation, or restriction to audio-linked events. Policy surprise magnitude, rather than lexical risk, is the consistently important predictor of absolute market movement.

Prepared and unscripted communication also fail to display the expected ordering. A model containing both speech sections estimates a Q&A coefficient of −0.023 and an opening statement coefficient of −0.213. The difference between them is not statistically significant. Chair-specific lexical distributions differ substantially, but a risk-by-Powell interaction provides no evidence of a distinct Powell-era slope. Exploratory measures based on explicit risk terms, semantic novelty, and changes between opening remarks and Q&A do not survive false discovery rate correction.

The paper makes three contributions. First, it isolates a lexical construct and avoids labelling transcript counts as vocal signals. Second, it evaluates incremental information across four asset classes with effect sizes, confidence intervals, permutation testing, influence checks, and leave-one-out prediction. Third, it treats the null result as evidence of the limits of conference-level dictionary measures. The result sets a reproducible benchmark for later answer-level semantic or acoustic work without extending the conclusion beyond the measure observed here.

The contribution is also methodological. Communication studies often face a high-dimensional set of plausible signals, outcomes, windows, and interactions. The present framework separates confirmatory tests from exploratory tests, reports effect sizes and confidence intervals, applies false discovery rate correction, and evaluates predictive performance. Such discipline is important because isolated marginal associations can emerge easily in a sample of fewer than one hundred press conferences. Robustness across asset classes, feature constructions, inference methods, and held-out events provides a more demanding standard for claims of predictive content.

The literature-derived hypotheses are developed in Section 2. The remainder of the manuscript proceeds as follows. Section 2 reviews evidence on central bank communication, press conferences, textual information, and non-verbal signals. Section 3 describes the FOMC corpus, market data, variable construction, and empirical strategy. Section 4 presents the primary results, predictive tests, diagnostics, and robustness analysis. Section 5 discusses the implications of the lexical benchmark and specifies the answer-level acoustic extension. Section 6 provides the conclusion.

2. Literature Review and Hypothesis Development

2.1. Central Bank Communication as a Monetary Policy Instrument

Modern monetary policy operates through expectations as well as through changes in administered interest rates and balance sheet instruments. Communication can affect the expected path of short-term rates, beliefs about inflation and output, perceptions of the central bank’s reaction function, and the compensation investors demand for bearing risk. Blinder et al. (2008) describe communication as an integral part of policy implementation rather than a supplementary exercise in transparency. Their synthesis also identifies an enduring empirical difficulty: a central bank message can clarify policy and reduce uncertainty, but it can also reveal previously unknown information and generate immediate repricing. Lower volatility is therefore not an automatic indicator of effective communication, and higher volatility is not necessarily evidence of failure. The direction of the response depends on what the market learns and how far that information departs from prior beliefs.

High-frequency research established that the information released at Federal Open Market Committee (FOMC) events cannot be represented by the unexpected change in the current federal funds rate alone. Gürkaynak et al. (2005) identify separate target and path factors in intraday asset price changes around FOMC announcements. The target factor captures news about the current policy rate, while the path factor is closely associated with the accompanying statement and expectations of future policy. Their estimates indicate that the statement-related factor accounts for more than three-quarters of the explainable movement in five- and ten-year Treasury yields around FOMC meetings. Communication therefore moves markets partly because it changes the expected sequence of future decisions even when the current decision is fully anticipated.

Equity market evidence reinforces the economic importance of policy surprises. Bernanke and Kuttner (2005) estimate that an unexpected 25-basis-point reduction in the federal funds rate is associated, on average, with an equity market increase of approximately 1%. Their decomposition points towards changes in expected excess returns and, to a lesser extent, expected dividends and real interest rates. Later work extends the set of policy instruments that must be represented in event studies. Swanson (2021) separates conventional policy, forward guidance, and large-scale asset purchases, showing that the latter two dimensions have distinct effects on financial markets. Vocal signal research must consequently control for multiple forms of monetary news rather than use a single rate-surprise variable as a sufficient representation of the FOMC event.

Written communication also contains information about the central bank’s assessment of economic conditions. Hansen and McMahon (2016) use computational linguistics methods to distinguish statements about the expected policy path from statements about the economic outlook. Their results assign a larger role to forward guidance language in explaining yields, although the estimated effects on real variables are comparatively limited. Shapiro and Wilson (2022) take a different approach, using FOMC language to infer the relative weights placed on inflation and economic slack. Taken together, these studies show that market participants can learn both about prospective policy and about the policymaker’s objectives or economic assessment. An observed association between vocal delivery and market volatility may therefore represent policy path news, macroeconomic information, a change in perceived confidence, or some combination of these channels.

2.2. Written Language, Novelty, Sentiment, and Uncertainty

The textual literature demonstrates that market reactions depend on how a message is framed and how far it departs from established communication. Ehrmann and Talmi (2020) measure semantic similarity across successive central bank statements. More similar statements are absorbed with lower financial market volatility after controlling for content, while material changes in wording produce higher volatility. The result provides a direct connection between communicative novelty and the cost of interpretation. It also implies that a vocal risk coefficient can be biased if unusual delivery coincides with an unusually novel statement. Textual novelty should therefore enter the empirical design separately from hawkishness, dovishness, or general sentiment.

Text-based uncertainty measures supply a second set of relevant controls. Baker et al. (2016) construct an economic policy uncertainty index from newspaper coverage and validate it against human-coded articles and identifiable policy events. Broad policy uncertainty can influence both the language used by policymakers and the volatility prevailing before an FOMC event. Controlling for such uncertainty helps distinguish a genuinely event-specific vocal signal from the possibility that the Chair and the market are responding to the same difficult macroeconomic environment.

Banerjee et al. (2025) show that the financial press provides an informative interpretive layer between official communication and market response. Their policy-specific dictionaries distinguish conventional policy, quantitative easing, and forward guidance. Surprises in the resulting sentiment index explain movements across major asset classes beyond market-based monetary policy surprises, including during the post-COVID period. The finding is important for vocal analysis because investors do not necessarily react to raw acoustic information in isolation. Media interpretation can amplify, translate, or stabilise the signal, and the response may continue after the press conference as coverage circulates.

Recent language model research expands the range of semantic measures available for FOMC documents. Shah et al. (2023) introduce a large annotated dataset of FOMC speeches, minutes, statements, and press conference transcripts and formulate hawkish–dovish classification as a finance-specific task. Nakayama and Sawaki (2023) apply zero-shot textual entailment to statements, minutes, press conferences, and speeches, illustrating how transformer models can detect policy tone without relying solely on fixed dictionaries. Kim et al. (2023) compare VADER, FinBERT, and a trial application of GPT-4, reporting both improved domain sensitivity and persistent difficulties arising from formulaic, deliberately unemotional FOMC language. Peskoff et al. (2024) use GPT-based measurement to study hawkish–dovish disagreement and find that public statements suppress much of the diversity visible in minutes and transcripts. These results support the use of modern semantic controls, but they also warn against treating any single model-generated score as ground truth.

