Soft Information in Lending: Why Relationship Lending Faded, and What Replaces It
Soft Information in Lending: Why Relationship Lending Faded, and What Replaces It
A loan officer who knows a thin-file borrower can see things the data cannot. In a stylized test, adding that judgement raised a model's Gini from 0.21 to 0.35 and cut defaults among approved loans from 9.00% to 7.27%. At about $400 per assessment, it pays for itself only on loans above roughly $46,000. That threshold, more than any technology, explains which loans still get a human, and why the middle of the small business market is underserved.
What you'll learn
Two kinds of information
Liberti and Petersen define hard information as information that can be reduced to numbers, transmitted without loss, and interpreted the same way by anyone who receives it: a credit score, a revenue figure, a payment history. Soft information is everything else: a loan officer's judgement of an owner's competence, knowledge of how a local business weathered the last downturn, the sense that a borrower's explanation of a missed payment is credible. It cannot be written down without losing much of what it is, and it is hard to verify by anyone who did not collect it.
Stein argued that the distinction shapes the organization of lending itself. Hard information can be passed up a hierarchy and acted on centrally. Soft information cannot, so lending that depends on it works best in small, decentralized institutions where the person who gathers the information also makes the decision.
This is why relationship lending has survived in some markets and vanished from others, and why its disappearance is not simply a story about technology replacing people.
What soft information is worth
Take a segment of thin-file applicants, where hard data separates good from bad borrowers poorly. In a stylized simulation, the hard-information model alone produces a Gini of 0.21. Adding a soft assessment from a loan officer who knows the borrower raises it to 0.35. At a 60% approval rate on a segment with an 11.5% base default rate:
| Information used | Gini | Default rate among approved |
|---|---|---|
| Hard information only | 0.21 | 9.00% |
| Hard plus soft information | 0.35 | 7.27% |
A 1.73-point reduction in default among approved loans is a large improvement. The question is what it costs. Suppose a soft assessment takes a loan officer's time worth about $400 per application, against about $15 for an automated hard-data decision.
- The benefit per application is 60% approved, times 1.73 points of default avoided, times an 80% loss given default, times the loan size: about 0.83% of the loan.
- The extra cost is $385 per application.
- So soft information pays for itself only on loans above roughly $46,000.
The threshold explains the shape of the market. Small consumer loans fell almost entirely to automated hard-data decisions because soft information was never economic there. Commercial lending above the threshold kept relationship officers. Per our market structure analysis, the middle, small business loans of $10,000 to $50,000, is where both approaches are marginal, which is why it is the segment most consistently underserved.
What the research found
Petersen and Rajan, studying small US firms, found that longer and broader lending relationships were associated with greater availability of credit, more than with lower prices. Berger and Udell found that firms with longer banking relationships paid lower rates and were less likely to pledge collateral on lines of credit. Both results fit the same picture: a lender that has accumulated soft information over time can extend credit on terms an outside lender, holding only hard data, could not justify.
The relationship is therefore an asset, and it has two owners. The borrower benefits from credit that hard data alone would not support. The lender holds information competitors lack, which gives it some ability to charge more later. Per our exit analysis, that second effect is the informational version of the loyalty penalty.
What is lost when it disappears
Per our branch analysis, branch closures and centralized credit decisions have removed the people who held soft information from much of the market. The loss is not evenly distributed.
- Borrowers whose quality is not in the data lose most. Per our legibility analysis, a business with irregular revenue and a strong owner looks the same in hard data as one with irregular revenue and a weak owner.
- Information that existed is not recorded. Per our institutional memory analysis, soft information leaves with the person who held it.
- Discretion moves elsewhere. Per our manual review analysis, some judgement survives as exception review, but reviewers without a relationship are applying a different kind of judgement.
Soft information has a dark side
The same discretion that captures a good owner's competence also captures a loan officer's biases. Soft judgement is harder to audit, harder to explain to a declined applicant, and per our fair lending analysis, harder to test for disparate treatment. Part of the shift to hard information was a deliberate response to documented discrimination in discretionary lending, and that motivation was legitimate. Any revival of soft information has to answer it.
