How Lumos Measures Small Business Default and Prepayment

Executive Summary

There is no single small business default rate. Depending on how the measure is built, the same SBA 7(a) portfolio over roughly the same period reads 3.32%, 3.66%, 3.83%, 4.80%, or 5.12%. All five are correct. They count different events, over different windows, against different denominators.

The same is true of prepayment, which reads 8.34%, 9.34%, or 9.66% depending on whether the figure is a single month, a trailing year, or a calendar year.

This page defines each measure, gives its current reading, and says when to use it. Every default and prepayment figure in a Lumos analysis is defined here, so a reader who finds two different numbers in two different articles can see immediately why.

The practical rule underneath all of it: a rate is only comparable to another rate built the same way. Most apparent contradictions in credit data are two correct measures being set against each other.

The Default Measures

What counts as a default. The default event is the same throughout in every measure below. What differs is the window it is counted over and the denominator it is counted against.

Two things can vary in a denominator, and confusing them causes most of the trouble. The first is whether each loan counts once or is weighted by its balance. Every measure on this page is loan-weighted: a $50,000 loan and a $2 million loan count the same. The second is whether a loan that was on the books for part of the year counts as fully exposed. That is the difference between measures 1 and 2 below, and it is worth about 13%.

1. Exposure-adjusted annual rate

What it counts. Loans entering default during a calendar year, measured against loan-months at risk and annualized. A loan booked in December contributes one month of exposure, not twelve.

Current readings. 3.83% for calendar 2025. The 2015 to 2019 expansion averaged 2.39%.

When to use it. Comparing years to each other, or comparing a segment to the portfolio. This is the default measure in most Lumos analysis because it does not penalize a year with heavy new origination.

2. Simple count share

What it counts. Defaults divided by loans, with no adjustment for when a loan entered the book.

Current readings. 3.32% for calendar 2025. The 2015 to 2019 expansion averaged 2.05%.

When to use it. When the same measure must be computed across two datasets and only one carries exposure data. It runs roughly 13% below the exposure-adjusted rate, because loans that were only at risk for part of the year sit in the denominator at full weight.

3. Rolling monthly rate by original term

What it counts. A monthly conditional default rate computed within each original-term bucket, then averaged across the year and weighted by loans. It is built to sit alongside the prepayment measures below, so the two exits can be compared on identical footing.

Current readings. 3.66% for calendar 2025, against a 2015 to 2019 average of 2.38%. Loans of twenty-five years and longer read 1.75%.

When to use it. Comparisons of default against prepayment, where both sides must be built the same way, and analysis of the seven original-term buckets this file carries. It is not the measure for sector work or for the twenty-five year comparison, which use measure 1 and cover a larger population: about 326,000 loans in 2025 against about 283,000 here, with different term boundaries. The same finding is therefore sized differently on the two measures. Long-term loans read at roughly half the portfolio rate here and roughly a quarter on measure 1. Both are correct; a comparison must stay on one.

4. Trailing twelve months, as published

What it counts. Defaults over the twelve months ending at the reporting date, against the loans on the books across that window. The window rolls forward with each report rather than resetting at a year boundary, so consecutive readings overlap by eleven months.

Current readings. 4.80% for the twelve months to March 31, 2026, and 5.4% for the first half of SBA fiscal 2026, October through March, annualized. Both come from our analysis of SBA default rates, which documents their construction, and neither is recomputed on the constructions above. The second is a partial-year reading on a fiscal window, so it is the fiscal-basis counterpart of measure 5 rather than a sixth construction.

When to use it. Reporting the most recent available conditions. It is the freshest reading and the least suited to year-over-year comparison, because its window moves.

5. Partial-year annualized

What it counts. Defaults recorded so far in an incomplete year, scaled to an annual rate.

Current reading. 5.12% for calendar 2026 through May 31, on the exposure-adjusted construction.

When to use it. Watching a year in progress. Never set it beside a completed year without labeling it, since seasonality and reporting lag both push it around.

Why they differ, in one place

The first three cover the same calendar year and differ only by construction, a spread of about half a point. The last two are higher mainly because conditions deteriorated through 2025 and into 2026, so a more recent window reads worse.

The Prepayment Measures

Prepayment is a loan leaving the portfolio through payoff, refinancing, or business sale rather than through default or maturity. The two exits compete, which is why the pair is more informative than either alone.

