How Do Fed Chairs Affect the Financial Market? A Fine-grained Monetary Policy Analysis Framework on Their Language (2024)

Sentiment Analysis in Finance

The term Sentiment Analysis was first defined by Nasukawa et al. (Nasukawa and Yi 2003) as determining the subjectivity polarity (positive or negative) and polarity strength (strongly positive, mildly positive, weakly positive, etc.) of a given review text. Sentiment analysis is increasingly recognized as a pivotal task within the financial domain, closely linked with market sentiment, which subsequently influences price movements. Early literature (Bharathi and Geetha 2017; Gorodnichenko, Pham, and Talavera 2023) predominantly focused on word categorization methods, such as the bag of words technique, to assess tonality. The introduction of the polarity lexicon by (Loughran and Mcdonald 2010) stands out as a significant advancement in the development of sentiment analysis methodologies for finance. Further, (Garcia, Hu, and Rohrer 2023) contributes new lexicons of positive and negative terms specifically for financial contexts. Nonetheless, these dictionary-based approaches face two principal challenges: 1) Context Insensitivity: They often neglect the subtleties of context, such as sarcasm or culturally specific expressions, resulting in imprecise sentiment analysis; 2) Inadequacy in Processing Complex Emotions: These methods are limited in their ability to interpret complex emotional expressions and fail to accurately discern sentiment intensities or nuances.Recently, large language models like Bert (Alaparthi and Mishra 2020) and RoBERTa (Liu etal. 2019) have demonstrated extensive capabilities, significantly enhancing a range of natural language processing (NLP) tasks. Despite their remarkable efficacy in a wide array of NLP applications, Pretrained Language Models (PLMs) encounter limitations in domain-specific tasks. This is primarily due to the marked disparity between the distribution of general training data and the specialized language of specific domains (Gururangan etal. 2020; Gu etal. 2022). An expression that holds a strong semantic orientation in one domain might necessitate a different interpretation in another. To mitigate this, several financial language models (Araci 2019; Yang, Uy, and Huang 2020) have been proposed to enhance performance in analyzing financial and economic texts. For example, recent endeavors, such as by (Shah, Paturi, and Chava 2023), have concentrated on monetary orientation analysis through the fine-tuning of large language models.

Policy Communication of Central Bank

Economists have long recognized the importance of central bank communication (Kohn and Sack 2003), making efforts to study its impact on financial markets both quantitatively and qualitatively (Blinder etal. 2008; Blinder 2018; Neuhierl and Weber 2019; Coenen etal. 2017). Despite all these efforts, the communication mechanism between the Fed Chairs and the market, an important channel of central bank communication, has not been fully explored and understood due to the technical limitations. Most related papers are restricted to the impact of the Chairs’ communication on the domestic financial market (Gorodnichenko, Pham, and Talavera 2023; Hansen and McMahon 2016). To fill in the gap, our paper investigates the impact of the Chairs’ language on the international financial market for the very first time, which contributes to the broad literature on cross-border transmission of monetary policy.

Problem formulation

We estimate the above specification for each horizon in a daily frequency, ranging from 0 day to 15 days. To illustrate the dynamic effects explicitly, we draw the estimated coefficients. 90 % bias-corrected confidence intervals are reported, shown by the dash lines in each figure.

Dataset Collection and Preprocessing

Monetary Policy HearingsOur dataset of this part includes samples from the United States Senate Committee on Banking, Housing, and Urban Affairs website¹¹1https://www.banking.senate.gov/hearings/, detailing hearings on financial and economic policies. An example is the July 13, 2017²²2https://www.banking.senate.gov/hearings/the-semiannual-monetary-policy-report-to-the-congress, hearing on the semiannual monetary policy report, featuring Janet L. Yellen, then Chair of the Fed, as the key witness. The website provides dates, times, locations, and additional resources like opening remarks and testimonies.

In our study, we scraped the United States Senate Committee on Banking, Housing, and Urban Affairs website to compile a list of all hearings. We then applied two criteria to identify the Monetary Policy Hearings: 1) the title must contain “monetary policy report,” and 2) the “Chair” must be listed among the attendees. From this, we extracted essential details from the hearing URLs, such as primary attendees, audio recordings, the chair’s testimonies, and meeting dates. For the formal statement documents (readouts) reported to Congress, we used Pdfplumber (Singer-Vine and The pdfplumber contributors 2023), a Python package, to extract the text and segment it into sentences. For meetings without transcripts, we employed an Automatic Speech Recognition (ASR) tool, WhisperX (Bain etal. 2023), for transcription. Our final dataset, as summarized in Table1, includes 42 Monetary Policy Meetings from 2003 to 2023, each with a PDF of the readout and, when available, an MP3 audio file. Notably, audio recordings were not available for all meetings; 12 out of 42 lacked audio files. We segmented a total of 36,203 audio pieces, averaging about 6.1 seconds each, mostly ranging from 3 to 6 seconds, showing a uniform distribution in duration.

