Conventional Small Business Loan Default Rates by Vintage: Results from a Five-Lender Lumos Prime+ Validation

Between 2019 and 2025, five conventional business lenders originated the loans that make up this analysis, and over that period the default risk Lumos Prime+ predicts on their new originations more than tripled. The same model, run across the same five portfolios, also placed roughly half of all historical defaults inside the riskiest fifth of applications.

This is the full set of results behind those two statements, with every origination year reported separately rather than rolled into an average. The evaluation covered 88,917 applications, and the result that carries furthest is where the model placed the loans nobody booked.

What Was Tested

A conventional business loan, for the purposes of this analysis, is a small business loan a lender holds on its own balance sheet without a government guarantee, which distinguishes it from the SBA 7(a) and 504 lending that most published small business credit research describes.

Lumos Prime+ predicts three measures over the full life of a loan: the probability that it defaults, the share of exposure not recovered if it does, and the expected loss those two imply. This validation tested the first of the three. Probability of default, or PD, is the likelihood a loan fails to perform over a stated horizon, and the horizon here is the loan's lifetime rather than the next twelve months. What counts as failing to perform is less uniform than the phrase suggests. The loans Lumos Prime+ learned from carry default indications set at 60, 90, or in some cases 120 days past due, depending on the institution that recorded them. The model has never been fitted to a single convention. The other two, loss given default (LGD) and expected loss (EL), were not evaluated in this study, and no claim about their accuracy should be read into what follows.

Five conventional business lenders contributed loan-level data. Each provided at least 500 loans containing at least 10 defaults, with four of the five providing more than 40 defaults, and each provided at least 15 declined applications, with four of the five providing more than 100. The pooled sample came to 88,917 applications: 54,538 that the lenders declined and 34,372 that they approved. Within the approved loans, 792 had defaulted at the time of analysis, a default rate near 2.3% that indicates how imbalanced this prediction problem is. On that approved book, a model that predicted no defaults at all would be right about 97.7% of the time and useless.

Every application was scored by Lumos Prime+ and sorted into four groups: applications the lender declined, approved loans that later defaulted, approved loans still performing after less than three years, and approved loans still performing after three years or more. Seasoning, the age of a loan relative to the window in which defaults typically occur, is the reason for that last split. The three-year line is where these portfolios begin to show most of their eventual failures.

Where the Model Placed Each Group

Across the pooled sample, the loans already known to have defaulted drew the highest median predicted PD of the four groups. Loans still performing after three years or more drew the lowest. That ordering is the basic result, and it holds in the direction a working model requires.

The two middle groups are the interesting ones. Unseasoned performing loans, which had not yet had time to fail, were scored closer to the known defaults than to the seasoned performers. On a twelve-month model that pattern would look like error. On a lifetime model it is the point, because a loan that is eighteen months old and current has not survived its risk period, it has not reached it yet. A shorter-horizon model would score those same loans as safe and be wrong about them later.

Declined applications behaved differently depending on how the data is cut, and the difference is worth stating precisely rather than smoothing over. Within each individual lender's portfolio, declined applications generally drew the highest predicted PDs of any of the four groups. Pooled across all five lenders, the defaulted loans drew the highest. Both statements are accurate. The most plausible reconciliation is that the five lenders operate at widely different average risk levels, so pooling their applications mixes populations whose scores are not directly comparable, and the per-lender view is the more meaningful one for a lender assessing its own book.

That per-lender result is the finding that speaks to underwriting. The model was never shown the outcome of any credit decision, and it independently assigned its highest risk scores to the applications each lender's underwriters had already turned down. Lumos Prime+ agreed with those underwriters using none of the judgment, relationship history, or narrative context they had. Reaching an underwriter's conclusion from loan and borrower data alone is what makes credit judgment repeatable across a portfolio, across underwriters, and across time.

Default Concentration in the Seasoned Portfolio

Ranking matters more than the absolute scores, because a lender acts on order rather than on levels. To test ordering, the 12,823 seasoned loans in the sample were sorted by predicted PD and divided into ten equal groups.

Defaults were not spread evenly across those ten groups. The riskiest decile by Lumos Prime+ ranking contained 35% of every default in the seasoned portfolio, and the second riskiest contained a further 13%. Performing loans, by construction, sat at roughly 10% per decile. The share of defaults rose steadily across the ranking rather than jumping only at the end, which is what separates a model that genuinely orders risk from one that merely flags a handful of obvious failures.

The practical version of that concentration is a trade between volume and losses. Declining the riskiest 20% of applications, following the Lumos Prime+ ranking, would have avoided about 50% of historical defaults in this seasoned data set.

