Tarun Khanna and Prithwiraj Choudhury use machine-learning technology to look for link between a CEO’s communications style and company performance.
EOs are communicators. Studies show that CEOs spend 85 percent of their time in communication-related activities, including speeches, meetings, and phone calls with people both inside and outside the firm.
Now, new research using machine learning is attempting a deep dive into the words and facial expressions of chief executives to see if leadership style can be correlated with a firm’s performance. The researchers believe their work could open new directions in big data analysis, combining image and textual analysis to create a more complete picture of how a chief executive influences firm performance.
“Machine learning is able to utilize data that is both large in size, but also in a different form than what would traditionally fit into an Excel spreadsheet,” says Harvard Business School’s Prithwiraj (Raj) Choudhury, the Lumry Family Associate Professor in the Technology and Operations Management Unit. “We are now able to work on all these rich new sources of visual data.”
In a forthcoming paper in the Strategic Management Journal, Machine Learning Approaches to Facial and Text Analysis: Discovering CEO Oral Communication Styles (pdf), the authors use a variety of techniques to group CEOs into distinct communication styles that, among other things, appear to be correlated with the financial performance of their companies.
“THERE IS A WHOLE OCEAN OF DATA OUT THERE THAT PEOPLE AREN’T USING”
The research was conducted by Choudhury and Tarun Khanna, the Jorge Paulo Lemann Professor in the Strategy Unit at HBS; Columbia Business School professor Dan Wang; and doctoral student Natalie Carlson.
The skills that a CEO needs to command a company are varied, but the ability to communicate effectively is near the top. “The task of being a leader in an organization is to pull together different resources to accomplish something productive,” says Khanna. “That means you are motivating people to do things, and the only way to motivate people is to communicate with them.”
As crucial as communication is to running an organization, however, it’s difficult to get an accurate read on differences in CEO communication style. The words they choose are crucial to meaning, of course, but they also express themselves through tone and nonverbal clues such as facial gestures. Further, these verbal and nonverbal cues differ across cultures and geographies.
“You might say something positive, but a negative facial expression may create the opposite meaning,” says Choudhury.
Finding CEOs to test
But where could the researchers find enough CEOs to conduct their study?
It turns out that Khanna was familiar with the answer. He and Geoffrey Jones, the Isidor Straus Professor of Business History at HBS, have been compiling an oral history project called Creating Emerging Markets, 130 video interviews of iconic business leaders from emerging markets discussing their careers and organizations in an unstructured format.
Those videos were analyzed using three machine-learning techniques.
First, the researchers looked at the words that CEOs chose, using statistical inference to eventually generate 100 topics as diverse as marketing, corporate boards, and personal family history. Each CEO was scored based on their tendency to stay on a particular topic, versus bouncing around from subject to subject, a measure they called “topic entropy.”
The second machine-learning technique also looked at words, but this time divided them by positive or negative valence, or how much the speaker vacillated between positive and negative emotions.
Finally, the researchers examined the nonverbal communication of CEOs by analyzing their facial expressions using a computer-vision application that rated them according to eight emotions: anger, contempt, disgust, fear, happiness, neutral, sadness, and surprise. This type of analysis had typically been done by human coders, but the researchers discovered that computer analysis proved to be a robust, yet fast and cheap alternative to generating emotional data.
Does CEO style correlate to performance?
With that data in hand, the researchers grouped CEOs into one of five distinct communication styles. Examining the various attributes, they gave each one a name. Those who used positive language and a range of facial expressions, for example, were deemed Excitable. Those who showed anger, contempt, and disgust but also a fair amount of neutral expressions were labeled Stern. CEOs with high topic entropy and who seemed to use happy and contemptuous facial expressed were called Rambling. Those with a range of expressions were called Dramatic, while those characterized by sadness and negativity were termed Melancholy.
The different styles only seemed to emerge, say Khanna and Choudhury, when both text and facial features were analyzed together. “We were able to take these multiple dimensions and construct these styles that were not possible before,” Choudhury says.
The researchers then looked for correlations between communications style and other aspects of the business. For example, those with a higher Dramatic score showed less merger and acquisition activity in the year following the interview. Also, the communication styles that emerged from jointly utilizing text and image data had more statistical power in explaining variability in the data, as compared to utilizing text or image data alone.
Choudhury and Khanna stress those results are only illustrative. Their main purpose in writing the paper, they say, is as proof of concept in opening up the conception of what kind of data machine learning can effectively analyze.
“There are so many new forms of data out there, and with computing power going up stratospherically in the last 20 years, there are now many opportunities for better analyzing many of the things going on in business,” Khanna says. YouTube, for example, could provide a rich trove of information to analyze CEO communication from speeches.
Additional uses for machine learning research
These additional data sources are especially useful in studying emerging markets where conventional data sources may be less easily available and researchers’ understanding of those markets are more rudimentary. Also, many of these geographies are rich in technology-based sources of data that can be studied, such as the mobile phone and Internet of Things explosions across many fast-growing Asian economies.
Beyond academic questions, business analysts could use these machine-learning techniques, which the researchers say are relatively simple and easy-to-use, to analyze voice patterns in earnings calls or nonverbal communication at shareholder meetings to add depth to their assessment of company performance.
“There is a whole ocean of data out there that people aren’t using,” says Khanna. “Using it could help examine many questions that are essential to business.”
Michael Blanding is a writer based in the Boston area.
Source: This article was first published Aug. 21, 2019 at https://hbswk.hbs.edu/item/machine-learning-can-help-us-understand-ceo-behavior. Reprinted here with permission from Harvard Business School.