Underwriting a Small Business Is Not Underwriting a Consumer

Small business lending borrows heavily from consumer methods, and at the smallest end that's correct rather than lazy — the borrower genuinely is a household with a trading name. The problem is the middle. A business too large to assess purely on its owner and too small to have the financial statements a commercial approach assumes is where most small business lending actually sits, and it's where most models perform worst. What separates the two is not size but four specific properties, and each of them breaks a method that works fine on consumers.

The spectrum, and the middle

Scale	Financially	Right method
Sole operator	A household with a trading name	Consumer methods, largely
A few employees	Partly separated	The difficult middle
Established small business	Mostly separate	Cash flow plus owner
Lower middle market	Separate	Financial statements

The second row is where the volume is and where methods fail. These businesses have real trading activity that isn't reducible to the owner's finances, and they don't have audited statements, a meaningful commercial file, or a track record long enough to model.

What tends to go wrong there:

• Consumer methods underweight the business entirely, missing concentration and trading deterioration.
• Commercial methods demand inputs that don't exist, so the applicant is declined for want of documentation rather than for risk.
• A blended model trained across the whole spectrum is fitted mostly to the ends and calibrated badly in the middle.

Which is the segmentation argument from our segmentation analysis: where a relationship genuinely differs by segment, a single model with a size variable won't capture it, and the test is whether segment-specific models outperform out-of-sample.

The owner-business boundary

The property with no consumer analogue.

Below a certain size the household and the business are one financial unit, and money moves both ways continuously — the owner funds the business when it's short and draws from it when they are. Per our structural analysis, a personal guarantee removes the containment that made the entity form meaningful, so most small business obligations are household obligations with extra paperwork.

What follows for assessment:

• The owner's personal file is a legitimate and frequently dominant input, and per our coverage analysis it's usually far richer than the commercial one.
• The owner's personal capacity matters — a business whose owner has substantial personal obligations has less resilience than its own figures show.
• Transfers between the two distort both pictures. Per our statements analysis, owner contributions inflate apparent revenue and draws understate available cash.
• A guarantee changes what you're underwriting, and it should change the assessment rather than sit as a document at the end.

The most useful single question: if the business stopped tomorrow, what would the owner's position be? That determines recovery on a guaranteed obligation and it isn't answered by any business metric.

And the boundary is diagnostic in itself. A business with no separation — mixed accounts, no books, no distinction between owner and entity — is a smaller effective credit than its revenue suggests, because there's no entity there to assess.

Concentration

The single largest difference in risk shape, and the one consumer methods have no equivalent for.

Per our concentration analysis, a small business may depend on a handful of customers, any of whom can leave without notice. That produces a far fatter tail of sudden severe revenue loss than consumer portfolios exhibit — and it's the most common proximate cause of a business failing to make payments it previously afforded comfortably.

How to assess it:

• Measure it directly from deposit data — what share of receipts comes from the largest source?
• Check the number of distinct payers, which is computable from bank data without asking.
• Check for a recent change in the mix, since a large customer's departure shows in the pattern before it shows in the total.
• Consider contract terms where they exist — a customer on notice is different from one who can stop instantly.
• Consider the counterparty, since concentration with a strong payer is a different risk from concentration with a weak one.

The second is underused and free. Counting distinct payers in the deposit stream gives a concentration measure with no additional data request — and per our affordability analysis, anything computable from data you already hold costs the applicant nothing.

The modelling implication: concentration is a variance property rather than a level property, so a model predicting expected loss from average characteristics will systematically understate it. It affects the tail, and the tail is where a small portfolio's results actually come from.

Seasonality

A property most consumer portfolios can ignore and commercial ones cannot.

Over a short window, a seasonal trough is indistinguishable from a decline. Which produces two systematic errors:

• A business assessed in its low season looks like it's deteriorating.
• One assessed at its peak looks stronger than it is, and the payment set against peak revenue arrives in the trough.

Both are systematic rather than random, which means they don't average out and they bias the portfolio depending on when applications arrive.

