Small Business Credit Risk: A Practical Guide to Measuring It

Published readings on small business credit risk run from 1.3% delinquency to 4.8% default right now, and every reading is defensible. They describe different borrowers, measured different ways. Choosing which one to benchmark against is a modeling decision rather than a lookup, and most of the confusion in this category comes from treating it as the latter.

Key takeaways

• No single rate exists. Readings run from 1.3% delinquency to 4.8% default right now, and most of that gap is measurement rather than credit. Delinquency counts loans behind on payments, default counts loans that have failed, and each series covers a different set of borrowers.

• Direction matters more than level. Setting those levels against each other settles nothing. Measuring each one against its own history is where the signal is, and on that basis the bank series and the guaranteed series disagree.

• Bank asset size is a flawed proxy for borrower size. Sorting Federal Reserve data by bank size gets closer to small firms without arriving, and a great deal of small business lending sits inside the hundred largest banks.

• Six measures track portfolio health, and one of them sits outside the portfolio. Five answer questions about the loans on the book: how risky the ones still there are, whether recent vintages are failing faster than older ones did at the same age, how the last two quarters of underwriting are performing, whether healthy borrowers can still refinance out, and what a failure costs once it happens. The sixth is the applications a lender turned down, and it is the only one that shows whether credit standards moved or the applicant pool got riskier.

• In standardized programs, lender variance outweighs market averages. Across 36 SBA lenders each holding more than $1 billion under identical program rules, 2025 default rates ran from 0.5% to 12.3%. That spread is close to four times the distance the program average itself moved across the whole cycle. Where a lender sits among its peers says more than how far it sits from a market average.

THE FRAMING · WHY ONE NUMBER CANNOT ANSWER IT

The most widely cited commercial credit figures are dollar-weighted across all commercial banks. Large corporate facilities carry most of the dollars in the banking system, so those series describe large-borrower performance. Small firms are counted in them, but their balances are too small a share of the total to move the average. That is not a flaw. It is what the series were built to measure. Evaluating small business credit risk against those series creates three problems, and they compound.

Measure mismatch: delinquency against default

Delinquency counts loans behind on payments. Default counts loans that have failed. A default rate should sit above a delinquency rate on the same book, so part of the spread between any two published figures is definitional before any economics enter.

Weighting mismatch: balances against loan counts

Bank delinquency series are calculated on loan balances. Portfolio default rates are frequently calculated on loan counts. The two diverge systematically in small business lending, because within any size band the loans that fail tend to be the smaller ones. A dollar-weighted view of portfolio quality will consistently understate how many borrowers were misjudged.

Population mismatch: conventional against guaranteed

This is the largest of the three and the least visible. A bank C&I book is anchored by the companies a conventional lender will serve on its own terms. A government-guaranteed book is not. The SBA 7(a) program carries a credit elsewhere test in statute: a borrower qualifies only if they cannot obtain credit on reasonable terms without the guaranty. That is not a quality judgment, and the distinction matters. The gap is usually about collateral coverage, operating history, or the term the borrower needs.

But it does mean the populations are structurally different, and their default rates should never be expected to converge. In three decades of overlapping data, the 7(a) portfolio has never run below bank C&I delinquency. A higher 7(a) rate is the arithmetic consequence of the population the statute selects, and it is a poor measure of whether the program is doing its job.

There is a competing reading worth stating at full strength, because it is the one policy critics make: elevated guaranteed defaults are evidence of loose underwriting, and the credit elsewhere test is a rationalization for it. The two explanations are separable. If program rules were setting outcomes, lenders operating under those rules would cluster, and the lender-level data below shows they do not, by a wide margin. A rule every lender follows cannot explain a spread that wide. What the statute explains is the floor beneath the spread, which sits higher than a conventional book's floor and always has.

The program works when a viable borrower obtains terms the conventional market would not have offered. Nothing in the structurally higher rate licenses lending to a borrower who cannot carry the debt, and the asymmetry is worth stating plainly: a default is shared between the SBA, which absorbs the guaranteed portion, and the lender, which absorbs the unguaranteed portion, while the borrower carries it whole, for years, through the personal guarantee and often the collateral behind it.

What the mismatches leave intact

Direction survives all three. Any series can be measured against its own history, and a series measured that way is comparable to another series measured that way even when the levels are not.