Press ConferenceThe Federal Open Market Committee (FOMC), the Fed’s key monetary policymaking body, consists of the Board of Governors’ seven members and the presidents of the twelve Reserve Banks. They meet approximately eight times a year to discuss economic conditions and develop strategies for maximum employment and price stability. Since 1994, and regularly after each meeting since May 1999, the FOMC issues statements outlining federal funds rate targets. These statements gained importance after December 2008 when the federal funds rate hit its effective lower bound. To improve transparency, Chairman Ben Bernanke began holding press conferences after four annual meetings starting in April 2011, a practice that expanded to after every meeting since 2019.

Our dataset, sourced from public records, includes conference call videos, audio recordings, and text transcripts from six international banks from 2011 to 2022. We specifically analyze Fed press conference videos from April 27, 2011, to January 26, 2022, covering 47 conferences. Differing from the conversation-level analysis in (Gorodnichenko, Pham, and Talavera 2023), we segment our text and audio data into sentences, rather than larger paragraphs, to preserve detailed information. This approach, focusing on sentence-level granularity, provides more precise semantic insights and variation, enhancing our model’s understanding and predictive accuracy. Our comparative analysis between sentence-level and conversation-level approaches shows that our fine-grained method is more effective. Besides, it also offers more context than phrases-based or words-based analysis. Our analysis reveals a higher proportion of neutral segments in the data, reflecting the cautious communication style of Fed Chairs, who often use neutral language to avoid unintended policy signals. This tendency towards neutrality is evident in our predictive results, aligning with the Chairs’ careful choice of words.

Acoustic and linguistic Analysis

Linguistic AnalysisPrevious studies have demonstrated that fine-tuned pre-trained language models (PLMs) significantly outperform both rule-based models and zero-shot models. Therefore, we employ the RoBERTa-large pre-trained language model as our foundational model, which we further fine-tuned for downstream classification tasks. Numerous studies have successfully fine-tuned PLMs for tasks related to emotion classification. However, our research diverges by aiming to extract policy-related information from text. Drawing inspiration from prior literature (Shah, Paturi, and Chava 2023), we do not employ traditional emotional characteristics such as neutral, positive, and negative. Instead, we fine-tuned the RoBERTa model using a recent open-source dataset (Shah, Paturi, and Chava 2023). This dataset comprises three categories: Dovish, Hawkish, and Neutral. Dovish sentences indicate an impending easing of monetary policy, whereas Hawkish sentences signal an upcoming tightening of monetary policy. Finally, The model achieved an average testing accuracy of 72.8% and an average testing F1 score of 72.4%, using a learning rate of 0.000001.

Acoustic AnalysisIn response to the absence of a speech database tailored to policy stances, we adopt the wav2vec2.0 model (Baevski etal. 2020) which is composed of multi-layer convolutional layers. In order to extract emotion-related acoustic features, the model is fine-tuned to predict the speaker’s emotional state on a subset of the Interactive Emotional Dyadic Motion Capture (IEMOCAP) (Busso etal. 2008) database, a prominent repository for English emotional speech, containing nine emotions. This database includes both scripted and improvisational sessions between two actors, making it a benchmark for English SER systems. It contains 12 hours of audiovisual recordings by ten actors, totaling 10,039 utterances, each annotated by three evaluators. The subset focuses on primary emotional states like neutral, happiness, sadness, and anger. The model achieved an average testing accuracy of 75.3% across these four emotion categories.

Policy Indicator

Comparative Analysis of Different Granularity

To evaluate the efficacy of the fine-grained method we employ in our framework, We first visualize the distribution of the sentiment indicators’ results derived from both fine-grained and coarse-grained methods. The coarse-grained method means we calculate the value of the sentiment indicator based on the conversational level as in (Gorodnichenko, Pham, and Talavera 2023). As shown in Fig.3, our findings reveal a high correlation in trends of sentiment across different granularities, suggesting that indicators from varied granularities yield consistent policy implications. Notably, the indicator based on fine-grained data demonstrates smoother trends with fewer extremes compared to its coarse-grained counterpart. This observation aligns with the expectation that as public figures, Chairs tend to maintain prudence and consistency in their communications.Furthermore, we perform a local-projection regression analysis on U.S. stock prices (SPY) to assess the performance of different granularity levels in predicting financial market volatility. Our results, illustrated in Fig.2, reveal that incorporating a fine-grained sentiment indicator significantly enhances the predictive capability of our framework. The sentiment’s coefficient became notably significant at a 10% confidence level with a 10-day lag. In summary, employing a fine-grained analysis method improves the accuracy and predictability of our framework, contributing to the resolution of critical economic issues.

Analysis of Domain-Specific Fine-tuning

In this section, we explore the impact of domain-specific fine-tuning. We use RoBERTa (Liu etal. 2019) as a baseline, which has been fine-tuned for three-category emotion recognition (positive, negative, and neutral) (Barbieri etal. 2020), in contrast to policy stance. We adjust our regression equation to incorporate specific sentiment measures. Our findings show that the financial market’s responses vary significantly between Fig.1(a) and Fig.4. Notably, the sentiment index based on positive and negative classifications proved consistently insignificant, even with control variables, suggesting it is not a reliable predictor of market volatility. However, one unit increase in the sentiment index using dovish-hawkish classifications increases SPY by approximately 500 bp after a 10-day lag, underscoring the value of domain-specific tuning. Consequently, we argue that the lack of a policy-specific speech corpus hindered our ability to achieve the desired analytical results from audio data, as discussed in Sec.Comparative Analysis of Different Modalities.