Transparency creates an additional institutional complication. Hansen et al. (2018) use a natural experiment and computational linguistics methods to examine how greater transparency altered FOMC deliberation. Their evidence is consistent with both discipline and conformity effects. Public communication may become more careful and uniform when participants expect disclosure, meaning that written text can understate internal disagreement or uncertainty. Vocal delivery during live questioning may be less completely standardised and may therefore reveal information not preserved in edited statements or delayed minutes.

2.3. The Press Conference as a Distinct Information Event

The post-meeting press conference is not a verbal restatement of the FOMC announcement. Its prepared opening remarks remain highly controlled, but the question-and-answer segment requires the Chair to respond in real time to unexpected prompts, reconcile the Committee’s position with personal interpretation, and explain uncertainty that may have been compressed in the written statement. The institutional separation between the statement window and the press conference window permits researchers to test whether the later event contains independent information.

Acosta et al. (2025) provide the most comprehensive recent infrastructure for such analysis through the US Monetary Policy Event-Study Database. Their database covers statements, minutes, and press conferences and includes a broad set of risk-free rates and risky asset prices. Their results show that press conferences contain important monetary news independent of the associated statements. Conference surprises have strong effects on Treasury yields and risky assets, and excluding them can omit a substantial part of the market’s response to FOMC communication. The result directly supports separate outcome windows for the statement and the press conference.

Narain and Sangani (2026) document a marked change under Chair Jerome Powell. Squared S&P 500 returns during Powell’s press conferences are more than three times their average level during the Bernanke and Yellen conferences despite similar volatility around the preceding statement release. Reversals are especially informative. Market movements during the press conference often run in the opposite direction to the initial response to the statement. Textual analysis links these reversals partly to language that departs from the statement, while Treasury option evidence suggests weaker resolution of uncertainty about the future path of rates. Their interpretation remains appropriately qualified because greater flexibility and clearer recognition of uncertainty may be valuable during rapidly changing conditions even when volatility increases.

Byun et al. (2026) reach a related conclusion using a finance-specific language model. FOMC statements often provide coarse hawkish, neutral, or dovish signals, whereas press conferences contain a more graduated mixture of policy language. The additional variation is correlated with future federal funds rates and with intraday asset price changes that reverse the statement window-related movement. Press conferences therefore fine-tune the policy message and help investors revise expectations about future policy. A vocal signal should be evaluated against that granular semantic information rather than against statement sentiment alone.

Cross-speaker evidence shows that consistency within the institution also affects transmission. Djourelova et al. (2025) analyse 481 time-stamped speeches by FOMC members and measure their textual alignment with the Chair’s recent communication. Average market responses to individual speeches are modest, but the conditional pattern is pronounced. Closely aligned speeches raise short-term rates, reduce breakeven inflation, and transmit like conventional monetary policy shocks. Divergent speeches lift the yield curve, inflation expectations, and equity prices, suggesting that investors read them as news about economic fundamentals rather than as a straightforward policy signal. Coordination strengthens monetary transmission, while visible disagreement can dilute it.

Schmanski et al. (2023) broaden the communication system further by examining official messages alongside news and social media discussion. Their framework draws attention to propagation and disagreement outside the immediate release window. Bauer and Wasserburger (2026) similarly emphasise that surprises around FOMC statements and post-meeting press conferences affect inflation expectations. The implication for the present study is that vocal delivery can influence volatility directly within the event window and indirectly through subsequent interpretation. Intraday outcomes provide the cleanest initial test, while longer windows can capture slower processing at the cost of greater exposure to confounding news.

2.4. Vocal and Non-Verbal Information

Gorodnichenko et al. (2023) provide the closest empirical foundation for analysing Federal Reserve voice. They apply deep learning to audio from the Q&A portions of FOMC press conferences and classify emotional tone using a large set of acoustic features, including Mel-frequency cepstral coefficients, chromagram information, and Mel-scaled spectral measures. Their design controls for policy actions and textual sentiment. A more positive vocal tone is associated with higher share prices, lower perceived interest rate risk, lower expected volatility, lower inflation expectations, and exchange rate movements. High-frequency estimates show an immediate stock price response at answer level, while part of the broader reaction develops more slowly as investors and media interpret the signal. Bond market responses are less uniform.

The economic interpretation of vocal positivity is not unique. A confident or positive delivery may operate as informal forward guidance by indicating a low probability of near-term tightening. It may communicate favourable private information about economic conditions. It may also reduce ambiguity by increasing perceived confidence in the policy path. Gorodnichenko et al. (2023) explicitly recognise these competing explanations. Their evidence establishes that voice contains incremental market-relevant information, but it does not convert vocal tone into a pure structural monetary policy shock.

Marchal (2020) provides complementary evidence from video rather than audio. Computer vision methods identify how frequently the Chair consults internal documents while answering reporters, producing an attention-based measure of discussion complexity. More complex discussions are associated with higher equity returns and a reduction in realised volatility from before to after the conference. The result is consistent with uncertainty resolution. Difficult questions can initially signal complexity, yet a substantive answer can release valuable information and lower residual uncertainty. Non-verbal intensity should therefore not be mechanically interpreted as risk increasing. The market response depends on whether the behaviour exposes unresolved uncertainty or helps resolve it.

Ahrens et al. (2025) extend communication research beyond scheduled FOMC meetings. Their time-stamped dataset of Federal Reserve speeches and supervised multimodal language model maps speech content into implied revisions of forecasts for inflation, output, and unemployment. Speech-implied forecast revisions explain realised volatility and tail risk in equity and bond markets. Chair speeches generate larger revisions and are unconditionally associated with stronger volatility and tail risk responses. Evidence that hawkishness systematically changes these effects is weaker and appears regime dependent. The paper demonstrates that market risk responds to the economic information embedded in speech, but it also underscores the need to separate semantic forecast news from acoustic delivery.

The combined evidence establishes three points. First, live communication contains information that is absent from written policy documents. Second, non-verbal cues can have economically meaningful effects after textual content and policy actions are controlled. Third, the sign of the volatility response is conditional rather than universal. Positive vocal tone can reduce perceived risk, complex interaction can resolve uncertainty, and forecast revisions can increase or decrease volatility depending on their content and the macroeconomic regime. A risk-specific vocal construct must capture the relevant acoustic state without assuming that all deviations from neutral delivery are destabilising.