THE PUBLIC DATA · WHAT EACH SERIES MEASURES

The Federal Reserve delinquency and charge-off family

The quarterly release, Charge-Off and Delinquency Rates on Loans and Leases at Commercial Banks, publishes business loan performance for three bank populations, each as a delinquency rate and a charge-off rate, each in seasonally adjusted and unadjusted form. The naming is systematic once the pattern is visible: DR or COR for the measure, BL for business loans, then the population, then S or N for the adjustment.

Three cautions govern the whole family.

Seasonal adjustment. The S or N suffix is not decoration. Comparing a seasonally adjusted figure against a not-seasonally-adjusted one produces a gap that is partly an artifact of the adjustment. The two conventions can differ by more than a tenth of a point on the same quarter and the same banks, which is enough to matter when an argument turns on the size of a gap. One convention, held across every series in the comparison.

Bank size is not borrower size. The non-top-100 cut is the most common attempt at segmenting toward small business, and it is a genuine improvement on the aggregate, because smaller banks hold proportionally more small business and regional commercial credit. It still does not isolate small firms. A community bank's business loan book carries mid-market and commercial-real-estate-adjacent credit, a substantial amount of real small business lending sits inside the top hundred through card-based lines and SBA origination, and the rate remains balance-weighted, so larger credits dominate within either cohort. It is one rung up the ladder rather than a small business series.

The premium the cut produces is worth tracking as a test of how much work it is doing. Non-top-100 banks have run above the aggregate by roughly half a point on average since 1987, and if bank-size segmentation were capturing small firm distress, that premium should widen as small firms deteriorate relative to larger borrowers. A premium sitting at its long-run average while small-firm-specific indicators move is evidence that the cut is not reaching the population in question.

None of which argues for discarding it. For a community bank asking how its own business loan book compares to institutions of similar size, this is the right reference and there is no public substitute. The bound is on what it can be asked to support: direction on a whole commercial book, not a level for the small business slice of it.

Delinquency is a stock, charge-offs are a flow. Reading the two together is worth the extra minute. Delinquency counts what is past due at a point in time, and charge-offs remove loans from that count, so a bank that resolves problem credits quickly reports lower delinquency on an identical book. Flat delinquency alongside rising charge-offs indicates resolution speed rather than improving health, and the flat number will conceal the move.

The surveys

Senior Loan Officer Opinion Survey (SLOOS). Published quarterly by the Federal Reserve Board, the SLOOS splits its C&I questions by borrower size, separating large and middle-market firms from small firms by annual sales. That split is why it matters more here than most macro indicators: it is one of the few mainstream forward-looking instruments that isolates small firms rather than averaging them away. Three of its readings move independently and should be read separately: whether standards have tightened, whether demand is arriving, and what banks expect of loan quality next. Tight standards with flat demand and deteriorating expected quality describes a different environment from tight standards with demand falling sharply.

Small Business Credit Survey (SBCS). Published annually by the Federal Reserve Banks, the SBCS is survey-based and applicant-level rather than loan-level. Its distinctive contribution is approval and financing-shortfall data: the share of applicants fully approved, partially approved, or denied. Lending data cannot capture demand that went unmet, and this does.

Kansas City Fed Small Business Lending Survey. Quarterly, and bank-reported rather than borrower-reported. It covers small business loan terms, rates, and standards as booked, and serves as a cross-check on the SLOOS with more granularity on pricing and structure.

Subchapter V filings, and the eligibility ceiling that moves

Subchapter V is the streamlined Chapter 11 restructuring path reserved for small businesses, with a debt ceiling capping who may elect it. Because that ceiling excludes larger filers by construction, the series cannot be diluted by large-corporate activity, which makes it structurally immune to the weighting problem and one of the cleanest small-firm solvency reads available. Epiq AACER publishes counts monthly.

The ceiling itself has moved four times, and any comparison crossing one of those moves compares different eligible populations.

Two consequences follow for anyone citing the series. Comparisons reaching back before June 2024 are broken. A filer carrying $5 million in debt could elect Subchapter V in early 2024 and cannot today, so the eligible population on either side of that date is not the same population. And because the ceiling contracted sharply in mid-2024, raw filing counts should have fallen mechanically on that change alone. Increases recorded since then have been recorded against a narrower eligible population, which makes them harder to explain away rather than easier.