Three boundaries belong on that figure. It is retrospective, calculated across loans whose outcomes are already known, and a prospective decline rule would not perform identically. It is drawn from pre-2023 originations, which are the only ones seasoned enough to support the calculation. And the 20% reduction in approvals is a gross figure that does not net out the previously declined applications Prime+ scored as safe, which a lender could have approved in their place. Whether that trade is worth making depends on a lender's pricing, cost of funds, and growth targets, which is a judgment no model makes.

The declined applications run the other way, and they are the half of this trade that gets overlooked. Lumos Prime+ scored 54,538 of them. They generally drew the highest predicted PDs within each lender's book, but they did not draw uniformly high ones: in every portfolio, part of the declined population scored below that lender's own approved median. A ranking that identifies the riskiest approvals identifies the safest declines in the same pass.

That points at a different use of the model than trimming a book. A lender working from the ranking can decline its riskiest approvals and review its lowest-scoring declines together, which changes what it books rather than simply how much. Whether total volume then holds flat, falls, or rises depends on how many declines clear that lender's threshold, and the pool is not small: these five lenders declined 54,538 applications against the 34,372 they approved. The 20% figure above measures only the first move.

One boundary is firm and worth stating plainly. A declined application was never booked, so it has no outcome, and nothing in this data establishes that a low-scoring decline would have performed.

The agreement result is the reason to look anyway. Across the declined population as a whole, Lumos Prime+ reached the same conclusion the underwriters reached, which indicates that most of what underwriting acts on is present in the data the model sees. A decline the model scores as low risk is an exception to that pattern, and an exception is worth examining when the pattern is otherwise strong. The competing reading is that the exception is the underwriter seeing something the file does not carry, which is exactly the case that would not show up in a backtest of approved loans. Both readings recommend the same next step, which is to look at those files rather than to assume in either direction.

What the ranking produces is a list worth re-examining, not a conclusion about loans that were never made. Settling it means approving some of them and waiting.

The broader argument about volume and loss moving independently is developed separately in The Growth vs. Risk Myth.

Defaults do not distribute themselves evenly, and a portfolio's worst tenth is not a tenth of its problem.

Rank Ordering by Origination Year

The standard measure of ranking quality is the area under the receiver operating characteristic curve, or AUC, which describes how reliably a model ranks a defaulted loan above a performing one. An AUC of 0.5 is a coin flip and 1.0 is perfect separation.

Reported by origination year, the results were uneven, and the pattern in that unevenness carries more information than the average does.

For loans originated before 2023, Lumos Prime+ reached an AUC of approximately 0.70, and above 0.70 in every year of that cohort except 2019. Whether 0.70 is a strong figure is better answered in operational terms than against a benchmark, and the decile ranking above is the answer: that figure is what put 35% of all defaults into the riskiest tenth of the book, three and a half times its share. It was earned under conditions that hold the number down rather than flatter it: out-of-time vintages, a book only partly seasoned, an event rate near 2.3%, and a scored population the lenders' own underwriting had already filtered. That last constraint matters most. AUC can only be computed where outcomes are known, so it was measured on approved loans, a population whose highest-risk applicants were removed before origination. Ranking within a book that has already been screened is harder than ranking across everything that walked in the door. Results produced without those constraints are not comparable to this one.

The 2019 exception, at 0.626, came from a concentration in industries that pandemic conditions hit disproportionately early in those loans' lives. Those industries defaulted at rates well above the 2019 cohort's other sectors, and nothing in a 2019 loan file marked which borrowers would be exposed to a shock that had not yet happened.

Two forces then pulled on that cohort in opposite directions. The shock produced defaults among borrowers that had looked sound at origination. Federal relief and widespread loan modification, including the Paycheck Protection Program, suppressed defaults among borrowers the model had scored as weak. That effect was large enough to register nationally, where the surge in business bankruptcies that the 2020 contraction would ordinarily have produced did not materialize. Rank ordering is measured against recorded outcomes, so both dynamics loosened the link between borrower quality and observed performance, which is the link AUC exists to measure. The pattern points to distortion in the recorded outcomes rather than a breakdown in ranking.

It is also confined to one vintage. The same model, applied to all seven without cohort-specific tuning, cleared 0.70 on three of the four seasoned years.

For originations in 2023 and later, AUC was 0.656. Those cohorts are young, and the section below explains why that number should be read as provisional.

Predicted Default Risk Is Rising

Average Lumos Prime+ predicted PD on new originations stayed below 0.5% for the 2019 through 2021 vintages. For 2024 and 2025 originations it exceeded 1.5%, more than triple the earlier level. On the 27 conventional business loans in this sample originated in 2026, average predicted PD exceeded 5%.

