Loan officers versus algorithms
Soft information and relationship lending appear on 26 recent papers in the finance and economics journals, but only 5 of them carry an algorithm, machine-learning, fintech or automation keyword. As banks automate small-business lending, what is lost when the loan officer's judgment goes, and can models learn it?
What prompted this#
Since October 2025, 26 papers carry a loan-officer, relationship-lending or soft-information keyword (15 in finance journals, 7 in economics): "soft information" on 19, "relationship lending" on 6. Only 5 of them also carry an algorithm, machine-learning, fintech, automation, digital or hard-information keyword.
Questions#
- Do loans approved by loan officers against the model's recommendation default more or less often than loans the model approved?
- Does automation cut lending to borrowers whose strengths are not in the data (new firms, immigrants, informal businesses), and does it reduce discrimination?
- Can text from loan applications and site visits give models the soft information officers use?
Data that could help#
Loan application files with both officer decisions and model scores from a lender, and loan performance.
Next step#
Partner with a lender that records officer overrides of model scores and compare the default rates of overridden and non-overridden loans.