Commercial indices and loan-level data

Equifax small business indices. A Small Business Lending Index, a Small Business Delinquency Index, and a Small Business Default Index, published monthly. These are among the closest available things to a public, non-SBA, small-business-specific default series and are worth tracking for direction, though the levels are constructed differently from an internal portfolio measure.

NFIB Small Business Economic Trends. Monthly, and sentiment-based. Adds credit availability and expectations readings that loan performance data cannot supply.

SBA 7(a) and 504 loan-level data. Public at the loan level, with three decades of history. Its value is not that SBA borrowers resemble everyone else, because by statute they do not. Its value is that it is the only large, public, loan-level small business credit performance dataset with full-cycle history, which makes it the only place several of the metrics below can be studied at all. The methods transfer to conventional books even where the levels do not.

Public does not mean ready. The data arrives as periodic extracts rather than as a maintained panel, and every survival-based measure in the next section requires assembling loans into consistent cohorts, tracking status across releases, and reconciling field definitions that have not stayed fixed for thirty years. The distance between the raw extract and a performance history that can be measured against is most of the work.

THE PORTFOLIO METRICS · WHICH QUESTION EACH ANSWERS

Public series describe a market. These describe a book. Six measures answer specific questions: four on how often and how soon loans fail, one on what a failure costs, and one on the applications that never became loans. Dispersion is different in kind. It is not a seventh measure but the set of cuts worth applying to all six, because a portfolio average hides where the answer sits.

Conditional and unconditional default rates

A default rate whose denominator considers the loans still active at the start of each period, rather than the loans originally booked, is a conditional or survival-based rate. It measures the risk of what is currently held. Dividing by the original cohort instead produces the unconditional or cumulative view, which measures how much a pool will lose across its life.

Both are defensible and they can differ enormously on identical loans. On a fully seasoned cohort the unconditional annual rate decays toward zero as the original book pays off or fails, while the conditional rate climbs, because borrowers with options leave and what remains is progressively the weaker half.

The gap is wider than intuition suggests. Measured on the SBA FY2016 cohort, which has now run its full course, the same loans read 0.1% unconditionally and 3.8% conditionally at year ten. Neither figure is wrong. The first says how much of what was originally booked failed that year, and it falls because there is progressively less of the original book left to fail. The second says how risky the remaining loans are, and it rises because the borrowers with somewhere else to go have already gone. A lender watching only the first line would conclude the cohort grew safer every year. The loans still on the books said the opposite.

The two are complements rather than rivals. Lifetime loss estimates are generally built from conditional rates: a period-by-period conditional probability of default, weighted by the probability the loan survived to reach that period, chained into a cumulative figure. The conditional curve is the input and cumulative loss is the output.

That output is what CECL reserving has to report. A lifetime expected credit loss is the survival-weighted probability of default multiplied by severity and by exposure at default, which means the conditional curve is one of three inputs rather than the answer itself. An estimate built on default probability alone, with severity held flat, will misstate the reserve in whichever direction the guaranty and collateral mix happens to run. Reserving lives at the output end. Understanding what is happening to a book right now lives at the input end.

The consequence for benchmarking is direct. Two lenders can post identical unconditional default rates while carrying entirely different risk, if one lender's healthy borrowers refinance away and the other's stay put. Confirming which convention an external benchmark uses is a precondition for comparing against it.

Vintage curves

Portfolio snapshots blend seasoned and new loans and mute the signal. Aligning cohorts at loan age zero and comparing them at the same age removes elapsed time from the comparison, and it is the single highest-value view for separating a problem in the book from a problem in the market. A cohort defaulting faster than its predecessors at the same age is an origination-quality finding. A portfolio average rising is not.

The comparison only works read down a column. In the SBA data, the FY2016 cohort stood at 3.0% at loan age two and the FY2024 cohort at 7.2% at the same age, on the same measure. Neither number means much beside a portfolio average that mixes both.

Early default

A loan that defaults within eighteen months of origination is an early default: the small business analog of early payment default in mortgage, and the fastest feedback loop available on underwriting quality. A vintage curve takes years to fill in, while early default reports on loans written two quarters ago.

The denominator boundary is where most implementations go wrong. Defining the young-loan pool as loans eighteen months old or younger inflates the base with two- and three-month-old loans that have not yet had the opportunity to fail, deflating the rate for purely mechanical reasons. A pool of loans under thirty-six months on book has largely seasoned through the full risk window.