Comparative Analysis of Different Communication Scenarios

In this study, we focus on the communication from two scenarios: the Q&A sessions at the post-FOMC press conferences and the semiannual monetary policy hearings. Both of these two channels play a vital role in the Fed’s policy communication: 1) the press conferences serve to prepare the public by shaping their expectations for the upcoming policy decisions (Ehrmann and Fratzscher (2007); Boguth, Grégoire, and Martineau (2019)), 2) the semiannual monetary policy hearings constitute an important communication channel between the Fed and Congress, by providing opportunities for the Fed to convey its policy intentions and economic outlook((Labonte and Makinen 2008)). The dynamics of sentiment orientation are presented in Fig.5. We observe that over a given period of time, the inclinations and dynamics of the monetary policy reflected in what the Chair said in both scenarios of meetings remain highly consistent. The analysis results align with the actual economic situation. For example, since 2010, the Fed’s monetary policy can be broadly categorized into the following stages: 1) From 2010 to 2014, the Fed supported economic recovery after the Great Financial Crisis with very low interest rates and quantitative easing, which could be reflected with the dovish sentiment result; 2) During 2015 and 2019, the sentiment tended to be a little bit more hawkish as the Fed tapered its asset purchases and gradually increased the federal funds rate; 3) Between 2020 and 2021, in response to the COVID-19 pandemic, the Fed cut rates to near zero and resumed quantitative easing to support the economy, as reflected in Fig.5 as a jump; 4) After 2021, as inflation pressures mounted, the Fed started discussing tapering asset purchases again, setting the stage for potential rate hikes, thus we can see a giant drop in all the four lines in Fig.5 accordingly, reflecting that the Chair Powell is getting much more hawkish in every talks he gave. In conclusion, our recognition of the policy stance is basically in line with the literature Bernanke (2020).

Meanwhile, some systematic differences are also shown in Fig.5. The sentiment of policy communication tends to be more neutral at press conferences, as the score of red line is closer to zero. We explain the reason that the Federal Reserve Act requires the Fed chair to report to Congress twice a year on monetary policy, provide explanations, and answer questions from members of Congress. The Chair needs to explain the Fed’s policy path in a transparent way so that he can assure Congress that he has been highly responsible, thus the sentiment value for other lines, besides the red one, can be volatile more according to the varying economic fundamentals. On the other hand, when communicating with the public, the Chair would try to be neutral to avoid unnecessary panic in the market. Thus, it’s natural that Chairs would sound relatively neutral in the press conferences.

Comparative Analysis of Different Modalities

To compare the effectiveness of our approach under different modalities, we conduct two parallel empirical regressions on text and audio modalities respectively, as shown in Fig.6. The result reveals that the Voice Tone index, calculated from acoustic sentiment tags, lacked statistical significance in forecasting the financial market. We speculate that the scarcity of speech corpus tailored to the financial sector notably hampers the method’s efficacy. Consequently, the model’s capacity to discern the underlying sentiment in the Fed Chair’s speech is compromised, thereby impeding its predictive power to financial market dynamics. Therefore, it is of significant importance to establish a speech dataset related to the financial sector in the future.

Analysis for Cross-border Transmission

The cross-border transmission of monetary policy is quite an important empirical issue in finance and economics. The existing literature in this field is mainly based on diverse interest-rate indicators to measure monetary policy ((Barbosa etal. 2018)), while the soft information channel represented by the Chair’s speeches has been underexplored. This paper considers the impact of the policy stance revealed in the speeches of the Fed Chair on other major economies, including China and the EU. We investigate the variance of three key financial variables, namely the stock price, the monetary-policy-related interest rate and the exchange rate between the national currency and the US dollars.

The impulse response results in Fig.7 show that the Fed Chairs’ language do have a spillover impact on the financial markets abroad. The sign of the coefficients is the same, while the magnitude diversify across China and the European Union. For China, the sentiment has an immediate impact on China’s short-term money market rates, leading to a 15 bp decrease within 10 days; the response of the stock market is slightly delayed, with over 500 bp increase in the stock price in about a week later; however, the sentiment does not show a significant impact on the exchange rate between US dollars and China RMB, which may because after the rapid and flexible adjustment of interest rates, the interest rate parity condition is met again, at which point the exchange rate does not need further adjustment. The responses of financial markets in the EU are basically consistent with China, whether in sign or magnitude. However, the duration of the spillover effect is different. In terms of the responses of the stock prices and interest rates, the effect diminishes in 10 days in China, while it continues to move the corresponding markets in EU in each horizon. In conclusion, the effect of the Chairs’ language would spill over to the financial markets in other countries, especially those major economies which are closely linked to the USA in such a globalized world.