2.5. Economic Mechanisms Linking Risk Communication to Market Responses

Four mechanisms organise the empirical evidence. The first is a policy path channel. Vocal tension, hesitation, or reduced confidence may cause investors to widen the distribution of expected future policy rates even when the modal path is unchanged. Gürkaynak et al.’s (2005) path factor, Swanson’s (2021) forward guidance factor, and the press conference evidence of Byun et al. (2026) all show that information about future policy can dominate news about the current rate decision. A vocal risk signal may alter the precision of the communicated path rather than its average hawkish or dovish direction.

The second is a central bank information channel. Jarociński and Karadi (2020) demonstrate that an announcement simultaneously conveys policy news and information about the central bank’s economic outlook. Their identification uses the sign of high-frequency interest rate and equity price comovement. A tightening shock raises rates and lowers equities, while favourable information raises both. Cieslak and Schrimpf (2019) likewise find that non-monetary news is present in approximately 40% of Fed and European Central Bank policy announcements and is even more prevalent in contextual communications such as press conferences. Vocal risk may therefore reveal concern about growth, inflation, financial stability, or model uncertainty rather than a planned policy action.

The third is an uncertainty resolution channel. Similar written statements are processed with less volatility, while material wording updates are harder to absorb (Ehrmann & Talmi, 2020). Complex Q&A interactions can nevertheless reduce realised volatility when they provide information that resolves previously unsettled questions (Marchal, 2020). Vocal arousal may increase volatility when it signals unresolved concern, but it may reduce volatility when it accompanies a candid and informative explanation. The net coefficient can be small if these opposing cases are aggregated at conference level.

The fourth is a credibility and coordination channel. A delivery style that appears inconsistent with the literal content may weaken the credibility of the message. Similarly, a speech that diverges from the Chair’s prior communication changes how markets classify the shock (Djourelova et al., 2025). The economically relevant variable may therefore be acoustic–semantic congruence rather than vocal tone alone. A calm delivery attached to risk-heavy language can reassure investors, while a strained delivery attached to reassuring words can indicate concealed uncertainty. Such mismatch offers a plausible explanation for why purely lexical risk measures or conference-level average tone may have limited predictive power.

2.6. Measurement of Lexical and Non-Verbal Communication

Vocal risk measurement must remain conceptually separate from textual sentiment. Acoustic features such as pitch level and dispersion, intensity, speaking rate, pause frequency, jitter, shimmer, spectral balance, and harmonics-to-noise characteristics describe delivery. They do not possess invariant economic meaning on their own. Speaker anatomy, age, microphone placement, room acoustics, compression, illness, and recording technology can shift these measures independently of policy concern. Speaker standardisation, recording quality controls, meeting fixed effects where feasible, and sensitivity tests across acoustic feature families are necessary before interpreting a composite index.

The deep learning approach of Gorodnichenko et al. (2023) shows the value of learning nonlinear relationships across many acoustic features. Its use of general emotion-recognition datasets also identifies a limitation relevant to the proposed study. Generic categories such as happy, sad, angry, fearful, or neutral are not identical to economically meaningful vocal risk. A Chair can sound serious without conveying uncertainty, and a technically calm answer can introduce highly destabilising information. Construct validity improves when acoustic features are calibrated against manually coded FOMC answers that distinguish concern, uncertainty, confidence, urgency, and ordinary formality.

Semantic controls should also be plural rather than model dependent. Dictionary-based risk counts are transparent but can miss negation, conditionality, and domain-specific context. Finance-tuned transformers improve contextual classification but may be unstable across prompts, training periods, or document types. The evidence in Shah et al. (2023), Nakayama and Sawaki (2023), Kim et al. (2023), and Peskoff et al. (2024) supports a validation strategy that compares dictionary, transformer, and human-coded measures. Agreement across methods strengthens interpretation. Disagreement should be reported rather than hidden through a single composite score.

Granularity is equally important. Conference averages combine prepared remarks, straightforward questions, difficult follow-ups, and answers addressing different economic topics. Gorodnichenko et al. (2023) estimate answer-level responses and also examine cumulative information during the conference. Their distinction is well suited to vocal risk. Answer-level models preserve local changes in delivery and permit alignment with minute-by-minute returns or realised variance. Conference-level models provide a lower-noise summary but risk cancelling episodes with opposite effects. A hierarchical design can use answers nested within conferences, with conference, Chair, topic, and time controls.

2.7. Identification and Market Response Outcomes

Narrow event windows reduce reverse causality because the policy decision precedes the market response, but they do not automatically identify the content of the shock. Jarociński and Karadi (2020) show that policy and information shocks can occur in the same announcement. Cieslak and Schrimpf (2019) further separate monetary, growth, and risk premium news. Bauer and Swanson (2023) document that high-frequency monetary policy surprises are correlated with macroeconomic and financial information available before the event. Their Fed-response-to-news interpretation implies that market participants may underpredict how strongly the FOMC reacts to public information. Orthogonalising surprises against pre-event data changes macroeconomic estimates materially, although high-frequency asset price estimates remain comparatively stable.

The identification problem is particularly relevant for vocal risk. Difficult macroeconomic conditions can generate both a strained delivery and high market volatility. A credible design should condition on the pre-conference volatility state, scheduled macroeconomic news, policy surprise factors, statement window returns, textual policy stance, textual risk, semantic novelty, and the economic topics discussed. Chair and era controls are also necessary because communication practice, press conference frequency, and recording technology differ across Bernanke, Yellen, and Powell.

Event window construction requires similar care. Lucca and Moench (2015) document a sizeable pre-FOMC equity drift, showing that FOMC day returns are not confined to the announcement itself. Acosta et al. (2025) separate statement, minutes, and press conference surprises and demonstrate that each event carries distinct information. Statement window outcomes should not be mixed with press conference outcomes when voice is the treatment of interest. A clean design can use statement window movements as controls and measure volatility from the beginning of the Chair’s spoken communication, ideally at answer or sub-answer frequency.

Realised volatility, implied volatility, absolute returns, squared returns, and tail risk measures capture different responses. Narain and Sangani (2026) focus on squared intraday returns and Treasury implied volatility. Marchal (2020) studies changes in realised equity volatility before and after the conference. Ahrens et al. (2025) examine realised volatility and tail risk across equity and bond markets. Gorodnichenko et al. (2023) analyse stock prices, the VIX and VIX futures, inflation expectations, exchange rates, and bond market variables. A multi-outcome design is therefore preferable, but the primary outcome and horizon should be prespecified to limit specification search.