Tracking it by count and by dollars separately is worth the duplication, because the two rank size bands differently. Count reports how many credit decisions went wrong. Dollars report what they cost. The dollar rate typically runs below the count rate, because within any size band the loans that fail tend to be the smaller ones, which means a dollar-weighted read of underwriting quality will understate how many borrowers were misjudged.

Payoff velocity and the competing-risks view

Every loan eventually leaves a portfolio, and most that leave early do so one of two ways. Default is involuntary and carries a loss. Voluntary prepayment does not: a borrower retires a note ahead of maturity, usually by refinancing into conventional debt or selling the business. Both draw from the same shrinking pool of survivors, which makes them competing risks. The rate at which the surviving pool defaults in a period is the conditional default rate (CDR); the rate at which it pays off voluntarily is the conditional prepayment rate (CPR).

Prepayment is the metric most credit teams do not track, and it is the only one of the two that reports on borrowers healthy enough to leave on their own terms. In a book under stress that is precisely the population most easily lost from view. Contained defaults alongside steady prepayment is what a healthy book looks like. Defaults rising while voluntary payoffs fall points to structural liquidity friction, and it also disposes of the most common objection to a rising conditional default rate, which is that healthy borrowers prepaid away and hollowed out the denominator. Were survivorship inflating the rate, the worst cohorts would have prepaid the most.

It is also stable enough to benchmark against. Voluntary prepayment on the SBA 7(a) book held between roughly 11% and 13% a year through both healthy expansions of the last three decades, which makes a departure from that band informative in a way few credit measures are.

Most institutions already hold payoff data in servicing systems without calculating a rate from it. The companion prepayment analysis sets out the construction.

Loss given default

A default rate reports how often loans fail. It says nothing about what a failure costs. Loss given default is the share of exposure not recovered once a loan has failed, net of recovery costs and of the time the recovery took, and it is the difference between a frequency measure and a loss estimate.

That distinction carries much of the weight in benchmarking. Two books with identical default rates can produce losses that differ by a multiple, because severity depends on what stands behind the loan. A government guaranty is the largest single driver in small business lending: on a 7(a) loan the SBA absorbs the guaranteed portion, so the lender's severity is calculated on the unguaranteed exposure alone. Collateral is the next. A loan secured by commercial real estate and a working capital line resting on a blanket lien over receivables recover on entirely different curves. Lien position, personal guarantees, and whatever going-concern value survives the default move it further.

Two measurement cautions apply, and both mirror problems already described above.

Recovery arrives over years rather than at the moment of default, so an LGD calculated without discounting for timing understates severity. A dollar recovered in year four is not a dollar recovered at charge-off.

And an LGD calculated only on resolved workouts is biased in the same way an unconditional default rate is. Files that resolve quickly are not representative of files still open, and the slow ones are generally the worse ones. Severity measured only on closed files will read better than the book deserves.

Dispersion

Portfolio averages conceal the actionable finding. Six cuts are worth maintaining permanently, and they differ in when a lender can act on what they show.

Three of the six behave roughly as expected. Sector dispersion routinely exceeds an institution's entire pricing grid. Geography repays county-level rather than state-level treatment where volume supports it. Vintage is the age-aligned comparison described above.

Loan size should be tested rather than assumed. In 2016 the relationship in the SBA book was a clean ladder: default rates fell almost perfectly in step with loan size, from 2.5% on loans under $150,000 down to 1.1% on loans above $3 million. That ladder has since broken. The curve is now U-shaped, with the largest credits having deteriorated fastest in relative terms and the safest band sitting in the middle of the distribution.

Conditional (survival-based) annual default rate by loan count, SBA 7(a). The $3M+ bucket moved from safest in 2016 to third-worst in 2025, quadrupling while every other band roughly doubled.

The mechanism bounds where that shape should be expected to appear. At the top of the size distribution the loans are largely changes of ownership, partner buyouts, and substantial commercial real estate, they carry heavy goodwill or fixed-asset weights with little slack, and in the 7(a) program they are overwhelmingly floating against prime. Debt service on several million dollars does not travel gracefully from 6% to 10.5%. A book of fixed-rate credits of similar size did not take that shock and should not show the same curve. Two other things partly account for it: loans above $3 million were only permitted from 2010, so the band has a young age distribution, and larger credits fail late rather than early, which means the shape is still filling in.

The point is not that every book is U-shaped. It is that a size policy inherited from the last decade assumes a monotonic relationship, and whether that assumption still holds is an empirical question with a knowable answer on any given book.

