Legibility: When the Problem Is Being Hard to Describe | HL Hunt
Legibility: When the Problem Is Being Hard to Describe
An institution can only act on what it can read, so anything that does not fit a field must be converted by the applicant, at their cost, with no guarantee it works. Three households each earning $50,400 a year meet three completely different processes: two pay stubs for one, twelve months of statements and a human reader for another. A meaningful share of what gets priced as credit risk is illegibility, and the two are conflated because they produce the same output and the decline reason rarely tells them apart.
What you'll learn
The core thesis
An institution can only act on what it can read. Its forms have fields, its models have inputs, and its procedures have categories. Anything that does not fit one of them has to be converted into something that does, and the conversion is done by the applicant, at their own cost, with no guarantee it succeeds.
Our position is that a meaningful share of what gets priced as risk is actually illegibility: the applicant is not harder to collect from, they are harder to describe. The two are routinely conflated because they produce the same output, which is a worse offer or no offer.
This is not a complaint about bureaucracy. Standardized categories are what make assessment possible at scale, and per our fixed cost analysis the alternative, which is individual judgment on every case, is more expensive and less consistent. The claim is narrower: the categories were drawn around the typical case, and the cost of being atypical is charged to the atypical person rather than treated as a cost of the system's design.
Three identical incomes
Take three households, each earning exactly $50,400 a year.
| Shape | Annual total | Monthly standard deviation | Months with nothing |
|---|---|---|---|
| Salaried, paid monthly | $50,400 | $0 | 0 |
| Base plus one annual bonus | $50,400 | $6,965 | 0 |
| Project-based, four payments a year | $50,400 | $6,607 | 8 |
The same money. Now the documentation:
- Salaried: two pay stubs. Done in a day, read by a system, no judgment required.
- Base plus bonus: stubs plus evidence the bonus recurs, which usually means two years of history and a letter.
- Project-based: twelve months of bank statements, two years of tax returns, often a profit and loss statement, and a human to interpret all of it.
The third household is not riskier in any way the arithmetic shows. It may well be, per our variance analysis, because volatility is a genuine risk factor that most underwriting treats as noise. But that is a separate finding requiring separate evidence, and in practice the decline is frequently not based on measured volatility. It is based on the file not fitting the form.
Per our irregular income analysis, this is the general position of a growing share of workers, and per our verification analysis the verification cost for them is several times higher for identical income.
Where illegibility is charged
| Circumstance | What the system reads | What is actually true |
|---|---|---|
| Several small income sources | Unverifiable income | Diversified income, arguably safer than one employer |
| No credit file | Unknown risk, per our thin file analysis | Possibly years of unrecorded on-time payment |
| Recent arrival in a country | No history | A full history held in a system that does not connect |
| Living with family, no lease | Address cannot be verified | A stable address of long standing |
| Business and household entangled | Unclear which obligations are whose | One economic unit, per our separation analysis |
| Income in cash | Undocumented | Received, spent, and real |
| A gap in employment | Instability | Caregiving, study, or illness, none of which the field captures |
The first row is the clearest case because the direction is arguably backwards. Per our concentration analysis, a business with one customer is understood to be riskier than one with five. A worker with one employer and a worker with five clients are assessed the opposite way, and the reason is not a view about risk, it is that one is easy to document.
The last row is where the field itself causes the problem. A form asking for continuous employment history has no way to record why a gap exists, so the gap is read as the only thing it can be read as. Per our caregiving analysis, the most common reason for such a gap is unpaid work that the system has no category for at all.
Why standardization is still right
The argument for the forms is strong, and treating this as pure institutional laziness gets the problem wrong.
- Consistency requires categories. Per our exception analysis, a system that assesses every case individually allocates outcomes by who presents best rather than by who qualifies, and that is demonstrably worse.
- Structured data is auditable. A decision made on fields can be tested for disparate outcomes. A decision made on a reader's impression cannot.
- Judgment is expensive. Per our manual review analysis, a human reading a file costs real money and the cost has to be recovered, usually from the applicant who needed the reading.
- Unstructured assessment is where bias lives. This is the strongest point: narrative files are judged by people, and the historical record of that is not good.
The fourth point should temper any enthusiasm for more human judgment as the remedy. The answer to illegibility is not to replace forms with impressions. It is to widen what the forms can read, which keeps the decision structured and auditable while admitting more true circumstances into it.
A falsifiable claim
Among applicants with equal measured income and equal measured volatility, those with non-standard income documentation should perform no worse than those with standard documentation. Any lender with a mixed book can test it: hold income level and twelve-month income standard deviation constant, then compare outcomes by documentation type. If non-standard documentation predicts worse performance after controlling for both, then the form is capturing real risk and the decline is justified. If it does not, the lender is pricing the cost of reading the file and calling it credit risk, and the two should be separated on the pricing sheet if they are to be charged for at all.