Broader comparative evidence cautions against assuming that findings transfer mechanically across institutions. Aguilar and Pérez-Cervantes (2022) use natural language processing and market data to study communication in Mexico, while Goodhead and Kolb (2025) separate communication shocks from conventional policy actions and trace their macroeconomic consequences. Institutional credibility, communication conventions, market depth, and press conference design can all affect the response. Focusing on Federal Reserve communications improves internal consistency, although it narrows external validity.

2.8. Research Gap and Contribution

Prior evidence shows that FOMC policy news, textual stance, semantic novelty, acoustic emotion, and visual behaviour relate to financial market outcomes. A narrower empirical issue remains open: does conference-level risk-related language add information about absolute market response magnitude after policy surprise controls? The question differs from tests of realised volatility and from tests of acoustic delivery.

Three points motivate the design. Risk dictionaries offer transparent replication and weak contextual interpretation. Conference averages combine answers with different communicative functions. Small event samples also make in-sample associations vulnerable to influential meetings and specification search. The present study therefore gives priority to construct clarity, policy surprise controls, effect-size reporting, and out-of-sample prediction.

The present study addresses the lexical part of the problem. It constructs a conference-level Q&A lexical risk index, compares it with an opening statement index, controls for press conference policy surprises and event characteristics, and evaluates predictive performance. Answer-level acoustic features, sentence-level semantic coding, realised volatility, and acoustic–semantic disagreement fall outside the reported estimates.

The coefficient therefore has an associational interpretation. Policy surprises, Chair effects, event controls, and narrow windows reduce several sources of confounding. They do not identify a causal effect of wording. Omitted pre-event volatility, same-day macroeconomic news, statement window reactions, and topic-specific information remain plausible common causes.

2.9. Study Hypotheses

H1. 

Higher Q&A lexical risk intensity is associated with larger cross-asset market response magnitude during the press conference window.

Risk-related wording might accompany new information, uncertainty, or larger policy path revisions. The directional hypothesis provides a direct test of the lexical measure used in the reported models.

H2. 

Q&A lexical risk intensity adds explanatory and out-of-sample predictive information beyond policy surprise magnitude and the prespecified event controls.

The hypothesis concerns incremental information after policy news enters the model. In-sample R², partial R², permutation inference, and leave-one-out prediction provide separate tests.

H3. 

The association between lexical risk and market response magnitude is stronger for unscripted Q&A speech than for prepared opening remarks.

Q&A answers involve spontaneous clarification and a wider range of topics. A joint model containing the Q&A and opening statement indices tests the proposed difference directly.

2.10. Synthesis of the Closest Empirical Evidence

Table 1. Closest empirical evidence and relevance to the present study.

Study	Communication Measure	Setting and Outcome	Main Result	Relevance to the Present Study
Gürkaynak et al. (2005)	Target and path surprise factors	Intraday US yields and equities	Statements contain a distinct path signal with strong long-yield effects	Requires policy path controls
Ehrmann and Talmi (2020)	Semantic similarity and wording changes	Central bank releases and volatility	Novel wording raises volatility, while similar releases are absorbed more easily	Requires semantic novelty controls
Jarociński and Karadi (2020)	Rate equity sign restrictions	High-frequency announcements and macroeconomic responses	Policy shocks and central bank information shocks have different effects	Prevents a single-shock interpretation
Cieslak and Schrimpf (2019)	Monetary, growth, and risk premium news	Four major central banks	Non-monetary news is common, especially in contextual communications	Motivates information and risk premium controls
Gorodnichenko et al. (2023)	Deep learning measure of vocal emotion	FOMC Q&A audio and multiple markets	Positive voice affects equities and perceived risk beyond text and policy actions	Establishes incremental vocal information
Marchal (2020)	Video-based attention and discussion complexity	FOMC Q&A and equity outcomes	Complex discussions are associated with higher returns and lower realised volatility	Shows that non-verbal intensity can resolve uncertainty
Ahrens et al. (2025)	Speech-implied forecast revisions	Time-stamped Fed speeches, realised volatility, and tail risk	Forecast revisions explain equity and bond risk, with larger Chair effects	Requires semantic forecast controls
Acosta et al. (2025)	High-frequency surprises by communication event	Statements, minutes, press conferences, rates, and risky assets	Press conferences contain independent and economically important news	Supports separate press conference windows
Banerjee et al. (2025)	Financial press sentiment surprises	Scheduled FOMC meetings and major asset classes	Media sentiment adds explanatory power beyond market surprises	Identifies an interpretation and propagation channel
Narain and Sangani (2026)	Intraday returns and conference-text departures	FOMC statements and press conferences	Powell conferences generate much higher volatility and frequent reversals	Motivates Chair effects and text–delivery comparison
Byun et al. (2026)	Language model sentiment	FOMC statements, press conferences, and future rates	Press conferences refine coarse statements and predict future policy	Requires granular conference semantics
Djourelova et al. (2025)	Speech alignment with Chair communication	FOMC-member speeches and market surprises	Alignment strengthens policy transmission, while divergence resembles fundamentals news	Supports acoustic–semantic congruence tests

synthesises the closest empirical evidence and shows how prior findings on policy signals, semantic content, vocal delivery, and press-conference dynamics inform the empirical design of the present study.

3. Materials and Methods

3.1. Study Design and Evidential Scope

The dataset supports an event-level analysis of lexical risk derived from official press conference transcripts and absolute financial market responses in a 70 min window. The reported variables do not include pitch, intensity, speech rate, pauses, disfluency, jitter, shimmer, or voice quality. Sixty-nine events have links to official recordings without speaker-segmented acoustic measurements in the analysis archive.

The paper therefore tests lexical risk and market response magnitude. It does not test acoustic vocal risk or realised intraday volatility. Terms such as lexical risk, market response magnitude, and absolute market movement describe the reported evidence throughout the revised text.

The principal empirical result is a bounded null. Q&A lexical risk does not add detectable explanatory or predictive information after policy surprise magnitude and event characteristics enter the model. The coefficient remains imprecise, the confidence interval spans negative and small positive values, and the 90-event sample offers limited sensitivity to small effects.

3.2. Event Universe

The event panel contains 93 Federal Reserve press conferences or closely related communications from 27 April 2011 to 17 June 2026. The corpus includes 93 official transcript files in both PDF and text form. Official recording links were matched to 69 events.

The prespecified main sample contains 90 scheduled press conferences. Two emergency communications in March 2020 and one Kevin Warsh event are excluded from the primary specification because their institutional setting differs from the regular press conference process. The main sample ends on 29 April 2026 because the final event in the full panel was not part of the scheduled estimation sample at the time of construction. Sixty-eight main-sample events have matched official audio links.