Rate structure is the cut most often missing from a segmentation scheme built before 2022. Fixed against floating, the index and the spread over it, and the reset schedule together determine how much of a policy rate move lands on a given borrower's debt service without anything happening to that borrower's revenue. In books dominated by floating structures tied to prime, that exposure is a risk dimension in its own right, and it is knowable at origination. Segmenting default outcomes by rate structure, holding vintage and sector constant, separates borrowers whose coverage ratios were rewritten by the rate cycle from borrowers whose businesses deteriorated.

Peer position is often more useful than any market index. Where lender-level data exists, and it exists most completely in government-guaranteed lending, the best and worst operator working under identical rules in one economy sit further apart than that program's average moved across the whole cycle. Both figures are the same measure on the same borrowers, which is what makes them worth setting against each other. Conditions move the average, and the spread around it is what conditions do not explain.

Conventional lender-level performance is rarely public, so the same measurement is harder to run outside a standardized program. The mechanism carries over regardless. Policy cutoffs, overlay choices, sector concentration, and servicing execution generate dispersion that a market average compresses, and a lender's place in that spread is more useful than its distance from the average.

The declined population

Everything above measures loans that were booked, which is half the record. The other half sits in the decline file, and it is the half that answers whether a book's deterioration reflects a moving bar or a worsening pool.

The two are indistinguishable from outcomes alone. A lender whose standards held while its applicant pool got riskier and a lender who lowered its bar against a stable pool will both report falling approval rates and rising defaults. They call for opposite responses.

Separating them requires scoring approved and declined applications together with one model held constant across periods, then measuring the approval rate within fixed risk bands rather than across the pool as a whole. Bands defined relative to each period, such as deciles, absorb the compositional shift they are meant to detect. If approval rates within bands held steady while the pool's composition moved toward the riskier bands, the bar did not move. If approval rates rose inside the riskier bands, it did.

The same exercise supports a single summary figure: the predicted default rate at which approval probability crosses one half, tracked by origination quarter. Drift upward in that implied cutoff is the clearest available evidence that a credit policy moved, and it is legible to a credit committee in a way a decomposition is not.

The common substitute does not answer the question. Tracking the distribution of predicted risk or bureau scores across approved files, whether through a population stability index or a simpler drift report, measures the output of a bar applied to a pool. A shift in that distribution is consistent with a moving bar, a worsening pool, or both, and nothing in the approved book distinguishes them. Stability indices are worth running for what they do measure, which is whether a model is seeing the population it was fitted on. They cannot attribute a change to policy or to demand.

Retaining decline records is the precondition, and it is where most institutions lose the ability to run this at all. Where the records do not exist, partial identification is still available from the approved book alone: a shrinking share of the lowest-risk approvals is more consistent with a deteriorating applicant pool than with a deliberately lowered bar, since a lender that loosened would add files at the riskier end without losing them at the safer end. That is weaker evidence and should be labeled as such. It is also a reason to begin retaining decline files now, because the test needs several years of history before it can say anything.

THE BENCHMARK · MATCHING A REFERENCE TO A BOOK

Four questions, in order.

Is the book guaranteed or conventional? Guaranteed portfolios select for borrowers the conventional market declined and should be expected to run higher permanently. Neither should be benchmarked against the other without stating the offset.

What loan sizes does it hold? A book concentrated under $350,000 has little in common with one concentrated above $1 million, in either default level or default timing. Larger credits tend to fail late, as rate resets work through coverage ratios, while smaller credits fail early. A benchmark that does not match the size distribution misleads about both magnitude and timing.

Delinquency or default, counts or dollars, adjusted or unadjusted? Establishing this before comparing accounts for a large share of apparent contradictions between sources.

And which question is the comparison meant to answer? Whether a market is deteriorating and whether a lender is underperforming that market require different references. The first wants direction in a population that resembles the book. The second wants a distribution of peer outcomes.

One caution governs all four rows. Default is a reporting classification rather than an observed fact. Whether a struggling borrower is recorded as defaulted depends on where an institution sets its delinquency threshold, how quickly it charges off, and how readily it modifies or restructures. Two identical borrowers at two lenders can produce different recorded outcomes, so confirming the definition precedes concluding anything about relative credit quality.