What the applicant pays
- Document assembly. Statements, returns, letters, reconciliations. Hours of unpaid work, repeated at each institution.
- Time to decision. Per our latency analysis, a file requiring a reader waits in a queue, and the wait is a price.
- A higher decline rate, which per our inquiry analysis means more applications, each recorded, which then makes the file look worse.
- Worse pricing when approved, because uncertainty that could not be resolved is priced as risk.
- Shopping is harder. The applicant who most needs to compare offers is the one for whom each comparison costs the most effort.
- Repetition. Per our portability analysis, the work of making yourself legible is not portable, so it is done again at every institution.
Item six compounds everything above it. A salaried applicant's legibility travels, because a credit file and a pay stub are universally readable. An atypical applicant's carefully assembled explanation does not travel at all, so every application is the first one.
Widening what can be read
The practical programme is to add fields and substitutions rather than to add readers.
- Accept bank transaction data as a first-class input, per our cash flow analysis, which converts an unreadable income shape into a measured one without a human in the loop.
- Measure the shape instead of rejecting it. Twelve-month mean and standard deviation of inflows describes any income pattern, including ones no employment field can express.
- Allow substitutable evidence, specified in advance. Per our sequencing analysis, a prerequisite with a stated second route is allocated by who qualifies rather than by who thinks to ask.
- Add fields for the common atypical cases rather than treating them as exceptions. Multiple income sources, caregiving periods, and recent arrival are not rare.
- Separate "cannot verify" from "verified as weak" in your decline reason codes, per our adverse action analysis, because conflating them makes the problem invisible in your own reporting.
- Count how often the only failure was documentation. That number is the size of this problem in your own book and almost nobody has it.
Item five is the one that makes the rest possible. An institution whose reason codes do not distinguish an applicant it judged risky from one it could not read has no way to discover how much of its decline volume is a documentation problem. Fixing the code taxonomy costs a configuration change and produces the number that justifies everything else on this list.
Scenarios and what we are watching
| Scenario | What we would expect |
|---|---|
| Transaction data becomes standard | The largest reduction in illegibility available, and it also reveals genuinely volatile applicants who previously passed |
| Document reading is automated | Cheaper reading of non-standard files, with the risk that an automated reader's categories are narrower than a human's and harder to appeal |
| Forms keep narrowing for conversion | Shorter applications raise completion rates and exclude anyone whose circumstances need more than the short form allows |
| Nothing changes | The cost of being atypical continues to be charged as credit risk, recorded as credit risk, and therefore never measured as what it is |
The third row is the trajectory we would flag, because it is driven by a metric everyone agrees on. Shortening an application improves completion, and completion is measured precisely. The applicants lost because the short form had no field for their situation are not measured at all, per our selection analysis, so the optimization proceeds on one half of the ledger and every iteration looks like an improvement.
Frequently asked questions
Whether an applicant’s circumstances fit the fields a system can read. It is distinct from risk but produces the same output, so the two get conflated.
On identical $50,400 incomes: two pay stubs for the salaried applicant, against twelve months of statements, two years of returns, and a human reader for a project-based one.
No. Categories are what make decisions consistent and auditable, and narrative assessment is where bias lives. The fix is widening what the forms read, not removing them.
Separate “cannot verify” from “verified as weak” in decline reason codes. It is a configuration change and it produces the number that sizes the problem.
Key takeaways
- Illegibility is distinct from risk: the applicant may be no harder to collect from, only harder to describe.
- Three households on identical $50,400 incomes face assessments ranging from two pay stubs to twelve months of statements plus a human reader.
- A worker with five clients is read as less documentable than one with a single employer, which is the opposite of how business concentration is assessed.
- Employment gaps are read as instability because the field has no way to record caregiving, study, or illness.
- Standardization is right: categories make decisions consistent and auditable, and narrative assessment is where bias historically lives.
- The remedy is widening what the forms can read, not adding human readers.
- Legibility does not travel, so an atypical applicant redoes the work at every institution while a pay stub is universally readable.
- Separating “cannot verify” from “verified as weak” in reason codes is a configuration change that sizes the whole problem.
This analysis is educational and does not constitute financial, legal, or compliance advice. All figures are stylized illustrations and do not describe any particular applicant or lender. Income documentation standards, verification requirements, adverse action reason code taxonomies, and permissible underwriting inputs vary by product, by institution type, and by jurisdiction and change over time.