3.3. Speech Segmentation and Lexical Measures

Each transcript was divided into the Chair’s prepared opening statement and question-and-answer responses. Risk language measures were computed only from the Chair’s words. Counts use the Loughran–McDonald financial dictionary (Loughran & McDonald, 2011) and are normalised per 1000 words.

The primary Q&A risk index is

RiskQA,i = [z(NegativeQA,i) + z(UncertaintyQA,i) + z(WeakModalQA,i)]/3

An analogous index was constructed for the opening statement. Standardisation was performed across events before averaging. A separate explicit risk measure counts direct terms such as risk, risks, uncertain, uncertainty, crisis, and recession. The most frequent explicit terms in the full corpus were RISKS (408 occurrences), RISK (403), UNCERTAINTY (217), CRISIS (190), UNCERTAIN (116), and RECESSION (111).

Dictionary counts provide transparent replication and limited semantic resolution. The index classifies terms by dictionary category and frequency. It does not infer whether a sentence negates a risk, places it in a conditional clause, refers to a past episode, or describes risk resolution. Expressions such as “risks have increased”, “risks have diminished”, and “risks are contained” therefore need not receive economically distinct treatment when the same risk terms appear. The study treats the index as a lexical frequency measure, not a sentence-level interpretation of communicative function.

3.4. Market Response Measures

Market data come from the US Monetary Policy Dataset press conference window. Four absolute responses are used:

• absolute S&P 500 return;
• absolute two-year US Treasury yield change;
• absolute ten-year US Treasury yield change;
• absolute US dollar index return.

The primary outcome is an equal weight composite:

MarketResponsei = [z(|SP500i|) + z(|UST2Yi|) + z(|UST10Yi|) + z(|DXYi|)]/4

Each component captures the response magnitude over the 70 min press conference window. Yield changes are expressed in percentage points in the source data. The composite avoids selecting a single asset after inspecting the results and captures the common magnitude of the cross-asset response. Principal component analysis provides an alternative data-driven aggregation.

3.5. Policy Surprise and Control Variables

The magnitude of the monetary policy surprise is calculated from the first two press conference window surprise factors:

PolicySurprisei = √(MP1i² + MP2i²)

The measure is standardised for regression analysis. The full control vector contains policy surprise magnitude, the Chair’s word count, an indicator for Summary of Economic Projections meetings, Chair indicators for Janet Yellen and Jerome Powell, a 2020 indicator, and a standardised linear time trend. Ben Bernanke forms the omitted Chair category.

The main regression is

MarketResponse_i = α + β RiskQA,_i + γ PolicySurprise_i + δ′X_i + ε_i

The coefficient of interest is β. A positive estimate aligns with H1. X_i contains the Chair word count, the Summary of Economic Projections indicator, Yellen and Powell indicators, the 2020 indicator, and the standardised time trend. Ben Bernanke forms the omitted Chair category. The coefficient describes conditional association after the listed controls. The design does not identify a causal effect of lexical wording. Figure 1 presents the most frequent explicit financial risk terms identified in the Chair’s Q&A responses, showing the relative prevalence of uncertainty, risk, and economic-condition language across the corpus.

3.6. Estimation, Inference, and Validation

The analysis estimates nested ordinary least squares specifications. M1 contains the Q&A lexical risk index alone. M2 adds policy surprise magnitude and Chair word count. M3 adds the Summary of Economic Projections indicator and Chair indicators. M4 adds the 2020 indicator and a standardised linear time trend. M5 includes both the Q&A and opening statement lexical risk indices. HC3 standard errors form the primary inference procedure because the event sample contains 90 scheduled conferences and several influential observations. Newey–West covariance estimates with one and four lags provide sensitivity checks.

Robustness tests exclude 2020, restrict the sample to audio-linked events, truncate the sample through 2025, remove the five largest Cook’s distance observations, restore emergency events, and winsorise the outcome. Alternative models use principal component, log-scaled, and rank-based outcomes. Leave-one-out cross-validation compares the control model with the model containing Q&A lexical risk. A 5000-draw permutation test evaluates incremental performance under random reassignment of the risk signal. False discovery rate adjustment applies to exploratory signal families. All reported tests use two-sided p-values.

3.7. Pre-Event Volatility Environment and Regime Specification

Pre-event volatility represents a plausible common driver of risk-related language and market response. Daily VIXCLS from FRED provides a public measure of near-term option-implied equity volatility (Chicago Board Options Exchange, n.d.). A reproducible regime extension follows Tong’s (1983) threshold autoregressive framework. The daily VIX series is fitted with a three-regime self-exciting threshold autoregression. Two estimated thresholds classify the previous trading day into low-, medium-, or high-volatility states before each press conference.

The event-level extension uses regime indicators and interactions with Q&A lexical risk:

MarketResponse_i = α + β RiskQA,_i + γ PolicySurprise_i + δ′X_i + λM DM,_i + λH DH,_i + θM(RiskQA,_i × DM,_i) + θH(RiskQA,_i × DH,_i) + ε_i.

Residual distributions from M4 are also compared across the three states. The current reproducibility archive does not contain the merged event-level VIX field, so the reported results do not include regime coefficients. The paper therefore treats pre-event volatility as an unmeasured contextual factor and avoids attributing the observed association to a causal communication channel.

3.8. Reproducibility

The analysis uses the event panel, transcript corpus, audio manifest, US Monetary Policy Event-Study Database workbook, and Loughran–McDonald dictionary supplied with the project. Core estimates were regenerated from the project analysis script. Extended diagnostics and sensitivity tests were saved as machine-readable CSV and JSON outputs. Numerical values are rounded for presentation. The VIX regime specification in Section 3.7 requires a separate event-level merge before estimation.

4. Results

4.1. Descriptive Evidence

4.1.1. Event-Level Distributions

The average absolute two-year and ten-year yield changes correspond to approximately 3.21 and 2.65 basis points. A small number of large events create right-skewed response distributions. The maximum composite response is more than 4.6 sample standard deviations above its mean, motivating robust inference and sensitivity analysis. See

Table 2. Event-level descriptive statistics for the scheduled press conference sample.