CALIBRATION · WHEN A MODEL STOPS DESCRIBING THE BORROWERS

A credit model encodes the borrower population it was fitted on. When that population shifts, the model does not report the shift. It produces confident scores against assumptions that have quietly expired.

The clearest available test is year-one default by cohort. Where recent vintages fail in their first full year at materially higher rates than the cohorts a model was trained on, the model's implied year-one assumption is stale and it will underprice current files without signalling that it is doing so. The SBA data shows how far that assumption can drift: the FY2016 cohort defaulted at 1.6% in its first full year and the FY2024 cohort at 3.4%. A model fitted on the earlier population encodes the lower figure, and encodes it silently. Loans failing that early were generally mispriced or misjudged at origination rather than overtaken by events, which makes year-one performance a direct read on whether the underwriting screen still works.

Two further questions are measurement questions rather than procurement ones, and they change how any model's output should be read. What population was it fitted on, small business loan performance or consumer credit behavior used as a proxy? And what horizon does it predict? A twelve-month PD and a lifetime PD score the same unseasoned loan differently, and the difference reads as model error when the horizon is not understood. The wider evaluation criteria, including training-cycle coverage, validation design, recalibration cadence, and the regulatory condition introduced by the SBA's March 2026 sunset of the FICO SBSS mandate, are treated in the build-versus-buy guide.

A second test runs continuously rather than by cohort, and it catches a failure the first one misses. Comparing realized default against predicted default on the approved book, by origination period, produces a calibration ratio. Where realized default rises roughly in step with predicted, the model saw what happened and its view of the population still holds. Where realized rises faster than predicted, something moved outside what the model observes. That residual is worth chasing, because the things a PD built on financial and bureau inputs cannot see are exactly the things that loosen quietly: thinner guarantor support, more generous add-backs in a cash flow reconstruction, weaker collateral positions, amortization stretched to manufacture coverage. A credit policy can hold a scoring cutoff perfectly constant while every one of those moves.

One guardrail on scope, because the point is routinely overstated in both directions. No model calculates debt service coverage, and a calibration test reports on the screen rather than on the underwriting behind a file. The case for a small business score on a commercial file is treated separately.

THE CURRENT READING · JULY 2026

This section is dated and refreshed quarterly, along with the current readings in the opening summary and the key takeaways. Everything else holds: the method does not move with conditions, and the figures illustrating it are drawn from cohorts and periods that have closed. At Lumos we do not forecast the macro economy, and nothing here should be read as a call on one.

Four readings, four populations, two measures.

The levels are not comparable. Measured against their own histories, the divergence is the finding: the bank series sit in the lower third of their own ranges while the 7(a) portfolio sits near the top of its own.

That comparison needs one correction applied evenhandedly, because the pandemic floor was artificial in the bank data as well as in the guaranteed data. DRBLOBS bottomed at 0.9% in the first quarter of 2021, the lowest reading in the thirty-nine years the series has existed, for the same reasons 7(a) defaults bottomed then: liquidity programs, forbearance, and direct payment relief. It has since risen by more than 90%, to 1.8%. Any claim that small-bank business delinquency has been flat holds only for the last four quarters. The fuller statement is that it normalized sharply off an artificial floor, stalled in early 2025, and has drifted slightly lower since.

The distinction that matters is where each series stopped. Small-bank delinquency normalized back toward its historical range and halted well inside it, below even the 1.9% it touched in the first quarter of 2020 and at 47% of its 2009 peak. The 7(a) portfolio normalized past its own historical range. Two populations recovering from the same suppressed floor, and only one of them kept going.

Reading the two bank series against each other adds a third finding, and it cuts against the most common way of using them. The premium that non-top-100 banks carry over the aggregate stands at roughly half a point, which is where it has sat on average since 1987. That premium fell to near zero across 2020 and 2021, when relief programs compressed every credit measure toward zero, and it has since reverted to its long-run level and stopped there. It is not widening. Were bank-size segmentation capturing small firm distress, the premium should widen as small firms deteriorate relative to large ones. It has not, which is the measured form of the caution in the second section. The cut sorts banks, the distress sits in borrowers, and at present the two do not line up.

The indicators built to see small firms point the same way, and the previous default analysis works through them in full. In brief: Subchapter V filings rose 50% year over year in the first half of 2026, recorded against an eligible population that narrowed in mid-2024; the January 2026 SLOOS found banks expecting quality to deteriorate for small-firm C&I loans specifically while holding steady for larger firms; and voluntary prepayment has fallen to roughly 8.4% against the 11% to 13% held through prior expansions, an inversion with one prior instance in thirty years of this data. The resemblance to 2008 is in the pattern, not in the cause, and nothing here argues the two periods share one.