Hardening soft information
The most important development is not the replacement of soft information but its conversion. Some of what a relationship officer knew can now be measured.
| What the officer used to judge | What can now be measured |
|---|---|
| How the business manages cash | Per our cash flow analysis, account data shows buffers, volatility and stress months directly |
| Whether the owner pays obligations | Reported tradelines in the business's own name |
| How the business weathered a bad season | Several years of transaction history |
| Whether the owner is reliable | Per our signaling analysis, a voluntarily built record of obligations met |
| Whether customers are stable | Receivables concentration from invoicing data |
Hardening lowers the cost of the information and so lowers the threshold. If a soft assessment costing $400 can be replaced by a data-derived feature costing $20 with most of its predictive power, loans far below $46,000 can benefit. Per our credit gap analysis, that is the most plausible route to closing the gap in the middle of the market. What cannot be hardened, such as judgement about an owner's character or a new business plan, remains expensive and remains confined to larger loans.
The case that soft information was never as valuable as claimed
Studies comparing discretionary and automated decisions have often found that loan officers' overrides of models perform no better, and sometimes worse, than the model alone, which suggests that much of what passed for soft information was noise or bias. That finding is real, and it is consistent with the analysis here: where hard data is informative, soft judgement adds little and its cost is not justified. The claim for soft information is narrower. It is that in segments where hard data is weak, such as thin files and young businesses, information held by someone close to the borrower adds discrimination that data cannot supply, and that the question is always whether the loan is large enough to pay for it.
Testable implications
- The share of loans decided with human judgement should rise with loan size, with a threshold that falls as information costs fall.
- Override performance should beat the model only in segments where the model's own discrimination is low.
- Branch closures should reduce credit most for informationally opaque borrowers and least for those with thick files.
- New hard data sources should increase lending most in the loan-size band just below the soft-information threshold.
The fourth implication is the one to watch. Per our renewal underwriting analysis, a lender's own repayment history with a borrower is the cheapest soft information to harden, because it already sits in the lender's systems.
Frequently asked questions
Hard information is quantifiable and transmissible; soft information is judgement that is hard to write down or verify by others.
Here only on loans above about $46,000, where 0.83% of the loan exceeds the $385 extra cost of a human assessment.
Longer relationships were associated with more available credit, lower rates on lines and less collateral, alongside some lender pricing power.
Partly. Much of what officers judged can now be measured, lowering the threshold, but judgement about character and new plans cannot.
Key takeaways
- Hard information is quantifiable and transmissible; soft information is judgement that is hard to record or verify.
- Soft information favors small, decentralized lenders where the person who collects it decides.
- In a stylized thin-file segment, soft judgement raised Gini from 0.21 to 0.35 and cut approved defaults from 9.00% to 7.27%.
- At $400 per assessment, it pays only on loans above about $46,000.
- That threshold explains why small consumer loans automated and larger commercial loans kept relationship officers.
- Research links longer relationships to more available credit, lower rates on lines and less collateral.
- Soft judgement also carries bias that is hard to audit, which was part of the legitimate shift to hard data.
- Hardening soft information through account, tradeline and invoicing data lowers the threshold for smaller loans.
References
- Petersen, M. A., & Rajan, R. G. (1994). The Benefits of Lending Relationships: Evidence from Small Business Data. Journal of Finance, 49(1), 3–37.
- Berger, A. N., & Udell, G. F. (1995). Relationship Lending and Lines of Credit in Small Firm Finance. Journal of Business, 68(3), 351–381.
- Stein, J. C. (2002). Information Production and Capital Allocation: Decentralized versus Hierarchical Firms. Journal of Finance, 57(5), 1891–1921.
- Liberti, J. M., & Petersen, M. A. (2019). Information: Hard and Soft. Review of Corporate Finance Studies, 8(1), 1–41.
This analysis is educational and does not constitute financial or legal advice. All figures are stylized illustrations from a simulated segment with assumed default rates, information quality, approval rates, assessment costs and loss given default, and do not describe any particular lender or portfolio. The cited studies are summarized for context; consult the original works for their data and conclusions.