Measure	Mean	Median	Standard Deviation	Minimum	Maximum
Composite market response index	−0.059	−0.234	0.638	−0.797	2.912
Absolute S&P 500 return, %	0.480	0.335	0.485	0.005	2.481
Absolute two-year yield change, percentage points	0.0321	0.0200	0.0354	0.0000	0.1826
Absolute ten-year yield change, percentage points	0.0265	0.0209	0.0210	0.0000	0.1059
Absolute DXY return, %	0.266	0.238	0.198	0.002	1.117
Q&A risk index	0.038	−0.150	0.815	−1.638	2.061
Q&A negative terms per 1000 words	13.919	13.226	3.730	7.699	28.895
Q&A uncertainty terms per 1000 words	10.965	10.499	3.106	4.967	20.174
Q&A weak modal terms per 1000 words	5.185	4.799	1.837	1.793	11.017
Policy surprise magnitude	0.0131	0.0075	0.0163	0.0000	0.0901
Chair word count	6871.8	6890.5	975.7	5079	9370

below.

4.1.2. Chair and Era Differences

Q&A risk distributions differ strongly across Chairs according to a Kruskal–Wallis test (H = 37.18, p < 0.001). The pattern may reflect differences in verbal style, economic conditions, transcription conventions, or the evolution of the press conference institution. Chair indicators and a time trend are therefore substantively important controls rather than optional additions. See

Table 3. Chair-level means for lexical risk, market response, and policy surprise magnitude.

Chair	Events	Mean Q&A Risk Index	Mean Market Response Index	Mean Policy Surprise Magnitude
Ben Bernanke	12	0.669	−0.333	0.0060
Janet Yellen	16	0.967	−0.138	0.0058
Jerome Powell	64	−0.353	0.089	0.0167

below. Furthermore, Figure 2 presents the distribution of Q&A risk language and market responses by Chair, highlighting clear differences in lexical risk profiles and the dispersion of market reactions across Federal Reserve leadership periods.

4.1.3. Bivariate Associations

The Q&A risk index has a weak negative Pearson correlation with the composite response (r = −0.161). Policy surprise magnitude is much more strongly related to the response (r = 0.605). Risk language is negatively correlated with policy surprise magnitude (r = −0.221) and Chair word count (r = −0.306). The four absolute asset responses correlate strongly with the composite, with correlations ranging from approximately 0.73 to 0.91.

The bivariate risk coefficient should not be interpreted causally. Its attenuation after policy surprise controls are introduced shows that communication content and policy news are not empirically separable without an explicit policy shock adjustment. Figure 3 traces the standardised Q&A risk language and market response indices across scheduled press conferences, showing limited co-movement overall alongside several episodes of pronounced market reaction.

4.2. Primary Regression Results

4.2.1. Nested Models

M1 produces a modest negative association that is not conventionally significant. Adding policy surprise magnitude reduces the coefficient by almost two-thirds in absolute value. The full M4 estimate is −0.063 response index units for a one-unit increase in the Q&A risk index. A one-sample standard deviation increase in the predictor corresponds to only −0.051 response index units. Its 95% confidence interval ranges from −0.226 to 0.100 per index unit.

Table 4. Nested regression models for the composite market response index.

Specification	Q&A Risk Coefficient	HC3 SE	p-Value	95% CI	R2	Adjusted R2
M1: risk only	−0.126	0.073	0.088	[−0.271, 0.019]	0.026	0.015
M2: policy surprise controls	−0.047	0.069	0.498	[−0.184, 0.090]	0.375	0.353
M3: Chair and event controls	−0.071	0.079	0.374	[−0.228, 0.087]	0.381	0.336
M4: full specification	−0.063	0.082	0.444	[−0.226, 0.100]	0.399	0.340
M5: Q&A and opening indices	−0.023	0.073	0.751	[−0.169, 0.123]	0.442	0.379

reports the nested regression estimates for the composite market response index, showing how the Q&A risk coefficient changes following the sequential inclusion of policy surprise, Chair and event controls, and opening indices.

The primary estimate does not support H1. The 95% confidence interval spans a moderate negative association and a small positive association. The approximate minimum detectable effect for 80% power equals 0.230 response index units. The sample therefore gives limited precision for small lexical effects. The estimate does not speak to acoustic delivery or realised volatility. Figure 4 illustrates the partial association between residualised Q&A risk language and the composite market response index after event and Chair controls, with the fitted line showing a weak negative relationship.

Policy surprise magnitude is the consistently important predictor. Its M4 coefficient is 0.368 with p < 0.001. A one standard deviation increment in policy surprise magnitude is associated with a 0.368-unit increase in the composite market response index, conditional on the remaining controls.

4.2.2. Incremental Explanatory Content

The control model has R² = 0.396. Adding the Q&A risk index raises R² to 0.399, an increment of only 0.0034. The risk coefficient’s partial R² is 0.0072. Both quantities indicate that the transcript-based risk index explains very little additional event-level variation once policy and event characteristics are included.

4.2.3. Prepared Remarks Versus Q&A

M5 includes both speech sections. The Q&A coefficient is −0.023 (p = 0.751), while the opening statement coefficient is −0.213 (p = 0.070). A direct test of equality gives

$$\hat{\beta }QA-\hat{\beta }Opening=0.190$$

with a standard error of 0.122 and p = 0.124. Evidence therefore does not support a stronger Q&A effect. The more negative opening estimate is suggestive at most and should not be treated as confirmation of a section-specific relationship.

4.3. Asset-Specific Outcomes

No asset-specific risk coefficient is statistically significant. Three estimates are negative and the two-year yield estimate is effectively zero. Policy surprise magnitude remains positive and statistically significant in every asset model, with coefficients of 0.209 for equities, 0.605 for two-year yields, 0.249 for ten-year yields, and 0.409 for the dollar index.

Isolated control coefficients should be interpreted cautiously because many coefficients are examined. The 2020 indicator is negative in the two-year model, and word count is negative in the ten-year model, but neither result forms part of a prespecified cross-asset hypothesis. See

Table 5. Asset-specific estimates under the full control specification.

Dependent Variable	Q&A Risk Coefficient	HC3 SE	p-Value	95% CI	Adjusted R2
Absolute S&P 500 response, standardised	−0.094	0.066	0.162	[−0.225, 0.038]	0.151
Absolute two-year yield response, standardised	0.010	0.109	0.930	[−0.208, 0.227]	0.545
Absolute ten-year yield response, standardised	−0.120	0.085	0.164	[−0.289, 0.050]	0.212
Absolute DXY response, standardised	−0.049	0.136	0.722	[−0.320, 0.223]	0.154

below.

4.4. Robustness and Sensitivity Analysis

4.4.1. Sample and Influence Checks

The five largest Cook’s distance observations occur on 2 November 2022, 15 June 2022, 1 February 2023, 19 June 2019, and 29 October 2025. Removing them changes the sign but not the substantive conclusion. Adding the emergency events also changes the sign. Such sensitivity argues against a stable directional effect rather than supporting either a positive or negative association. See

Table 6. Robustness tests for the Q&A risk language coefficient.

