Technology is changing lending: Implications for research

Abstract

Costello, Down, and Mehta (2020) trace their slider intervention to deviations from the credit line amount recommended by a credit scoring model. The deviations are followed by larger delinquency declines and bigger sales orders, and Costello et al. interpret these results using discretion-based theories. However, incremental deviations are concentrated on newer clients rather than those the lender has accumulated soft information about. Deviations also appear larger for public than private borrowers. My discussion evaluates whether these results align with discretion-based theories, and explores alternative interpretations based on salience and unique aspects of the trade credit setting. Differences in interpretation aside, the evidence is informative about technological advances in commercial lending. I conclude with an overview of several recent advances and discuss the implications for lending research.

Introduction

Recent advances in technology have transformed how lenders originate commercial credit and monitor borrowers. Information sharing technologies have proliferated as the costs of gathering and verifying information have declined (Djankov et al., 2007; Liberti et al., 2020). As a result, lenders have access to a growing array of credit scores and default prediction models, all seeded with increasingly timely and rich firm, industry, macroeconomic, and household data. When screening and monitoring borrowers, lenders are also expanding the types of information they access, including digital footprints (Berg et al., 2020), social media (Lin et al., 2013), and collateral surveillance. Investments in technology also enable lenders to share data from internal reporting systems with their own loan officers and branches (Campbell et al., 2019; Kang et al., 2020).

These developments are important because information asymmetries prevent lending (Stiglitz and Weiss, 1981). Commercial lenders engage in screening and monitoring to help overcome these asymmetries, and face tradeoffs surrounding the information sources they access (Liberti and Petersen, 2019). Much lending research is devoted to understanding the role of auditor-verified financial statements. As Watts and Zimmerman (1983) argue, verified financial statements are used when they are “an efficient method of monitoring contracts between managers and those supplying capital” (p. 626).

Although publicly held borrowers undergo audits by mandate, audits are costly and U.S. privately held borrowers can bargain with their lender over reporting and verification requirements (Allee and Yohn, 2009; Minnis, 2011; Cassar et al., 2015; Lisowsky et al., 2017; Duguay et al., 2020). Therefore, the continued demand for verified financial statements is far from guaranteed: “In general, of course, it will pay the owner-manager to engage in bonding activities and to write contracts which allow monitoring as long as the marginal benefits of each are greater than their marginal cost” (p. 326 Jensen and Meckling, 1976; emphasis added). The arrival of new information sources such as business credit scores can reduce the marginal benefit of verified financial statements to lenders. Tellingly, one of the world's largest business credit scoring companies advertises to lenders by saying “financial statements not required” (Figure 1). In light of this, research studying the development and use of such alternative information sources is important to understanding the continued role of verified financial statements in commercial lending.

Costello, Down, and Mehta (2020) (henceforth CDM) analyze lending decisions and contract performance using a randomized control experiment with a sample of trade creditors using Credit2B. Credit2B is similar to vendors of other data studied in the literature including PayNet and Dun & Bradstreet. They are a third-party information intermediary (they do not lend), and they combine lender-provided information with external data on both the economy and specific borrowers to create credit scores and other data products. CDM's experiment involves assigning half of sample lenders (the treated group) a “slider” that enables them to deviate from the credit line amount recommended by the Credit2B “machine”-generated model. The control group can also deviate, but does not receive the slider reminding them of this capability.

Studying loan officer behavior and trade credit in particular is difficult because of the paucity of public data. Additionally, credit line originations and outcomes are influenced by a variety of factors that can be hard to observe and disentangle. CDM deserve credit for bringing the Credit2B data into the literature and constructing a randomized experiment. Their estimation is performed within borrower-month and borrower-lender relationship. In this way, the analysis accounts for unobservable time-varying shocks to the borrower's demand for credit, as well as time-invariant aspects of the lending relationship.

CDM find that treated lenders deviate more, and their deviations are associated with a greater decline in delinquencies, on average. Treated lenders also experience a significantly larger increase in sales orders. The sizable origination and performance responses they document are important given the links between 1) information asymmetry and access to credit, and 2) borrower delinquencies and financial stability. The results also reinforce how, holding constant borrower-lender reporting, shifts in the other information lenders use generate meaningful fluctuations in lending outcomes.

CDM use relationship lending frameworks to interpret the deviations occurring during their experiment (Sharpe, 1990; Boot, 2000): “If discretion allows loan officers to incorporate private information, then it may help overcome information asymmetry problems that are prevalent when dealing with opaque borrowers” (p. 4).

The discretion literature commonly uses borrower ownership type, relationship length, and size as opacity proxies to evaluate when a loan officer's private information comes into play (Petersen and Rajan 1994; Berger et al., 2005; Degryse and Ongena, 2005; Sutherland, 2018; Liberti and Petersen, 2019). CDM find nearly three-times larger deviations for newer than longstanding clients, and deviations appear larger for public than private borrowers.¹ Deviations also result in a larger decline in delinquencies. While these results are fascinating in their own right, they are not in line with loan officers using soft information they have accumulated over time to finance opaque, constrained firms and invest in relationships with them. As a result, it is unclear how CDM's evidence speaks to discretion-based theories or the interaction between technology and relationship lending.

Section 2 considers whether salience or trade credit theories might shed light on the collection of results. Salience is relevant because discretion plays a major role in the pre-experiment period: loan officers originate credit lines that rarely correspond to the model recommendation. The average pre-experiment absolute deviation from the model exceeds 50% and the majority of credit lines display a material degree of deviation. Then, the slider, and its introduction in the experiment, reminds the loan officer of the abundant discretion they already have. Such a reminder would be less necessary for private borrowers whose soft information is so important, but could cue the loan officer to consider deviating from the model for public borrowers and new clients when they might not have previously. Perhaps unintentionally, the authors' and Credit2B's communications with loan officers provided additional cues to loan officers about using discretion where it might not otherwise apply.

It is also possible that discretion-based theories developed for banking do not transfer so easily to this specific experiment and trade credit setting. In trade credit, the parties are bargaining over both the purchased good and its financing. Additionally, the financing terms can differ from those in most commercial loans.

Setting aside differences in interpretation, the changes in credit line amounts and delinquencies CDM document are significant, worthy of investigation, and add to recent evidence on how technological advances are transforming commercial lending. In particular, their evidence on borrower social media presence is novel and informative about how AI models process new information sources. Section 3 outlines several implications of these advances for research on commercial lending markets. I focus on work concerned with the role of verified financial statements in debt contracting and the interaction between transparency, competition, and financial stability. Section 4 concludes.