Everyone Looking at the Same Thing: The Monoculture in Credit Assessment | HL Hunt
Everyone Looking at the Same Thing: The Monoculture in Credit Assessment
Most lenders assess applicants using the same underlying data, attributes computed the same way, and models trained on similar populations to predict similar outcomes. Which means independent institutions, each deciding in good faith, reach highly correlated conclusions about the same person. That has a consequence nobody chose: for the population the data assesses poorly, exclusion is absolute rather than competitive. There is no lender who sees you differently, because there is nothing different to see with. Diversity of judgment was itself a form of access, and it was removed as a side effect of getting assessment right.
In this report
Where the convergence comes from
Four layers, each independently sensible, compounding into one.
| Layer | Variation across lenders |
|---|---|
| Underlying data | Very little — the same bureaus, the same furnishers |
| Attributes | Some, per our attribute analysis — and drawn from the same raw data |
| Models | Different specifications, similar inputs and similar targets |
| Policy rules | Genuine variation, applied to the same score |
Row one is doing most of the work. Per our coverage analysis, the system records credit accounts and little else — so every lender is looking at the same narrow slice of a person's financial life, and no amount of modelling sophistication generates information that isn't in the data.
Two lenders with completely different models will still rank a population similarly if they're reading the same file. Model diversity is not input diversity, and only the second produces genuinely different answers.
And per our vendor analysis, a substantial share of the market uses purchased scores — so for many decisions the convergence isn't approximate, it's literal: the same model is deciding.
Model diversity is not input diversity. Two different models reading the same file rank people the same way.
Decline as verdict
The consequence for the individual, and it's a different thing from what a decline is supposed to be.
A decline is meant to be one institution's assessment. The framing behind adverse action requirements, behind the advice to shop around, and behind the idea of a competitive credit market is that lenders differ.
Where they don't:
- The decline is close to a property of the applicant rather than of the relationship.
- There's no second opinion available, because the second opinion consults the same source.
- The reasons given are accurate and unactionable in a specific way — per our explainability analysis, "insufficient credit history" is a true statement about a file that no lender can look past.
Which sharpens the finding in our rationing analysis. That report noted the exclusion error is invisible because it generates no data. This adds that it's also absolute: the excluded aren't distributed across lenders with different appetites, they're excluded everywhere at once.
And per our thin-file analysis, the population this bites hardest on is the one whose performance is better than the file suggests. Their reliability exists; nothing in the market can see it.
Why shopping around fails
The practical consequence, and it contradicts standard advice.
"Apply elsewhere" assumes variation. Where assessment converges:
- The second application reads the same file.
- It produces a similar assessment.
- It declines.
- An inquiry is added, per our inquiries analysis.
- The third application sees a file with more inquiries.
- Which is itself a negative, so the answer gets worse.
Searching for a second opinion actively worsens your position in a market where no second opinion exists. That's a genuinely perverse structure — the recommended response to a decline makes the next decline more likely — and per our decline guide, it's why the right response is to address the stated reason rather than to reapply.
Where variation does survive, it's worth naming because it's the exception: lenders with genuinely different criteria, different data, or an existing relationship with you. Per our renewal analysis, an institution holding your behavioural data is the one place a meaningfully different view exists.
Tightening together
The market-level consequence, and it's the correlation finding from our correlation analysis operating on the supply side.
When lenders share inputs and methods, they respond to conditions in the same direction at the same time.
- Deteriorating data affects everyone's models simultaneously.
- Utilization rising across a population lowers scores everywhere at once.
- Similar model structures respond similarly to the same shifts.
- Everyone tightens together, which contracts supply just as demand rises.
Which is the mechanism behind the finding in our shock analysis — that credit contracts precisely when it's needed. That report treated it as a property of credit; this identifies why it's synchronized rather than staggered. If lenders assessed differently, some would tighten and some wouldn't, and the aggregate contraction would be smoother.
And a second-order effect worth naming: the household experiences simultaneous withdrawal as personal. Three lenders reducing lines in the same month reads as a judgment about you, when it's three institutions responding to the same data movement.
Correlated errors
The property that makes this more than an access story.
Shared inputs mean shared blind spots. Whatever the standard file fails to capture, it fails to capture for everyone — so an error in the common view is an error across the whole market rather than a competitive opportunity for whoever avoided it.
Where this shows up:
- Obligations the file doesn't hold, per our coverage analysis — every lender underestimates the same borrowers' commitments.
- An attribute definition change, per our attribute analysis, which shifts behaviour at every lender consuming it.
- A furnisher reporting incorrectly, which propagates to everyone.
- A population the data misrepresents, which is misassessed uniformly.
- Per our definition analysis, a shared outcome convention, which means everyone is predicting the same possibly-wrong thing.
The implication for a single lender is uncomfortable. Your model's validation will show it performing in line with the market, because the market shares your errors — so "our performance matches benchmark" is consistent with the whole benchmark being wrong in the same direction. Per our validation analysis, that's a limitation of comparative benchmarking that doesn't get stated.
And per our override analysis, the evidence that the common view misses things already exists — near-cutoff approvals outperform predictions, consistently, at multiple lenders. That's the monoculture's blind spot showing through the one mechanism that partially escapes it.
What standardization bought
Stated properly, because this report would be badly misread as an argument for going back.