Robustness Specification	Events	Q&A Risk Coefficient	HC3 SE	p-Value	Interpretation
Main scheduled sample	90	−0.063	0.082	0.444	Primary null estimate
Exclude 2020	83	−0.062	0.086	0.469	Nearly unchanged
Audio-linked events only	68	−0.093	0.105	0.380	No evidence in future acoustic subsample
Events through 2025	87	−0.077	0.084	0.363	Similar conclusion
Exclude five influential events	85	0.029	0.061	0.643	Sign reverses and remains null
Include emergency events	92	0.075	0.216	0.730	Sign reverses and uncertainty widens
Winsorised response outcome	90	−0.051	0.077	0.511	Outlier treatment does not rescue effect

below.

4.4.2. Alternative Inference

Inference remains unchanged under short-lag heteroskedasticity and autocorrelation-consistent covariance estimates. See

Table 7. Sensitivity to alternative covariance estimators.

Covariance Estimator	Coefficient	SE	p-Value
HC3	−0.063	0.082	0.444
Newey–West HAC, 1 lag	−0.063	0.073	0.390
Newey–West HAC, 4 lags	−0.063	0.067	0.350

below.

4.4.3. Alternative Variable Construction

The market response principal component explains 69.9% of the variance across the four absolute asset responses. Its loadings are positive and broadly balanced: 0.450 for equities, 0.542 for two-year yields, 0.497 for ten-year yields, and 0.508 for the dollar. The lexical risk principal component explains 68.5% of the variance in negative, uncertainty, and weak modal frequencies. None of the alternative constructions yields evidence of a risk effect. See

Table 8. Alternative constructions of the outcome and risk predictor.

Alternative Model	Risk Coefficient	p-Value
Market response first principal component	−0.123	0.352
Log-scaled response magnitude	−0.079	0.485
Rank-based response model	−0.026	0.826
First principal component of lexical risk categories	−0.045	0.504

below.

4.4.4. Chair Tenure Heterogeneity

The Q&A-risk-by-Powell interaction is small and non-significant. The baseline risk coefficient is −0.044 (p = 0.613), and the incremental Powell-era coefficient is −0.036 (p = 0.830). The data therefore do not support a different relationship during the Powell tenure, despite clear differences in the unconditional distribution of risk language across Chairs.

4.5. Predictive Validation

Leave-one-out cross-validation evaluates whether adding risk language helps predict an unseen event. See

Table 9. Leave-one-out predictive performance.

Model	LOOCV RMSE	Out-of-Sample R2
Controls only	0.5341	0.2913
Controls plus Q&A risk	0.5373	0.2828

below.

Adding the Q&A risk index worsens root-mean-square prediction error by 0.60% and reduces out-of-sample R² by 0.0085. A 5000-draw permutation test gives p = 0.438. The index offers no detectable incremental predictive value at the event level.

Out-of-sample performance defines the predictive claim directly. The model containing Q&A lexical risk has higher leave-one-out RMSE and lower out-of-sample R² than the control model. The present measure therefore lacks detectable incremental predictive content in the 90-event sample.

4.6. Diagnostics

Residuals are strongly non-normal according to the Jarque–Bera test (JB = 375.58, p < 0.001), mainly because a few policy events generate very large market responses. HC3 inference, winsorisation, rank-based estimation, and influence exclusions address the practical consequences without assuming Gaussian residuals.

The Breusch–Pagan test does not detect systematic heteroskedasticity (LM = 6.45, p = 0.597). Durbin–Watson is 2.28, providing no indication of positive first-order residual autocorrelation. Five observations exceed the conventional Cook’s distance threshold of 4/n. Maximum leverage is 0.334, and maximum Cook’s distance is 0.126.

Variance inflation factors are 9.31 for the Powell indicator, 5.49 for trend, 2.89 for the Yellen indicator, and 1.90 for the Q&A risk index. High overlap between Chair tenure and calendar time is unsurprising. The target risk coefficient itself does not exhibit serious multicollinearity. Removing the tenure and trend structure would be inappropriate because lexical style changes sharply across Chairs.

4.7. Exploratory Signal Tests

Exploratory models examine direct risk vocabulary, changes from the opening statement to Q&A, and semantic novelty. False discovery rate correction is applied across these tests. See

Table 10. Exploratory communication signals with false discovery rate adjustment.

Exploratory Predictor	Coefficient	Raw p-Value	FDR-Adjusted p-Value
Q&A explicit risk vocabulary	−0.139	0.065	0.228
Absolute Q&A risk change	−0.091	0.082	0.228
Opening explicit risk vocabulary	−0.162	0.114	0.228
Opening semantic novelty	not substantively different from zero	0.311	0.467
Signed Q&A risk change	not substantively different from zero	0.504	0.605
Q&A semantic novelty	not substantively different from zero	0.628	0.628

below.

No exploratory signal survives correction. The negative direct risk coefficients could be consistent with risk resolution language, in which explicit discussion reduces uncertainty, but the evidence is too weak for that interpretation to be presented as a finding.

Separate models for the three dictionary components also remain null. See

Table 11. Individual lexical risk components with false discovery rate adjustment.

Lexical Component	Coefficient	Raw p-Value	FDR-Adjusted p-Value
Negative	−0.075	0.389	0.778
Uncertainty	−0.045	0.588	0.783
Weak modal	−0.005	0.937	0.937
Explicit risk	−0.139	0.065	0.261

below.

5. Discussion

5.1. Main Findings and Scope

Q&A lexical risk does not predict larger cross-asset market responses after policy surprise magnitude and event characteristics enter the full specification. The estimate equals −0.063 with p = 0.444 and a 95% confidence interval from −0.226 to 0.100. Incremental R² equals 0.0034, and leave-one-out prediction error rises by 0.60% after adding the lexical index.

The result concerns transcript-derived lexical frequency and an index of absolute market movement. Acoustic delivery, realised intraday volatility, implied volatility, and answer-level reactions remain outside the reported tests. The paper therefore draws no conclusion about pitch, intensity, timing, pauses, disfluency, or other non-verbal features.

The null result still sets a useful benchmark. Any richer semantic or acoustic measure should add information beyond policy surprises and the transparent dictionary score, and it should improve performance for unseen events. The result also shows why lexical risk, acoustic emotion, policy stance, and non-verbal behaviour require separate labels.

5.2. Hypotheses and Predictive Evidence

H1 receives no support. The full coefficient has the opposite sign from the directional prediction and wide sampling uncertainty. Sign reversals after influence exclusions and after restoring emergency events also argue against a stable directional relationship. See

Table 12. Assessment of the study hypotheses against the available evidence.

