Today we are pleased to present guest contributions written by Dimitrios Kanelis (Westfälische Wilhelms-Universität Müun) and Pierre Sieklos (Wilfrid Laurier University and CAMA at ANU). The opinions expressed here are their own and do not reflect the official opinions of the organizations to which the authors are affiliated.
The obvious source of non-written information is a person’s voice because voice does not necessarily convey the same message as words. Gorodnichenko et al. (2020) provides evidence of how the Fed chair’s tumultuous sentiments affected stock prices in the days following the FOMC press conferences. Complementary studies by Curti and Kazinnik (2021) and Alexopoulos et al. (2022) Estimating the real-time effects of the Fed Chair’s facial emotions on stock prices.
in our area paperWe estimate the impact of vocal sentiment and language during the question-and-answer sessions of the ECB press conferences under Mario Draghi’s presidency on the yield curves of the four largest economies in the eurozone as well as margins versus German yields. We conduct an event study and create a new dataset consisting of simultaneous audio and text data in time for press conferences between May 2012 and October 2019.
One of the challenges is that Draghi answers several questions in a row on completely different topics. Therefore, we exploit an interesting feature of the transcripts of the ECB press conference. The ECB staff defines the contact points and their structures in writing. By following this structure, we adjust the phonemic data for each answer and establish synchronicity between the phoneme and the words. Then we implement the fully convolutional neural network (FCN) of García-Ordás et al. (2021), which has a special feature for processing audio files of non-fixed length.
To measure the unwritten language content used in press conferences, we apply Fin-BERT. This large-language neural network model can analyze economic and financial texts and goes beyond the word-counting approach of dictionary methods. We use data from France, Germany, Italy and Spain.
To identify sentiment and create a numerical variable to estimate our event regression, we use methods developed in SER, a subfield of machine learning (Pérez-Espinosa et al., 2022). Recently, economists have begun to use SER for the Fed Chair’s vocal sentiment analysis to estimate effects on asset prices (Gorodnichenko et al. (2020); Alexopoulos et al. (2022)).
In contrast to Gorodnichenko et al. (2020), we use a fully convolutional neural network (FCN) based on García-Ordás et al. (2021), which generates higher out-of-sample accuracy and has clear advantages when measuring sentiment during question-and-answer sessions, which are characterized by responses varying in length. These feelings are neutral, calm, happy, sad, angry, and stunned, and are available in two different intensities (non-masculine emotional intensity and strong emotional intensity).
During the European Sovereign Debt Crisis (ESDC), vocal sentiment was found to be consistently negative. The exception is the press conference on August 2, 2012, which is the most positive moment during the crisis and can be seen a few days after Draghi’s famous ‘whatever’ speech which is considered a turning point during the ESDC. After the end of ESDC, a temporary increase in vocal sentiment could be observed before becoming more negative again through most of 2014, which was a challenging year for the ECB Governing Council due to a low inflation environment and economic growth, increased financial fragility and risks of erratic inflation expectations. The introduction of the Asset Purchase Program (APP) is in line with a more positive sentiment, possibly due to Draghi’s success in pushing through unconventional monetary policies despite the controversy surrounding the policy within the board (Brunnermeier et al., 2016). The decline in average vocal sentiment is seen again during 2018, a time of increasing policy challenges and hitting a new low when the European Central Bank resumes quantitative easing, just a few months after the Governing Council kicked off the start of its exit attempt.
For clarity, the slide below is an example containing positive, neutral, and negative sentiments (all possible in the same press conference) that Draghi said according to our methodology. Readers are asked to indicate whether they agree with the emotion displayed by Draghi as determined by the methodology used (answers are provided in the caption below the slide). Actual illustrations are also selected for comparison.
NB: The first soundtrack is neutral as of April 27, 2017; The second is positive of 1St press conference on future directions on July 4, 2023; The last clip is negative and is from April 15, 2015 when an activist jumped on the table.
Among the hypotheses tested was whether positive emotions increase yields while negative emotions decrease yields. We estimate a significant positive effect of the change under the press conference on spreads, ie a positive change in the statement leads to a larger spread while a negative restatement leads to a lower spread. Marginal plots measure the effects of vocal emotion given a given level of positivity in the language used during a question-and-answer session. This allows us to quantify more subtle effects on return margins than regressions alone can provide. The figure below shows the marginal impact of acoustic sentiment on Italian ten-year bond spreads.
NB: These plots depict the marginal effect of a change in VO at time t given a given level of PositivityAN, also at time t, on the ten-year government bond spread of Italy. We report the change in the margins of the basis points (ie in the range -25 to +25).
The interaction between vocal feelings and language during a question-and-answer session has a significant and asymmetric effect on Italian bond spread. The combination of negative vocal emotions and negative language increases the prevalence of the disease. For example, Italian ten-year bonds also react differently, depending on the degree to which the sound and language signals conflict with each other. The affirmative-positive combination of sound and language and the negative-negative combination increase diffusion, while conflicting signals decrease diffusion. In the case of Germany (not shown), non-written communication has a positive effect on productivity. However, this effect is limited to the short end of the yield curve and is asymmetric with respect to the type of acoustic emotion. More positive communications indicate an increase in German Bund yields. The paper also discusses the impact of Draghi’s sentiment and the content of the ECB’s press conferences on French and Spanish bond yields.
One of the central themes of the paper is a reminder that communication goes beyond just words. Future research should look at how financial markets perceive and process acoustic signals during crises such as the COVID-19 pandemic or surging inflation since 2021 and how central bankers’ emotions can influence asset prices during times of heightened economic and geopolitical uncertainty.
The paper is available at:
Alexopoulos, M., Han, X., Kryvtsov, O., and Zhang, X., 2022. More Than Words: Communications of Federal Reserve Chairs During Congressional Testimonials. Worksheet.
Brunnermeier, M.K., James, H., Landau, J.P., 2016. The Euro and the Battle of Ideas. Princeton University Press.
Curti, F., and Kazinnik, S., 2021. Let’s face it: Measuring the impact of nonverbal communication at Federal Open Market Committee press conferences. Working paper draft: March 18, 2022.
García-Ordón, MT, Alaiz-Moretón, H., Benítez-Andrades, JA, García-Rodríguez, I., García-Olalla, O., Benavides, C., 2021. Sentiment analysis in non-fixed-length sounds using a grid fully convolutional neuron. Processing and monitoring of biomedical signals 69.
Gorodnichenko, Y., Pham, T., Talavera, O., 2020. The Voice of Monetary Policy. Draft Working Paper: June 14, 2022, forthcoming in American Economic Review.
Perez-Espinoza, H.; Zatarin-Kabada R, Baron-Estrada ML 2022. Emotion recognition: from speech and facial expressions. Chapter 15. Published in: Torres-García, Alejandro A. and Reyes-García, Carlos 1 and Villasenor-Pineda, Luis and Mendoza-Montoya, Omar (Ebs.): Bioprocessing and classification using computer learning and intelligence. Elsevier.
This post is written by Dimitrios Kanellis And Pierre Siclos.