Measuring executive personality using machine‐learning algorithms: A new approach and audit fee‐based validation tests

• Published date: 01 March 2020
• Author: Rafael Rogo,Christine Wiedman,Ray Zhang,Jiri Novak,Karel Hrazdil
• Date: 01 March 2020
• DOI: http://doi.org/10.1111/jbfa.12406

DOI: 10.1111/jbfa.12406

Measuring executive personality using

machine-learning algorithms: A new approach and

audit fee-based validation tests

Karel Hrazdil1Jiri Novak2Rafael Rogo3Christine Wiedman4

Ray Zhang1

1Simon FraserUniversity, Burnaby, Canada

2Charles University, Prague,Czechia

3Indiana University, Bloomington, USA

4University of Waterloo,Waterloo, Canada

Correspondence

KarelHrazdil, Simon Fraser University,8888

UniversityDrive, Burnaby, BC V5A 1S6, Canada.

Email:karel_hrazdil@sfu.ca

Fundinginformation

CzechScience Foundation, Grant/AwardNum-

ber:15-13040S; CPA Education foundation of

BC;Social Sciences and Humanities Research

Councilof Canada, Grant/Award Number:

31-R640084

Abstract

We present a novel approach for measuring executive personality

traits.Relying on recent developments in machine learning and artifi-

cial intelligence, we utilize the IBM Watson PersonalityInsights ser-

vice to measure executive personalities based on CEOs’ and CFOs’

responses to questions raised by analysts during conference calls.

We obtain the Big Five personality traits – openness, conscien-

tiousness, extraversion, agreeableness and neuroticism – based on

which we estimate risk tolerance. To validate these traits, we first

demonstrate that our risk-tolerance measure varies with existing

inherent and behavioural-based measures (gender, age, sensitivity

of executive compensation to stock return volatility,and executive

unexercised-vested options) in predictable ways. Second, we show

that variation in firm-year level personality trait measures, includ-

ing risk tolerance, is largely explained by manager characteristics,

as opposed to firm characteristics and firm performance. Finally,we

find that executiveinherent risk tolerance helps explain the positive

relationship between client risk and audit fees documented in the

prior literature. Specifically,the effect of CEO risk-tolerance – as an

innate personality trait – on audit fees is incremental to the effect

of increased risk appetite from equity risk-taking incentives (Vega).

Measuring executivepersonality using machine-learning algorithms

will thus allow researchers to pursue studies that were previously

difficult to conduct.

KEYWORDS

big five, machine learning, personality, risk tolerance

JEL CLASSIFICATION

M12, M42, G30

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and repro-

duction in anymedium, provided the original work is properly cited.

c

2019 The Authors. Journal of Business Finance & Accounting published by John Wiley & Sons Ltd

J Bus Fin Acc. 2020;47:519–544. wileyonlinelibrary.com/journal/jbfa 519

520 HRAZDIL ET AL.

1INTRODUCTION

Upper echelons theory predicts that organizational outcomes are ‘reflections of the values and cognitive biases

of powerful actors’ and that individual executives have a significant influence on corporate policies and activities

(Hambrick & Mason, 1984, p. 193). While accounting is subject to greater regulation than other corporate activities,

research in accounting finds that top managers, particularly CEOs and CFOs, also exert significant influence on

financial reporting decisions (Plöckinger,Aschauer, Hiebl, & Rohatschek, 2016). As research on upper echelons theory

has grown, so too has the interest in developing proxies for the individual traits of executives. Over time, several

broad approaches have emerged. The ‘black box’ approach captures characteristics of individual executives using

managerial fixed effects (Dejong & Ling, 2013; Dyreng, Hanlon, & Maydew, 2010; Ge, Matsumoto, & Zhang, 2011).

While individual managers do appear to have unique financial reporting ‘styles’, this approach is unable to identify

the specific personality traits or to articulate how they relate to accounting outcomes. Another approach is to infer

personality from the actions that managers take, such as option-exercisebehaviour (Hribar & Yang, 2016; Malmendier

& Tate, 2005, 2008), or to use demographic characteristics to predict personality traits (Bamber, Jiang, & Wang,

2010). However, managerial actions are often influenced by many other factors, and demographic characteristics

such as age, gender and education have shown limited ability to explain managerial fixed effects because they do

not effectively capture the underlying personality characteristics (Ge et al., 2011). A third approach is to measure

personality traits such as the Big Five1through the administration of surveys known as personal inventory scales.

While research shows these measurements to be reliable, administering these types of tests on a large scale is not

feasible. As Ham, Lang, Seybert, and Wang(2017, p. 1090) note, ‘executives are understandably unwilling to complete

surveys or questionnaires to directly measure personality traits such as narcissism’. Plöckinger et al. (2016) call

for additional research to continue the development and validation of meaningful measures to enable closer links

between managerial idiosyncrasies and financial reporting choices, and our research is motivated by this call. The

purpose of this paper is to propose a new approach to measure executivepersonality in a large sample setting.

In contrast to previous studies, we rely on recent developments in machine learning and artificial intelligence

to measure personality traits. Specifically, we utilize the IBM Watson Personality Insights service (Watson PI) to

process transcripts of the Q&A sessions of conference calls made by CEOs and CFOs. This machine-learning software

infers personality scores for the Big Five personality dimensions from textual information using an open-vocabulary

approach. Researchers in psychology,psycholinguistics and marketing theorize that language can provide insight into

a speaker’s personality type, thinking style, and emotional states. The frequency with which speakersuse certain cate-

gories of words, for instance, can provide clues to these characteristics,and word usage in written communications can

predict aspects of personality (Fast & Funder, 2008; Hirsh & Peterson,2009; and Yarkoni, 2010). Our study focuses

on these Big Five (OCEAN) traits, as they portray basic underlying trait dimensions of personality (Goldberg,1990)

and are recognized as genetically based, relatively stable and generalizable across cultures (Cobb-Clark & Schurer,

2012; Costa & McCrae, 1997). Based on prior research that provides relativelyconsistent guidance on the relationship

between the Big Five personality traits and an individual’s appetite for risk (Borghans, Heckman, Golsteyn, & Meijers,

2009; Clarke & Robertson, 2005; Judge & Bono, 2000; Judge & Cable, 1997; Nicholson, Soane, Fenton-O’Creevy,

& Willman, 2005), we then combine the OCEAN personality traits to derive a measure for CEO and CFO risk

tolerance (RT).

Giventhe novelty of using machine learning to measure executive personality traits, we validate the WatsonBig Five

personality traits, including risk tolerance, in several ways. First, consistent with prior literature(Barnea, Cronqvist,

1The Big Five, also known as the five-factor model (FFM) developed by Norman (1963) and Costa and McCrae (1997), is one of the best-studied and the

most widely used personality models to describe how individuals generally engage with the world. The model’s dimensions are captured by the mnemonic

OCEAN, where ‘O’ stands for Openness (the extent to which a person is open to experiencing a variety of activities); ‘C’for Conscientiousness (tendency to

act in an organized or thoughtful way);‘E’ for Extraversion (tendency to seek stimulation in the company of others); ‘A’ for Agreeableness (tendency to be com-

passionate and cooperative toward others); and ‘N’ for Neuroticism (emotionalrange, the extent to which a person’s emotions are sensitive to the person’s

environment).