That last figure rests on 27 loans. It is reported for completeness rather than as a finding, and a sample that size cannot settle what 2026 conditions look like.

The distinction the rest of the trend depends on is between predicted and realized. The rising numbers above are model outputs on newly originated loans, not observed defaults. Realized cumulative default rates in this sample moved differently: 3.27% for 2019 originations easing to 2.93% for 2022, then 1.69% for 2023 rising to 1.99% for 2025.

Those two series appear to contradict each other, and the strongest alternative explanation deserves a direct answer. If predicted risk is climbing while observed defaults on recent vintages are lower, one possibility is that the model has drifted upward and is now over-predicting on new originations. The competing explanation is seasoning: the 2023 through 2025 cohorts have not lived long enough to produce their defaults, so their cumulative rates are necessarily depressed and will rise as the loans age.

The seasoning explanation fits the rest of the evidence better. AUC on those same young cohorts is also below the seasoned cohorts' level, which is what a shortage of realized outcomes produces rather than what genuine model decay looks like. Measurement pulls the same way, since healthy borrowers who refinance away leave a portfolio without defaulting and reshape the denominator these rates are calculated on, a point developed in the analysis of SBA 7(a) prepayment speeds. Seasoning is the better reading of the evidence, though it is not yet a settled one, and it makes a specific prediction: as the 2023 through 2025 cohorts mature, both their realized default rates and the measured AUC on them should rise. If neither does, the drift explanation gains ground. The 2023 cohort crosses the three-year seasoning threshold during 2026 and is the first that can settle the question.

The recent cohorts are not safer. They are younger.

Scope of the Findings

Five lenders is a small number of institutions, even at 88,917 applications, so the absolute default levels reported here should not be read as representative of conventional small business lending generally. The rank ordering results are a different matter, because they compare loans within each lender's own portfolio and do not depend on the five being a representative sample.

There is a further point in the five-lender design that reads as a limitation and works as the opposite. Default is not an observed fact like death. It is a reporting classification, and whether a struggling borrower carries it depends on where the lender sets its delinquency threshold, how quickly it charges off, how readily it modifies or restructures, and the regulatory treatment in force at the time. Two identical borrowers at two lenders can produce different recorded outcomes, and a borrower who falls behind and later catches up may be recorded as a default under one lender's definition and never under another's. A model fitted inside one institution learns that institution's conventions along with its credit risk, and a model scored inside one institution is graded against them. Lumos Prime+ was fitted inside none of them, and the loans it learned from were labeled under several different conventions rather than one. Ranking consistently across five sets of conventions is evidence that what Lumos Prime+ measures is the borrower rather than the bookkeeping. It is also the general case of what 2019 showed acutely, when relief and forbearance rewrote those conventions across the whole market at once.

The findings concern PD only. Loss given default and expected loss, which Lumos Prime+ also predicts, were not tested in this study.

The 50% capture figure is retrospective, drawn from pre-2023 loans, and gross of the volume a lender could recapture from safe declines.

The trend in predicted default risk is a model output. It is a forecast of credit deterioration, not a measurement of it, and the cohorts that will confirm or contradict it are still seasoning.

At Lumos, we do not forecast the macro economy, and none of the above should be read as a call on where conditions go next. Findings on the SBA side, where the 7(a) portfolio's twelve month default rate reached 4.8% in March 2026, are examined separately in SBA 7(a) Default Rates Hit 4.8%. That is a different population operating under a different mandate, and the two sets of figures should not be blended, though they point the same direction.

Lenders weighing bureau scores, internal scorecards, and lifetime models against one another will find the evaluation criteria set out in our guide to building or buying a small business credit risk model. Those criteria apply to Lumos Prime+ as much as to anything else.

What the analysis does support is narrower than a marketing claim and more useful than one. Across five independent portfolios, Lumos Prime+ ranked defaulted loans above performing ones at an AUC near 0.70 on seasoned vintages, concentrated 35% of all defaults into a single decile, and reproduced underwriters' own decline decisions without being shown them. Those three results were produced on lenders' own books rather than on a development sample, and they are what everything above rests on.

Every number here is a prediction with a date attached, and the cohorts that would confirm or break it are already on the books.

The designed version of this report, with the full figures formatted for circulation to a credit committee, is available as a PDF. Lumos will also retro-score a lender's own historical originations and declined applications, which produces the equivalent of the analysis above on that lender's portfolio rather than on these five. Request a portfolio retro-score.

Frequently Asked Questions