Assessing Human Information Processing in Lending Decisions: A Machine Learning Approach

• Published date: 01 May 2022
• Author: MIAO LIU
• Date: 01 May 2022
• DOI: http://doi.org/10.1111/1475-679X.12427

DOI: 10.1111/1475-679X.12427

Journal of Accounting Research

Vol. 60 No. 2 May 2022

Printed in U.S.A.

Assessing Human Information

Processing in Lending Decisions: A

Machine Learning Approach

MIAO LIU ∗

Received 1 December 2020; accepted 28 January 2022

ABSTRACT

Effective financial reporting requires efficient information processing. This

paper studies factors that determine efficient information processing. I

∗Carroll School of Management, Boston College.

Accepted by Rodrigo Verdi. This paper is based on my dissertation. I am grateful to

my dissertation committee members Philip Berger, Christian Leuz (chair), Sendhil Mul-

lainathan, and Valeri Nikolaev for their guidance and support. I also appreciate helpful com-

ments from an anonymous referee, Ray Ball, Pietro Bonaldi, Jonathan Bonham, Matthias

Breuer, Robert Bushman, Jung Ho Choi, Anna Costello, Hans Christensen, John Galle-

more, Pingyang Gao, Joao Granja, Anya Kleymenova, Yun Lee (discussant), Rebecca Lester,

Ye Li, Jinzhi Lu, Yao Lu, Mark Maffett, Charles McClure, Danqing Mei, Michael Minnis,

Maximilian Muhn, Sanjog Misra, Thomas Rauter, Ethan Rouen, Haresh Sapra, Andriy Shk-

ilko (discussant), Douglas Skinner, Gurpal Sran, Andrew Sutherland, Rimmy Tomy, James

Traina,Felix Vetter, Xian Xu, Anastasia Zakolyukina, workshop participants at Chicago Booth,

Harvard Business School, MIT, Boston College, University of Florida, Rochester Univer-

sity, UT Austin, 2021 JAR Conference, Singapore Management University Emerging Scholar

Forum, CMU emerging scholar session, University of Miami Winter Conference on Ma-

chine Learning and Business, and LBS Trans-Atlantic Doctoral Conference. I am indebted

to the company executives and many loan officers who preferred to remain anonymous

for helpful discussions and access to data. I do not have a financial interest in the out-

comes of this research. I gratefully acknowledge financial support from the University of

Chicago Booth School of Business. Research support from the Sanford J. Grossman Fellow-

ship in Honor of Arnold Zellner is gratefully acknowledged; any opinions expressed herein

are the author’s and not necessarily those of Sanford J. Grossman or Arnold Zellner. This

study is exempt from further review by the Institutional Review Board at the University of

Chicago. Any errors are my own. An online appendix to this paper can be downloaded at

http://research.chicagobooth.edu/arc/journal-of-accounting-research/online-supplements.

607

© 2022 The Chookaszian Accounting Research Center at the University of Chicago Booth School of

Business

608 m. liu

exploit a unique small business lending setting where I am able to observe

the entire codified demographic and accounting information set that loan

officers use to make decisions. I decompose the loan officers’ decisions into a

part driven by codified hard information and a part driven by uncodified soft

information. I show that a machine learning model substantially outperforms

loan officers in processing hard information. Loan officers can only process

a sparse set of useful hard information identified by the machine learning

model and focus their attention on salient signals such as large jumps in cash

flows. However, the loan officers use salient hard information as “red flags”

to highlight where to acquire more soft information. This result suggests that

salient information is an attention allocation device: It guides humans to al-

locate their limited cognitive resources to acquire soft information, a task in

which humans have an advantage over machines.

JEL codes: C55, D9, D83, G4, G21, M41

Keywords: human information processing; cognitive constraints; soft infor-

mation acquisition; salience; attention allocation

1. Introduction

Research has shown that financial reporting facilitates decision making

and affects a wide range of financial and real outcomes (Beyer et al.

