What sparked your interest in bringing textual analysis into the classroom, and why do you think it has become such a relevant tool for finance professionals today?

Let me start with the second question. Textual analysis helps us to better understand investor behavior and market outcomes. Also, research has shown that manager sentiment, i.e., how positive or negative managers talk during earnings conference calls, is predictive for future firm performance. Therefore, unstructured text data is an excellent complement to the quantitative financial data that have been used for decades in the industry.

To prepare the Master in Finance students best for their future careers, it is important to give them the opportunity to acquire a comprehensive methodological toolbox for analyzing real-world data. Furthermore, on a more personal note, I find textual analysis fascinating and often use it in my research. Thus, I am more than happy to share my knowledge about and fascination for the topic in the classroom.

Companies often adapt their communication to avoid negative sentiment or downplay bad news. What insights can textual analysis reveal about these strategies?

Interestingly, we observe that the sentiment of earnings conference calls has become more positive in recent years, whereas the sentiment of annual reports has become more negative. In other words, there is a growing gap between the sentiment of earnings calls and sentiment of annual reports. This gap may originate from firms adding more and more risk factors to their annual reports to lower litigation risk and, at the same time, portraying an increasingly brighter picture of the future economic prospects in earnings calls.

The course covers a wide range of texts, from annual reports to earnings calls. Which of these sources do participants usually find most surprising in terms of their informational value?

From my perspective, students are very interested in earnings conference calls because most students have only little or no previous knowledge of the earnings call setting. Therefore, looking at verbatim transcripts of what the top executives of publicly listed companies from all around the world said about their firms’ financial conditions and business outlooks creates a lot of research ideas in students’ minds.

Annual reports are, however, also interesting. For example, there is a nice paper called “Lazy Prices” that was published in the Journal of Finance in 2020 that shows that subtle text changes in firms’ annual reports are a strong predictor for future firm performance. For students, it is surprising to learn that simply comparing the text of a firm’s current annual report to the text of its previous year’s report can make a successful investment strategy.

How do you see students using the skills they gain, whether in research, investment analysis, or corporate roles?

Let’s start with research. I had the pleasure of supervising several students who used textual analysis methods in their master thesis to measure firms’ risk exposure. For example, one student developed a dictionary for inflation risk and analyzed how firms disclose inflation risks in their annual reports and how investors respond to it. Being able to measure firms’ risk exposure more precisely is not only interesting for academia but also helpful for investment analysis. Many asset managers are already using text data to extract signals about firms’ future performance. For corporate top executives, it is important to understand that markets may hang on their words.

Large language models like ChatGPT are changing the field rapidly. Which developments do you expect will shape the future of textual analysis in finance?

The recent large language models are a great tool for researchers and industry professionals as they allow them to extract more precise signals from unstructured text data. However, the models also come with problems. Let me provide two examples: First, they can be a black box, i.e., it is hard to tell how a model arrived at its assessment. As a researcher, it is, however, important to know the reasons for the model’s classification, particularly when testing economic theories. Second, large language models may suffer from look ahead bias, i.e., they may use information that would not have been available to a user in real time. Assume that I ask a model to decide whether an article about Wirecard AG that was published in early 2020 is positive or negative. The model may decide to label the article as negative, not because of its content but because it knows that Wirecard AG filed for bankruptcy in June 2020.

Leaving out the financial texts, what other texts do you love to spend your time with?

As a finance professor, I read a lot of finance research papers ranging from the master theses of GBS students to the latest publications in the Journal of Finance. However, I assume that these types of papers also count as financial texts…

Outside of finance, I listen – I am more an audio book person – to non-fiction books. The two most recent books that I listened to were “Runnin' Down a Dream: How to Thrive in a Career You Actually Love” by Bill Gurley and “Atomic Habits” by James Clear. I highly recommend both.