What small-business lending actually loses

Approval is not disbursement, and status is not loss

Before any rate, two counts have to be separated. 1,961,455 7(a) loans were approved, but 242,790 of those were cancelled and 22,918 more were committed without a disbursement date ever being recorded, so the population that could fail is 1,695,747. The separation is in the file rather than inferred: across both programs the 4 never-funded status rows cover 304,743 loans and carry 0 disbursement dates between them, which the table below shows row by row. What the file does NOT carry is a disbursed amount, only the approved one, so every dollar denominator on this page is approved dollars on loans that were disbursed, and is labelled that way. The status column has 5 tokens in 7(a) and 6 in 504, listed below; only CHGOFF is treated as a realised loss.

Source: SBA, 7(a) and 504 FOIA loan data Status tokens are the file's own, including the space-separated "P I F" that 7(a) writes and 504 does not. The build enumerates both sets and aborts on a token it has not been reasoned about, because a new status could be a loss category these rates would ignore. Methodology

What EXEMPT does to the denominator

297,494 7(a) loans carry the status EXEMPT, of which 297,492 were disbursed, 17.5% of the disbursed book and $168.9B of approvals. 0 of them carry a charge-off date, so the file records no realised loss on any of them. They therefore sit in the denominator of every rate here as loans that have not been charged off, and never in the numerator. That is the honest treatment of a status the file does not resolve, but it is worth stating plainly rather than burying: if any of them are in fact unrecovered, these rates are floors.

Duplicates and the rows that do not add up

The 7(a) files carry 1,918 exact whole-row duplicates and the 504 files 69, counted across every published column including the borrower fields, because dropping those columns first would make two different borrowers on identical terms look like one duplicated row. They are kept: the published file is the datum. Four smaller inconsistencies are carried rather than cleaned. 58 charged-off loans have no ChargeOffDate, so they sit in no age window and are absent from every fixed-age numerator. 182 have a zero or missing charge-off amount. 146 carry the charge-off status with no disbursement date and are excluded from every rate. And 6,984 loans have a GrossChargeOffAmount larger than their GrossApproval, which is why no dollar loss rate here is bounded by 100%: the file gives no field that would decompose the excess.

Source: SBA, 7(a) and 504 FOIA loan data Methodology

The clock, not the credit

A 7(a) loan that fails does not fail quickly. Taking the 23 vintages, FY1991 to FY2013, in which every disbursed loan has had 12 full years, 1,064,643 loans in all, the cumulative charge-off rate is 0.12% at age 1, 1.82% at age 2, 10.95% at age 5, and 17.41% at age 12. Read the other way: a vintage observed for 5 years has shown 62.9% of the losses these same vintages went on to record by year 12, and one observed for 2 years has shown 10.5%. That is the whole problem with a naive vintage curve, and it is worth a chart of its own before any rate is drawn.

Source: SBA, 7(a) and 504 FOIA loan data Ages are calendar years from ApprovalDate, so age 5 means the loan's own fifth anniversary rather than a fixed day count. The two programs season at different speeds: by age 5, 7(a) has realised 62.9% of its 12-year charge-offs and 504 only 33.6%. Methodology

What the naive rate does to recent vintages

The chart below draws both measurements on the same vintages. The columns are the fixed-age rate, on the 30 vintages FY1991 to FY2020 in which every disbursed loan has had 5 years. The dashed line is the naive lifetime-to-date rate on every vintage, including the 6 the clock has not finished with. On the seasoned years the line sits above the columns, because lifetime includes charge-offs that landed after year 5. On the censored years it collapses: FY2026 shows a naive rate of 0.00% against 4.01% for FY2020, and that fall is not an improvement in underwriting. It is that FY2026 has had none of its 23,243 disbursed loans reach the window.

Source: SBA, 7(a) and 504 FOIA loan data A vintage counts as comparable only when NOT ONE of its disbursed loans is still inside the 5-year window. Partial seasoning is not near enough: the loans that qualify in a half-observed year are the ones approved earliest in it, which is a biased sample of that year rather than a smaller one. Methodology

The vintages that can be compared, and the regimes inside them

On the 30 comparable vintages, the 5-year charge-off rate runs from 2.59% in FY2020 to 27.60% in FY2007, a range of 25.0 percentage points. Measured in dollars rather than loans the peak is 11.84% in FY2007, the same vintage. Counts and dollars differ because the loans that fail are not the average loan: FY2007 charged off 27.60% of its loans but 11.84% of its approved dollars. Nothing here should be read as a smooth trend across 30 years, because the program's own terms did not hold still.

Source: SBA, 7(a) and 504 FOIA loan data The last two columns of a censored row are deliberately not comparable with the rest: "lifetime to date" is printed for every year because it is what a naive reading would produce, and the censored rows are exactly where that reading goes wrong. Methodology

The guarantee moved, so the exposure did too

The SBA-guaranteed share of approved dollars is not a constant. Across the 36 fiscal years it runs from 69.4% in FY2007 to 84.2% in FY2021, and it moves by 3 percentage points or more from one year to the next in 6 of them. The two largest moves are +10.7 points in FY2021, then -11.0 points in FY2022. Those are policy, not lending: Congress temporarily raised the maximum 7(a) guaranty to 90% under the 2009 Recovery Act, extended it through the end of calendar 2010, and raised it to 90% again from late December 2020 through 30 September 2021 under the Economic Aid Act, after which it reverted. A long charge-off series that ignores this is comparing loans on which lenders retained very different amounts of their own money.