[2010], Dechow, Ge, and Schrand [2010], Leuz and Wysocki [2016],

Roychowdhury, Shroff, and Verdi [2019]). The extent to which the

objective of financial reporting is met, however, depends on how decision-

makers process the information. One strand of research in economics

and psychology highlights human weaknesses in processing information,

because of cognitive constraints (Blankespoor, Dehaan, and Marinovic

[2020]) and behavioral biases (Kahneman [2011]). Another strand in

accounting and finance, however, emphasizes humans’ strong ability to

discover new information (Goldstein and Yang [2017]), especially soft

information (Liberti and Petersen [2018]). In this paper, I investigate what

factors determine these seemingly contradictory strengths and weaknesses

in information processing by loan officers in their lending decisions, using

a machine learning model–based decision rule as a benchmark.

To examine the efficacy of this decision making, I rely on more than

30,000 detailed loan contracts from a large Chinese small business lender

during the period of 2011 to 2015. Loan officers observe hard information

on borrowers’ demographics, accounting reports, credit history, and ex-

ercise discretion to acquire additional soft information by making phone

calls. As a newly founded company in 2011 in a nascent market, the lender

has no internal credit model or credit rating from a third party to aid loan

officers in making decisions. Although the loan officers might not pro-

cess hard information efficiently, as predicted by theories of cognitive con-

straints and behavioral biases, their efforts to collect soft information by

interacting directly with borrowers should improve their lending decisions,

as demonstrated in other similar settings (e.g., Petersen and Rajan [1994,

a machine learning approach 609

1995]). In addition, the officers determine when and how to acquire soft

information after observing hard information, suggesting the two types of

information may interact. I use this setting to study which factors impede

efficient hard information processing and how these factors further affect

soft information acquisition.

Three challenges emerge when assessing information processing effi-

ciency. First, it is difficult to study how information users process informa-

tion if the underlying information is unobservable, as is usually the case.

Second, which information to collect is often a choice. Consequently, dif-

ferent people might appear to process information differently not because

they have varying processing skills but because they have different informa-

tion sets. Third, evaluating errors requires a benchmark. In other words,

what is the “correct” way to process a given piece of information? Although

a benchmark can often be established in a laboratory (such as the correct

answer to a test), one is usually missing in the real world.

I combine my unique setting with a novel research design to overcome

these challenges. Two key features of the setting help address the first two

challenges. First, the data allow observation of the loan officers’ entire hard

information set about a borrower, sidestepping the unobservability prob-

lem. Second, borrowers are randomly assigned to loan officers, meaning

each officer has the same pool of borrowers on average. As a result, any

systematic difference in lending decisions across officers stems from their

differing abilities in processing information and not because they are en-

dogenously matched with different types of borrowers, helping overcome

the second challenge that the information set is usually an endogenous

choice.

To address the third challenge, I use a machine learning model as a

benchmark to assess human decisions. I split my data randomly into a train-

ing sample and a hold-out sample. I train a machine learning model on

the training sample to predict a borrower’s repayment and design a fea-

sible lending decision rule by reallocating larger loans to borrowers who

are more likely to repay, as predicted by the model. Next, using the hold-

out sample, I show my first result that this model-based decision rule can

boost the lender’s profit by at least 38%, making it a valid benchmark for

examining loan officers’ limitations in processing hard information.

To make the machine learning model a benchmark, I must address the

fact that, although it only uses codified hard information as an input, loan

officers can acquire additional soft information, mainly by making phone

calls to the borrower. I decompose officers’ decisions into a part driven

by codified hard information and a part driven by uncodified soft infor-

mation. Specifically, I fit a separate machine learning model for each loan

officer, this time to predict the officer’s lending decisions based on hard

information. Unlike the first model, which predicts borrower repayment,

the purpose of this model is to mimic how each loan officer processes hard

information. Soft information is then captured by the residual, as it rep-

resents variation in officers’ decisions that cannot be explained by hard