Impact of mortgage soft information in loan pricing on default prediction using machine learning

• Published date: 01 March 2023
• Author: Thi Mai Luong,Harald Scheule,Nitya Wanzare
• Date: 01 March 2023
• DOI: http://doi.org/10.1111/irfi.12392

ORIGINAL ARTICLE

Impact of mortgage soft information in loan

pricing on default prediction using

machine learning

Thi Mai Luong | Harald Scheule | Nitya Wanzare

Finance Discipline Group, UTS Business

School, University of Technology Sydney,

Sydney, New South Wales, Australia

Correspondence

Harald Scheule, Finance Discipline Group,

UTS Business School, University of

Technology Sydney, PO Box 123, Broadway,

Sydney, NSW 2007, Australia.

Email: harald.scheule@uts.edu.au

Funding information

Australian Prudential Regulation Authority;

Brian Gray Scholarship of the Australian

Prudential Regulation Authority; Hong Kong

Institute for Monetary Research

Abstract

We analyze the impact of soft information on US mortgages

for default prediction and provide a new measure for

lender soft information that is based on the interest

rates offered to borrowers and incremental to public

hard information. Hard and soft information provide for

a variation in annual default probabilities of approxi-

mately 3%. Soft information has a lesser impact over

time and time since origination. Lenders rely more on

soft information for high-risk borrowers. Our study

evidences the importance of soft information collected

at loan origination.

KEYWORDS

credit risk, default, hard information, lending, mortgage,

prediction, pricing, soft information, yield spreads

JEL CLASSIFICATION

G01, G20, G21, C51, C55

1|MOTIVATION

Mortgage lending plays a crucial role in financial markets and accounts for a high proportion of commercial

banks' balance sheets.

1

Home ownership and finance play key roles in consumers' lives. However, mortgage

lending was found to be a critical cause of the Global Financial Crisis of 2008–2009. Financial institutions rely

Received: 7 August 2021 Revised: 17 June 2022 Accepted: 6 August 2022

DOI: 10.1111/irfi.12392

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which

permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no

modifications or adaptations are made.

© 2022 The Authors. International Review of Finance published by John Wiley & Sons Australia, Ltd on behalf of International Review

of Finance Ltd.

158 International Review of Finance. 2023;23:158–186.

wileyonlinelibrary.com/journal/irfi

on credit scoring of borrowers to determine their ability to repay debt for loan pricing at origination and refi-

nance, loan provisioning and calculating bank capital requirements.

Lenders include both “hard”and “soft”information to assess the likelihood and magnitude of future loan losses

and then make lending decisions. We follow the definition by Stein (2002) in our paper and define hard information

as information that is measurable, digitally stored, and publicly verifiable by all lenders. Such information is easily

incorporated into scoring models and applied to a large number of borrowers. On the other hand, individual lenders

collect private soft information based on personal interactions, customer visits, and trust relationships between

lenders and customers. This kind of information is difficult to quantify and verify but remains significant in credit

assessment processes.

To illustrate, a borrower submits her mortgage loan application to a lender and loan officers analyze hard infor-

mation such as the FICO score, debt-to-income (DTI) ratio and loan-to-value (LTV) ratio. They further collect soft

information from interviews or other sources, such as transactional records and determine lending terms. Soft infor-

mation may result in the rejection (acceptance) of a loan application if the loan officer believes that the borrower is

riskier (less risky) than a threshold. In addition, the assessment of risk will be embedded in the pricing process,

whereby customers may have to pay a higher interest rate for a higher risk. It is apparent that soft information gath-

ered by lenders adds a human touch to the approval and pricing process.

2

This paper provides a new measure for lender soft information based on the interest rates offered to borrowers

incremental to measurable hard information. We rely on the credit spreads embedded in loan rates offered to bor-

rowers following the accept decision of lenders. Through the credit spread channel, we identify and measurethe net

position of soft information in which the more adverse the soft information is, the higher the credit spread on loans

is, and hence the higher the default likelihood is. This approach is novel as it defines soft information as the variation

after controlling for hard information. It includes soft information that may be observed through proxy variables such

as geographic distance between borrowers and lenders as well as soft information that is not observed. Further, this

paper also tests the additional accuracy of our soft information measure in predicting credit risk.

The econometric contributions of this approach are twofold. First, we are able to measure unobserved soft

information. Prior literature considers observed soft information via proxy variables but has not considered soft

information that is not observed by such variables. Second, soft information is measured in our paper on a continu-

ous scale and on an interpretational level of the credit spread. A higher value corresponds to a higher credit spread.

Prior literature has confirmed the existence of soft information but not the degree to which soft information predicts

credit risk.

Relative to the existing finance literature, we make the following contributions in this paper. First, we provide

evidence that soft information is predictive for mortgage default risk. Hard and soft information provide for a varia-

tion in annual default probabilities of 3%. Second, we find that soft information is less predictive over time and time

since origination. This may signal a shrinkage of soft information collection and depreciation since loan origination.

Third, we document that lenders rely more on soft information for high-risk borrowers as more soft information is

collected and priced for borrowers when information is more binding as information has a greater sensitivity on

default risk.

This paper is organized as follows, Section 2reviews related literature. Section 3outlines the model framework.

Section 4describes the data used, main empirical analysis and robustness checks. Section 5provides an economic

impact analysis and Section 6discusses our findings and concludes.

2|LITERATURE REVIEW

The existing literature supports the importance of soft information in business and consumer lending. There is a vast

literature on relationship lending to firms.

3

Examples include Stein (2002), DeYoung et al. (2008), Degryse and Van

Cayseele (2000), Chakraborty and Hu (2006), and Brick and Palia (2007). For example, Stein (2002) argues for the

LUONG ET AL.159