A Generative Deep Learning Framework for Detecting Senior Executives' Characteristics and Its Fintech Applications


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date21 Aug 2020


According to the upper echelon theory, organizational outcomes are partially predicted based on the characteristics of the top executives. Therefore, it is of great business value to analyze these characteristics. However, it is nearly impossible for the analyzers to directly meet with the executives of interest. With the rise of the Internet, one can perceive top executives' personal characteristics through what they have posted on the Internet. However, academic research focusing on the analysis of top executives' characteristics through their social media behaviors is still limited. The largest challenge lies in the variety modalities of information co-existing on the Internet, including text, emojis, images, and videos. In this dissertation, we follow the design science methodology and propose several novel deep learning-based methods to address the above-mentioned challenges and detect executives' characteristics, e.g., Big Five personality, perceived age, perceived trustworthiness, and perceived IQ. 

The first study designs a novel data-driven analytics methodology that is underpinned by an attention-based deep learning model for mining the personalities of executives based on their social media textual posts. In addition, this research examines the usefulness of the proposed method in a real-world Fintech scenario, i.e., the prediction of corporate policy and performance. Under controlled experiments, the proposed data-driven analytics methodology outperforms other state-of-the-art machine learning methods on personality detection by at least 7 % over two benchmark datasets. Based on another real-world dataset collected from social media accounts of senior executives from S&P1500 companies, further experiments verify that executives' personality traits detected by the proposed method have predictive power on the corresponding corporate performance. The managerial implication of this study is that firms can apply the proposed analytics methodology to uncover the personality traits of their senior executives, and hence predict corporate performance.

The second study aims to extend the personality detection method proposed in the first study, which only deals with textual data, to process multimodal information, including both text and images. The challenge is to coordinate varying levels of noise and conflicts between modalities. Controlled experiments are conducted to demonstrate that by combining text and image information from the social media posts of CEOs together, the performance of the personality detection method is further improved. Moreover, the study examines the perceived risk propensity conveyed by executives’ social media behaviors being negatively related to corporate credit rating. The managerial implication of this study is that analysts or investors can apply the proposed analytics methodology to uncover the personality traits of CEOs of interest, and hence analyze corporate credit ratings.

The third study involves comprehensive analyzing startup entrepreneurs’ online impression conveyed by their social media behaviors. Following the design science methodology, the main contribution of this study is to propose a novel online impression analytic approach, named ExeAnalyzer, for executives to predict the impression made on others by the executives’ social pages. Specifically, ExeAnalyzer is designed to perceive online impression in terms of the executives’ social media profile images, posts, and other sociometric attributes. Under a dataset containing 7086 startup companies in AngelList, a popular equity crowdfunding platform, and empowered by the proposed artifact, this study conducts an empirical analysis and determines that the CEOs’ online impression of dominance is negatively related to their financing performance. Moreover, the experiments demonstrate that the detected features have predictive power on the early success of startups in the financial market.

    Research areas

  • Fintech, Multimodal Data Analysis, Deep Learning, Corporate Finance