Two constructions are in use. The readings below come from the rolling monthly rate by original term, which pairs with default measure 3. Sector-level prepayment, used in the industry benchmark, is built like default measure 1: exposure-adjusted, annualized, loan-weighted, on a calendar year. The two reconcile to within about 0.4 points, the residual reflecting different populations and the difference between an annual rate and a mean of monthly rates.

Our analysis of SBA 7(a) prepayment speeds reports monthly readings, so its figures are the first row of this table rather than the third or fourth. A single month is volatile and should not be compared to a multi-year average. The 2022 peak was the highest annual reading since the series begins in 1996, and pandemic-era liquidity contributed to it, so measuring a decline from that year overstates it.

Which Base Year

Choosing the comparison year matters as much as choosing the measure, and in one case it changes the sign of the answer.

2021 is a real base and the wrong one for measuring divergence. Relief lending suppressed default readings across both bank and SBA data. Bank delinquency at institutions outside the hundred largest reached an all-time low that year, and any percentage change measured from it largely captures relief programs running off. This is a real reading of recovery from a trough and the wrong instrument for measuring whether two groups diverged.

2019 is a clean year but not always a normal one. It is pre-pandemic and unaffected by relief, which makes it a common reference. But for some segments it sat at the top of a trend rather than in the middle of one. Prepayment on the largest, longest 7(a) credits ran 7.76%, 8.37%, 9.09%, 10.56%, and then 13.02% across 2015 to 2019, so a comparison anchored on 2019 alone shows a steep decline where a comparison against the five-year average shows none.

The 2015 to 2019 expansion average is the most defensible anchor for level comparisons. It contains no recession, ends before the pandemic, and averages away single-year peaks. Where a Lumos article compares against a single year instead, it says so.

The Bank Series

Federal Reserve delinquency and charge-off series are quoted in some Lumos analysis as a contrast to loan-level data. Three things about them.

They are seasonally adjusted quarterly readings, not annual rates, so they are not comparable in level to any SBA figure on this page. A delinquency rate counts loans currently past due; a default rate counts loans entering default over a period.

They are stratified by bank size, not borrower size. The split between the hundred largest banks and all others is a split between lenders, not between the borrowers those lenders serve.

Delinquency and charge-offs answer different questions. A delinquent loan sits in the numerator while it remains on the balance sheet. Once charged off, the balance leaves the numerator and the denominator together. The two series can sit on opposite sides of their expansion range at the same time, and currently do.

Why none of this can produce a small business delinquency rate is the subject of our analysis of what C&I aggregates can and cannot measure.

Predicted Default at Origination

One figure in Lumos analysis is not a rate at all. Predicted probability of default at origination, reported for 88,917 applications across five non-SBA lenders, is a model output describing expected performance for a loan being written, not a count of loans that have failed.

It answers a different question from everything above and is not comparable to any of it in level. What it supports is a statement about direction across vintages. Seasoned performance on conventional small business books is not published by anyone, which is why the conventional evidence takes this form.

The Lender Panel

Some analysis reports the distribution of results across SBA lenders rather than a portfolio total. Those are conditional annual default rates for individual lenders, and the panel is not fixed: 194 lenders carry a 2019 reading and 199 carry a 2025 reading.

A percentile from that panel describes where a lender sits among peers. It is not comparable in level to a portfolio rate, which is a single weighted figure across all loans.

Reading Two Lumos Articles Together

If two Lumos articles give different numbers for what looks like the same thing, work through this order.

Check the measure. Exposure-adjusted, simple count, rolling monthly by term, trailing twelve months, and partial-year annualized all describe defaults, and none equals another.

Check the window. A calendar year, a fiscal half-year, a trailing twelve months, and a multi-year average are four different windows.

Check the base. A change measured from 2021 and the same change measured from 2019 or from the expansion average can differ in size and occasionally in direction.

Check the population. The full 7(a) portfolio, a term or size bucket, a single industry sector, and a lender panel are four different populations.

Two figures that survive all four checks and still disagree are a genuine problem. In our experience most do not get past the first two.

Sources and Methods

Readings on this page are current as of the 2025 performance year and the first quarter of 2026. Updated as each measure is refreshed.