Corresponding Quantlets to master's thesis 'Face value of companies: deep learning for nonverbal communication'
Abstract:
As a side effect of digitalization, a massive amount of unstructured data is generated every day. Unstructured data comprises video, speech, text, and image data, which are easy to interpret for humans - but can be challenging for computers. Financial research has been much engaged in recent history with decision-making based on textual or sentiment analysis. Textual analysis is based on the verbal part of communication, but in human interaction on a face-to-face level, nonverbal communication can play an equally important role in supporting a message. The interpretation of emotions in facial expressions is a major component of nonverbal communication. Deep learning is a versatile technique, that is used in numerous applications, providing somewhat cognitive capabilities for machines. This thesis describes how to build a deep convolutional neural network with the ability to detect emotions in faces. Different approaches in deep convolutional model designs are tested and evaluated. The results are then used to evaluate videos of the regular press conference of the European Central Bank between January 2011 and September 2017. This processing step results in emotional-scores of facial expressions from 70 press conferences and more than 200,000 single pictures. It is investigated whether information of nonverbal communication, measured in levels of emotional excitement, can be linked to the movements of the Euro Stoxx 50 index. This `face value' is compared to the value of speech and accompanying research. Using image data from press conferences as source of unstructured data and transferring of nonverbal communication to stock markets are both topics that, to the best of found knowledge, have not yet been focused upon in research.