Making better decisions through tracking emotion

Making better decisions through tracking emotion

Deniz Iren gave a talk during the Institute of Data Science Research Seminar Series on emotion tracking. Deniz is a researcher at the Business Intelligence and Smart Service Institute and has a computer science background with a strong emphasis on business application. The Research Seminar Series is bringing together data scientists from the various schools at Maastricht University at one location to learn from each other. Deniz presented ELISA, an emotion-tracking dashboard. The scientific question was to discover if emotion contain hidden information, and the business application relates to reducing uncertainty when making financial decision by being able to explore relations between emotions and stock prices.

Context: Who wants to track whose emotions?

With GDPR, the new European privacy protection law, right around the corner, the question of whose information is being tracked might be the most pressing. The context of the study is a big pension fund. Pension funds need to invest in companies in order to have enough money to pay out the pensions they promised. They prefer long-term investment compared to short-term gains. For pension funds to make investment decisions quarterly earnings become a main source of informaiton.

Investment in companies is not based on gut feelings, but after careful analysis of all potential information. But, while rational-based financial theory assumes perfect information and unlimited time, the reality, as expressed by behavioral finance, is different. For example, your degree of risk aversion influences where and how much you are going to invest.

Changing the unknown knows to know knows

To make sustainable investment decisions, investment analysts rely on quarterly earning reports. These are written reports and webinars during which companies report their earnings for the past quarter and participants can ask questions. In this situation, it is clear that the person who is hosting the webinar and presenting the report has more information than the participants. For example, new product launches, a failed service development, or the recent loss of a big client are events that are not necessarily made public and therefore not known to everybody.

Behavioral finance, the branch of finance that recognizes the human nature in humans, argues that there is an information gap between the person who is reporting the quarterly earnings, the host, and the people who are listening to it, the participants. This information gap is best characterized by looking at the three types of knows: Known knows, unknown knows, and unknown unknowns. While at first sight confusing, it’s pretty logical. The known knows are all the events and aspects about a company that are known. This is the information in the report and talked about on media. The unknown knows are the things that are known only to some, but not by the person in question (e.g., the investment analyst). If somebody has that information, there are methods to get to it. Finally, the unknown unknows are what insurance companies often call “acts of God”. Nobody knows them, and hence there is no way to find out.

To make better financial decisions it helps to have more information. Hence reducing the information gap by changing the unknown knows to know knows will make a difference. This is the job of ELISA.

emotion-tracker-information-gap

Building an emotion tracker

The scientific question behind ELISA, the emotion tracker, was to find out if emotions contain hidden information. The short answer is yes.
Several steps were undertaken by Deniz and his team to create ELISA. The data was in audio and text format and taken from the quarterly earnings presentation. ELISA was based on VERA, which by now with an accuracy of 60 % detects emotions in speech. VERA’s aim is to recognize emotions in call-center talks. To build ELISA, the team combined Aenas, Weavenet, and PyAudioAnalysis. The user interface was build using Pythons dashboard tool dash. Very briefly, these were the steps that the team undertook:

  1. Using the sentence as the unit, the text and audio was synchronized.
  2. Using an army of workers, the emotions in the audio was labeled. The emotional classes defined by Dr. Ekman were used, with the addition of neutral (which in Deniz’s words are non-emotion).
  3. The input generated in the previous step was used to create a dashboard which allowed for exploration of emotions together with financial performance of the company in question.

Deniz did a short presentation of the dashboard. The key fields a user could indicate where keywords (e.g., Brexit, Italy, fraud, risk, stability), the name of the company, and where this keyword should appear (e.g., question of participant, answer). The output was a table indicating for each quarterly earnings, how frequently a specific emotion was present. It was also possible to further explore it, by visualizing emotions together with financial performance, such as stock prices.

Based on the exploratory result there are some indications that emotions contain hidden information about events that can have a positive or negative impact on a company’s future financial performance. For Deniz’s team, the next step is to conduct time series analysis to reduce the uncertainty inherent in their exploratory analysis.

Do you like what you read? Do you want to explore the scientific backing of your arguments or learn more about a topic ? Let’s talk.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Close Menu