The conventional evidence is what establishes that none of this is an artifact of the guaranteed population. Across 88,917 conventional applications at five lenders, held without a government guaranty, average predicted probability of default on new originations rose from below 0.5% for the 2019 through 2021 vintages to above 1.5% for 2024 and 2025. Realized defaults on those recent cohorts read lower, which is a seasoning effect rather than a contradiction: the loans are younger, not safer.

Risk concentrates in the same places in both books: in the 2022 through 2024 vintages, in a handful of sectors, at both ends of the size distribution, and most of all in particular lenders. Across 36 guaranteed lenders each holding more than $1 billion, 2025 default rates ran from 0.5% to 12.3%. The program average moved 3.2 points across the whole cycle, from its 2021 trough to March 2026. The gap between the best and worst lender in a single year is nearly four times that.

THE TAKEAWAY

The question of how small business credit is performing is badly posed whenever it is asked at the aggregate level. Four defensible readings currently sit between 1.3% and 4.8%, each correct about a different borrower, and the first analytical act is deciding which of them describes the book in hand.

What travels across all of them is method. Measure each series against its own history rather than against each other. Read delinquency next to charge-offs, conditional rates next to cumulative ones, and default next to prepayment, because in each pair the second number tells the first what it means. Align vintages at loan age before concluding anything about origination quality. And treat peer outcomes as the benchmark that matters, because inside one population the best and worst operator sit further apart than the average moved across the entire cycle.

The lenders running at the bottom of that spread are not luckier. They are reading a different set of numbers.

FAQ

How is small business credit risk measured?

At the market level, by selecting a data series whose borrower population resembles the book in question and reading its direction against its own history rather than comparing levels across populations. At the portfolio level, with six measures answering six questions. Four cover how often and how soon loans fail: a conditional default rate for the risk of what is currently held, vintage curves aligned at loan age for whether origination quality has shifted, early default for underwriting quality on loans written in the last two quarters, and payoff velocity for whether healthy borrowers can still exit. A fifth, loss given default, covers what a failure costs once it happens, since two books with identical default rates can lose different amounts. The sixth sits outside the book entirely, in the applications a lender declined, and it is the only one that separates a change in credit standards from a change in the applicant pool. Dispersion cuts by sector, loan size, geography, vintage, rate structure, and peer position belong alongside all six, because portfolio averages conceal the findings that can be acted on.

What is the small business loan default rate?

There is no single figure, and treating one as authoritative is the most common benchmarking error in this category. Published readings differ by measure, delinquency against default; by weighting, balances against loan counts; and above all by population, guaranteed against conventional and large borrowers against small ones. As of mid-2026 defensible readings ran from 1.3% for business loan delinquency at all commercial banks to 4.8% for the trailing twelve-month default rate on the SBA 7(a) portfolio, with conventional small business performance between them. Those figures describe different borrowers and should not be compared as levels.

What is the difference between a delinquency rate and a default rate?

Delinquency counts loans behind on payments at a point in time. Default counts loans that have failed, on a definition the reporting institution sets. A default rate should sit above a delinquency rate on the same book. Delinquency is also a stock that gets cleaned by charge-offs, so an institution that resolves problem credits quickly reports lower delinquency on an identical book, which is why delinquency and charge-off series are best read together.

Why does loss given default matter when comparing small business portfolios?

Because a default rate reports how often loans fail and says nothing about what a failure costs. Loss given default is the share of exposure not recovered after a loan fails, net of recovery costs and of the time recovery took, and two books with identical default rates can produce losses that differ by a multiple. In small business lending the largest driver is whether a government guaranty stands behind the loan, since on a guaranteed loan the lender's severity is calculated on the unguaranteed exposure alone. Collateral type, lien position, and personal guarantees move it further. This also matters for reserving: a lifetime expected credit loss under CECL is the survival-weighted probability of default multiplied by severity and by exposure at default, so an estimate built on default frequency alone will misstate the answer.

Does segmenting Federal Reserve data by bank size isolate small business lending?