Hypothesis	Empirical Finding	Interpretation
H1: Higher Q&A lexical risk is associated with larger market response magnitude	Full coefficient −0.063, p = 0.444	Not supported in the present sample
H2: Q&A lexical risk adds information beyond policy and event controls	Incremental R2 = 0.0034, partial R2 = 0.0072, LOOCV error rises 0.60%	Not supported at conference level
H3: The lexical risk relationship is stronger in Q&A than in prepared remarks	Q&A −0.023, opening −0.213, coefficient difference p = 0.124	Not supported

below.

H2 receives no support from either fit or prediction. Adding Q&A lexical risk raises in-sample R² from 0.396 to 0.399, produces partial R² of 0.0072, lowers out-of-sample R² from 0.2913 to 0.2828, and raises leave-one-out RMSE from 0.5341 to 0.5373. The permutation p-value equals 0.438.

H3 receives no support from the prepared-versus-unscripted comparison. In M5, the Q&A coefficient equals −0.023 and the opening statement coefficient equals −0.213. Their difference equals 0.190 with p = 0.124. Conference-level lexical frequency therefore does not identify a stronger Q&A relation.

5.3. Policy Information, Volatility Environment, and Identification

Policy surprise magnitude remains the strongest empirical predictor. Its full-model coefficient equals 0.368 with p < 0.001, and the asset-specific coefficients remain positive across equities, two-year and ten-year Treasury yields, and the dollar. The large two-year response fits the close link between short maturity yields and revisions in the expected policy path.

The attenuation of the lexical risk coefficient after policy controls enter the model illustrates the identification problem. Its magnitude falls from −0.126 in the bivariate model to −0.047 after policy surprise magnitude and word count are included. Communication content and policy news therefore overlap empirically in the event window.

Pre-event volatility remains an additional common cause. High VIX conditions might coincide with more risk-related wording, larger policy surprises, and larger market responses. Section 3.7 formalises a three-regime SETAR extension using the previous trading day’s VIXCLS close (Chicago Board Options Exchange, n.d.; Tong, 1983). The current archive lacks the merged event-level VIX field, so no regime interaction enters the reported coefficient. The estimates should therefore be read as conditional associations under the available control set.

5.4. Lexical Measurement and Conference Aggregation

Negative point estimates do not establish a stabilising effect. Risk discussion might coincide with clarification, acknowledgement of already priced information, or policy explanations that reduce residual uncertainty. Sampling error also remains substantial.

Dictionary frequency does not recover communicative function. Negation, conditional clauses, temporal references, comparisons, and risk resolution statements alter meaning around the same vocabulary. A finance-specific transformer or a manually coded answer sample would provide a stronger contextual validation layer than raw counts alone.

Conference aggregation also mixes local reactions. One answer might reveal a concern, and a later answer might resolve it. Assigning one 70 min outcome to the full Q&A session removes the temporal order needed to separate those functions. Answer-level observations nested within conferences would preserve local variation and permit conference-level controls.

The 90-event sample further limits interaction tests and small effect detection. The approximate 80% minimum detectable effect of 0.230 response index units should guide interpretation of null coefficients. The absence of evidence for small associations does not supply evidence about unmeasured acoustic or sentence-level semantic effects.

5.5. Relation to Acoustic and Non-Verbal Evidence

The lexical result addresses a different construct from the acoustic result of Gorodnichenko et al. (2023). Their measure draws on audio features and broad vocal emotion. The present index draws on transcript categories for negative, uncertainty, and weak modal language. Evidence from one channel should not be assigned to the other.

Marchal (2020) and Ahrens et al. (2025) also show why economic content and non-verbal behaviour require separate measurement. Visual interaction, forecast information, and acoustic delivery each differ from a dictionary count. The present null result therefore sets a lexical boundary without testing those channels.

A later multimodal design should align each Chair’s answer with audio, transcript text, topic labels, and one-minute or finer market data. Speaker-standardised acoustic variables and contextual semantic measures should enter the same temporal unit before any acoustic–semantic interaction is estimated.

5.6. Limitations and Next-Stage Analysis

Five limitations define the current evidence. The study measures lexical risk, not acoustic delivery. The outcome measures absolute market movement, not realised volatility. Conference aggregation hides answer-level timing. The 90-event sample limits precision for small effects and interactions. The available control vector omits pre-event VIX, same-day macroeconomic announcements, statement window returns, and topic-specific information.

Future text validation should combine a manually coded answer sample with finance-specific transformer measures, semantic novelty, and hawkishness or dovishness. Human coding should distinguish risk revelation, risk resolution, reassurance, conditional statements, and historical references. A hierarchical answer-level model should nest answers within conferences and align each answer with the market movements observed immediately afterwards.

Future acoustic work should use matched official recordings, speaker segmentation, within-speaker standardisation, recording quality controls, and features covering pitch, intensity, speech rate, pauses, disfluency, jitter, shimmer, and harmonics-to-noise ratio. A completed VIX merge should also implement the SETAR regime specification in Section 3.7 and compare M4 residuals across low-, medium-, and high-volatility states.

6. Conclusions

The study examines whether risk-related words in Federal Reserve press conferences predict the magnitude of financial market responses. The primary sample contains 90 scheduled FOMC press conferences. The Q&A lexical risk index combines negative, uncertainty, and weak modal frequencies, while the outcome combines absolute movements in the S&P 500, two-year and ten-year Treasury yields, and the US dollar over the 70 min press conference window.

The full-model Q&A coefficient equals −0.063 with p = 0.444 and a 95% confidence interval from −0.226 to 0.100. Incremental R² equals 0.0034, partial R² equals 0.0072, and leave-one-out prediction error rises by 0.60%. Asset-specific models and the reported sensitivity checks give the same substantive result. The 90-event sample leaves limited precision for small effects.

Policy surprise magnitude remains the strongest predictor across the composite and asset-specific models. The estimates therefore support a narrow conclusion: conference-level lexical risk adds little information about absolute market response magnitude after the available policy and event controls enter the model.

The study does not identify a causal effect of wording. It also does not test acoustic vocal signals or realised intraday volatility. Pre-event volatility and other contextual variables remain incompletely measured, and conference aggregation might combine risk revelation with later risk resolution. The Supplementary Materials provide the supporting documentation, reproducibility files, and additional material associated with the empirical analysis reported in the study.

Future work should move to answer-level timing, contextual text validation, acoustic measurement, and a merged pre-event VIX series. The three-regime SETAR specification in Section 3.7 provides a reproducible test of volatility state heterogeneity once the event-level VIX field enters the analysis archive.