How to handle it:

1. Use a full annual cycle where the data allows.
2. Compare to the same period in prior years where it doesn't.
3. Identify seasonality by industry, which is frequently predictable from the segment alone.
4. Assess affordability against the trough, per our statements analysis — the payment has to survive the worst month.
5. Structure around it where the product allows, since a seasonal business may be a good credit with a seasonal payment schedule and a poor one with a flat schedule.

The fifth is the underused response. Treating seasonality as a risk to price rather than a pattern to structure around declines businesses that would perform well on a schedule fitted to their cash cycle.

The data problem

What's available differs fundamentally from the consumer case.

Input	Consumer	Small business
Credit file	Rich, standardized	Thin, inconsistent coverage
Income	Verifiable from source	Revenue, needing interpretation
Obligations	Mostly on file	Frequently not on any file
Financials	Not applicable	Frequently unavailable or unreliable
Bank data	Useful	Usually the best input available

Row three is the trap. Per our coverage analysis, a great deal of small business borrowing doesn't appear on a commercial file at all — so an assessment relying on the file will miss obligations that bank data reveals immediately, including the daily-debit advances our advance analysis describes.

Which makes bank data the primary input rather than a supplement, per our cash flow analysis. It shows revenue, obligations, concentration, seasonality, and the balance pattern in one source.

Its limits, stated honestly: it shows cash rather than profitability, it's distorted by owner transfers, it can be split across accounts you don't see, and a short history makes seasonal patterns unreadable. The classification rules that separate revenue from transfers are model inputs needing the same governance as any other, per our attribute analysis.

Segmentation that matters

Commercial portfolios differentiate on dimensions consumer portfolios barely use.

• Industry, which drives failure rates, seasonality, margins, and asset intensity — and is frequently the strongest single segmentation variable available.
• Age of business, where early years carry materially different risk.
• Size, per the spectrum above.
• Business model — recurring versus project-based revenue behave differently under stress.
• Asset intensity, which determines what's recoverable.

Per our segmentation analysis, test whether these are segments or variables: include them with interactions first, and split only where segment-specific models beat the combined one out-of-sample. Industry frequently earns a split; the others frequently don't.

And a boundary warning: industry classification is itself unreliable — self-reported codes are frequently wrong or stale, and a business's actual activity may not match its registered category. A segmentation resting on a misclassified input produces confident errors, so verify the classification before building on it.

Outcomes look different

What happens after default differs enough to matter for the economics.

• The business can cease to exist, which has no consumer equivalent — there's no entity left to pursue.
• The guarantee becomes the recovery route, which means recovery depends on the owner's personal position rather than the business's.
• Collateral may be recoverable, depending on asset intensity.
• Deterioration can be faster. A business losing its main customer can go from current to failed within a quarter, where consumer deterioration is usually slower.
• Cure rates differ from consumer patterns, so per our definition analysis, applying a consumer outcome definition to a commercial portfolio mismeasures both the timing and the severity.

The fourth has a direct implication for monitoring. Per our early warning analysis, a portfolio that can deteriorate within a quarter needs monitoring at that frequency — and bank data makes it possible, since revenue and concentration changes appear in the deposit stream weeks before a payment is missed.

Fair lending considerations

An area that has developed and needs current advice rather than assumption.

Points to raise with counsel and compliance:

• Which requirements apply to your business lending, which depends on product, size, and jurisdiction, and has changed.
• Data collection and reporting obligations for small business applications.
• Adverse action requirements, per our notices guide, where the applicable standards may differ from consumer lending.
• Whether using the owner's personal characteristics raises questions your consumer program already addresses.
• Whether industry or geographic segmentation creates proxy concerns.

The fifth deserves specific attention because industry is the strongest segmentation variable here and geography correlates with protected characteristics in ways this desk's testing analysis describes. A segmentation that improves prediction and correlates with geography needs the same scrutiny as any other variable, and the commercial context doesn't remove the question.

Test for disparities by segment rather than only in aggregate, and document the business justification for each segmentation decision at the point you make it.

Frequently asked questions

---

This guide is educational and does not constitute legal or compliance advice. Fair lending requirements applicable to business credit, small business data collection and reporting obligations, and adverse action standards vary by product, institution, and jurisdiction and have changed in recent years. Consult qualified counsel and your compliance function about the requirements applicable to your lending, and your model risk function about validating any segmentation.