Only partially. The non-top-100 cut, DRBLOBS seasonally adjusted or DRBLOBN not, improves on the all-commercial-bank aggregate because smaller banks hold proportionally more small business and regional commercial credit. It still sorts by bank asset size rather than borrower size: community bank business books include mid-market and commercial-real-estate-adjacent credit, substantial small business lending occurs inside the top hundred, and the rate remains balance-weighted so larger credits dominate within either cohort. It is one rung up the ladder rather than a small business series. Comparing a seasonally adjusted figure from one series against a not-seasonally-adjusted figure from another adds a further artifact and should be avoided.

Can a conventional small business book be benchmarked against SBA data?

Not on levels. Government-guaranteed programs select for borrowers who could not obtain credit on reasonable terms without the guaranty, so their default rates should be expected to run structurally higher and permanently so. What transfers is method and direction. Vintage alignment, conditional rate construction, early default definitions, and competing-risks analysis all apply to conventional books, and SBA loan-level data is the largest public dataset with full-cycle history on which those methods can be studied.

Which indicators give the earliest warning of small business credit stress?

Four, none of which require waiting on a delinquency report. Early default reports on underwriting quality within roughly two quarters. Payoff velocity indicates whether healthy borrowers still have exit options, and falling payoffs alongside rising defaults point to structural liquidity friction rather than a book turning over normally. The expected-loan-quality question in the SLOOS isolates small firms specifically and is forward-looking. And Subchapter V filings are published monthly and cannot be diluted by large-corporate activity, subject to the eligibility ceiling caveat above.

What is a conditional default rate?

A conditional or survival-based rate uses a denominator of loans still active at the start of each period rather than the original cohort, so it measures the risk of the surviving portfolio independent of how quickly healthy loans pay off. The previous default analysis treats the question in full, including how conditional rates chain into cumulative loss.

What is an early default rate?

A loan defaulting within eighteen months of origination, measured against a denominator of loans under thirty-six months on book so the pool has seasoned through the full risk window. The construction, and the reason the denominator boundary matters, are set out in the previous default analysis.

Does a credit score built on consumer data work on a small business file?

The vintage evidence argues against relying on one alone, and since March 2026 a lender's own model is permitted subject to its primary federal regulator, provided it does not rely solely on consumer credit scores. The small business credit score analysis covers what that leaves.

About this analysis

Method and metric definitions here draw on loan-level data in the Lumos Data Portal, which holds more than 2 million SBA loans and 30 years of performance history, and on validation of Lumos Prime+ across 88,917 conventional small business applications at five lenders, on loans held without a government guaranty.

The same analytics can be pointed at a single institution's book rather than at the market. Portfolio Insights applies probability of default, loss given default, and expected loss at the loan level across small business portfolios, SBA and conventional alike, with risk migration and segmentation by vintage, industry, and geography. Prime+ scores applications at origination. The Business Report returns a qualification grade with sector and geographic risk context before an analyst opens a file.

Lumos will retro-score a lender's own historical originations and declined applications, which produces the vintage, early-default, and dispersion views described above on that institution's portfolio rather than on market data.

Notes & sources

Measurement. Conditional default rates are survival-based by loan count, with each period's denominator considering the loans active in that period. Conventional figures come from a lifetime-horizon PD model; the loans it learned from carry default indications set at 60, 90, or 120 days past due depending on the recording institution. Realized conventional default rates on 2023 and later cohorts are depressed by incomplete seasoning. Federal Reserve figures in the current reading are seasonally adjusted throughout.

Federal Reserve. Charge-Off and Delinquency Rates on Loans and Leases at Commercial Banks: DRBLACBS, DRBLOBS, DRBLT100S, and charge-off companions. Senior Loan Officer Opinion Survey, January 2026. 2026 Report on Employer Firms. Federal Reserve Bank of Kansas City, Small Business Lending Survey.

Bankruptcy. Epiq AACER and the American Bankruptcy Institute, small business filings, first half 2026. Eligibility thresholds under the Small Business Reorganization Act, the CARES Act, and 11 U.S.C. § 104.

Lumos. SBA 7(a) loan-level data from the Lumos Data Portal; conventional figures from the five-lender Prime+ validation. Analysis by Lumos.

Companion analyses. SBA 7(a) default rates and vintage curves · prepayment speeds and the competing-risks view · conventional default rates by vintage · FY2026 program performance

Last updated July 2026. The current reading section is refreshed quarterly, together with the figures in the opening summary and key takeaways.