Standardized, portable assessment was an enormous improvement and the reasons are not marginal:
- It made credit available without a relationship. Per our verification analysis, a portable file lets a stranger lend to a stranger — which is the foundation of most consumer credit.
- It removed a great deal of arbitrary and discriminatory judgment. Local discretion was not neutral, and replacing it with a documented rule was a substantial gain for people who were on the wrong side of that discretion.
- It made decisions reviewable. A rule can be tested for disparate impact; a loan officer's impression cannot.
- It lowered costs enormously, per our pricing analysis — and those are the fixed costs that dominate small-loan pricing.
- It made assessment contestable, since a file can be corrected and a judgment can't.
None of this is in tension with the argument. The claim is that standardization had a cost — the loss of independent judgment as a route to access — that nobody accounted for because it was invisible and because the thing replaced was genuinely worse in other respects.
The answer isn't less standardization. It's more standardized inputs — which is precisely what our coverage analysis argues for, and which would produce variation without reintroducing discretion.
What genuine variation requires
The constructive conclusion, and the test is sharp.
Variation requires different inputs, not different weightings of the same inputs. Two lenders weighting the same attributes differently will disagree at the margin; two lenders looking at different things will disagree about specific people.
What produces real variation:
- Cash flow data, per our cash flow analysis — it disagrees with the file about particular applicants, which is the whole point.
- Rent and utility payment history, per our coverage analysis, which assesses people the file can't see at all.
- Behavioural data from an existing relationship, per our renewal analysis.
- Channel and context, per our channel analysis, where an embedded lender knows something about the transaction.
- Different outcome definitions, which change what's being predicted rather than how well.
And the commercial argument is stronger than the social one. A lender using a genuinely different input identifies people the rest of the market can't see — which is an underwriting advantage on a population nobody is competing for, at prices set by everyone else's inability to assess them.
That's the clearest statement of why this matters to practitioners rather than only to policy. In a monoculture, the returns to a different input are unusually high, because the competition literally cannot make the same decision.
The strongest objections
"Lenders do differ — criteria and appetites vary widely." True at the margin and the variation is mostly in where the line is drawn rather than in how people are ranked. A subprime lender and a prime lender disagree about which scores to accept and largely agree about who scores what. Variation in cutoffs is not variation in judgment, and it's the ranking that determines whether a second opinion exists.
"Alternative data is growing." It is, and that's the report's recommendation rather than a refutation. The relevant question is what share of decisions currently rest on genuinely different inputs, and for most consumer credit the answer remains small. The trend is the right one and the current state is the subject.
"You haven't measured the correlation." Correct, and it's the central limitation. Measuring how correlated lenders' decisions actually are would require data no single institution holds — you'd need the same applicants assessed by multiple lenders, which only happens in marketplaces and isn't published. The argument is from the shared input structure, not from measured agreement.
Testable implications
- Applicants declined by one lender should be declined by others at rates far above chance, controlling for score — measurable in marketplace data where the same application reaches several lenders.
- Lenders using genuinely different inputs should disagree with the market more than lenders using the same inputs differently weighted.
- Approval rates should move together across lenders in response to common data shifts, more than fundamentals justify.
- Near-cutoff declines should outperform predictions at multiple lenders simultaneously, which is the shared blind spot and is already partly evidenced.
- Attribute definition changes should shift decisions market-wide, observable as synchronized approval-rate movements with no macro cause.
- Lenders adding a genuinely new input should find underpriced applicants — the commercial prediction, and the one a single firm can test.
The first is the direct test and the data exists in marketplace platforms. If an applicant declined by one lender is declined by the next four at rates approaching certainty, then shopping around is not a remedy and consumer guidance that recommends it is describing a market that doesn't exist.
The conclusion we'd hold: credit assessment converged for good reasons, and convergence removed something that was doing real work. Independent judgment was a route to access for people the data assesses badly, and it was eliminated by a process that improved almost everything else. The fix is not discretion — it's giving lenders different things to look at.
Frequently asked questions
Most lenders using the same data, similar attributes, and similar models, so independent institutions reach highly correlated conclusions about the same person.
It removes the second opinion. A decline becomes closer to a verdict than an opinion, and exclusion is absolute rather than distributed across lenders with different appetites.
Substantial ones — portability, the removal of arbitrary local judgment, reviewability, and enormous cost reduction. The argument is that it had an unaccounted cost, not that it was a mistake.
Different inputs rather than different weightings — cash flow, rent and utility history, relationship behaviour. These disagree with the file about specific people, which is what a second opinion requires.
Key takeaways
- Model diversity is not input diversity; different models reading the same file rank people the same way.
- Where assessment converges, shopping around adds inquiries and produces the same answer, making the search self-defeating.
- Shared inputs synchronize tightening, which is why credit contracts all at once rather than gradually.
- Shared blind spots mean a model performing in line with benchmark is consistent with the benchmark being uniformly wrong.
- Standardization removed arbitrary discretion and made decisions contestable — the fix is more standardized inputs, not less standardization.
- In a monoculture the returns to a genuinely different input are unusually high, because competitors cannot make the same decision.
This report presents an analytical framework and the authors' interpretation; it is not financial, legal, or policy advice. The degree of correlation between lenders' decisions is argued from the structure of shared inputs rather than measured, and no estimate of it is offered here; data that would establish it is not generally published. The implications identified as testable are hypotheses.