Source: SBA, 7(a) and 504 FOIA loan data | CRS R41146, SBA 7(a) Loan Guaranty Program | SBA Policy Notice 5000-1098, Recovery Act 90 percent guaranty The file records the guaranteed amount, never the statutory authority behind it, so the years above are read off the data and the statutes are cited rather than inferred from it. Methodology

A gross charge-off is not a bank loss

GrossChargeOffAmount is the whole charge-off on the loan, and SBA guarantees most of a 7(a) loan. Across the disbursed book, $519.3B was approved and $385.4B of that was SBA-guaranteed at approval, 74.2%, leaving $133.9B of unguaranteed exposure with the lenders. Splitting each realised charge-off in the same proportion as that loan's own approval-time guarantee puts roughly $17.1B of the $23.9B on SBA and $6.7B on lenders. That split is an ESTIMATE and the word is doing work: the file records the whole-loan charge-off and never how a given loss was actually shared, recovered against, or repurchased. It is the right order of magnitude for who is exposed; it is not an accounting of who paid.

Source: SBA, 7(a) and 504 FOIA loan data | CRS R41146, SBA 7(a) Loan Guaranty Program The guarantee ratio is capped at 1 per loan; 1 row in the file records a guaranteed amount above its gross approval and is capped rather than dropped. Methodology

The lenders, and why a certificate is not a firm

1,766,098 7(a) loans, 90.0% of the file, carry an FDIC certificate, and 3,923 of the 3,929 distinct certificates resolve to a row in this platform's FDIC institution table, covering 99.996% of those loans. A further 40,566 rows carry an NCUA charter number instead, of which 586 of 590 charters resolve, and 680 rows carry both. 155,471 loans carry neither, 7.9% of the file and $57.8B of approvals: those are lenders outside both registries, and no lender-level figure below includes them.

Source: SBA, 7(a) and 504 FOIA loan data | FDIC, BankFind institution directory | NCUA, quarterly call report data Approved dollars in this table cover every row carrying the key, including cancelled and never-disbursed ones, unlike every rate elsewhere on this page. The FDIC institution table joined here includes closed and merged institutions, which is why the match rate is as high as it is: a certificate whose bank failed decades ago still resolves. Methodology

The certificates that do not resolve

6 of the 3,929 certificates find no row in the institution table. Between them they carry 69 of the 1,766,098 loans that carry a certificate at all, 0.0039% of them, of which 58 were disbursed. They are few enough to print rather than characterise, so here they are with the name the loan file gives them. The institution table covers FDIC-insured institutions, so an identifier belonging to a trust company, a securities affiliate or another non-insured entity has nothing to resolve to; the miss is a fact about the identifier, not a defect in the join, and none of these rows is dropped from the file-level counts above.

Source: SBA, 7(a) and 504 FOIA loan data | FDIC, BankFind institution directory Where a certificate carries more than one recorded name, the name shown is the alphabetically first of them, so the table is reproducible rather than dependent on row order. Methodology

What a certificate identifies, and what it does not

An FDIC certificate identifies a CHARTER. It survives a rename and it disappears when the charter is absorbed, so a certificate-level series is a charter's history, not a firm's. Two counts make the point. Of the 3,929 certificates in the file, 3,898 carry exactly one lender name across their whole history, 99.2% of them, and among the 1,908 that originated across 20 or more fiscal years the figure is 1,905. Names in this file are therefore as of the extract, not as of origination: a certificate whose earliest loan is FY1991 is labelled with whatever that charter is called today. The lender table below prints each certificate's first and last approval fiscal year beside its current name, so wherever a name postdates the loans sitting under it, that is visible on the row rather than hidden. Nothing below should be read as "this firm's 36-year record". Separately, 478 of the matched certificates belong to institutions the FDIC no longer lists as active, and only 1,254 certificates have a last approval fiscal year of FY2025 or later, FY2025 being the last complete fiscal year in the extract. Most of the certificates in this file stopped originating under this program long before the extract was pulled.

The largest 7(a) lenders

Ranked by approved dollars on disbursed loans, over certificates with at least 500 loans inside the 5-year window. That floor leaves 216 certificates of 3,929, holding 86.5% of all loans in the window and 82.9% of their approved dollars. The rate column is the fixed-age rate on that certificate's own seasoned loans, so it is comparable across lenders in a way the lifetime column is not.

Source: SBA, 7(a) and 504 FOIA loan data | FDIC, BankFind institution directory The name shown is the FDIC's for the certificate, not the filer's text in the loan file, because the filer text is the same as-of-extract name and the FDIC directory is the registry this platform already keys on. The "approval FYs" and "names in file" columns are there so the entity-drift caveat above is visible on every row rather than asserted once. Methodology

Lenders are far apart, and size is not the reason

Across the 216 certificates above the floor, the 5-year charge-off rate runs from 0.94% to 37.76%, with a median of 4.89% and a middle half between 3.67% and 6.63%. The top of that range is 7.7 times the median. Counting loans and counting dollars also rank lenders differently, and the gap is arithmetic rather than interpretation: where a lender charges off a larger share of its loans than of its dollars, the loans that failed were smaller than its average. The widest such gap above the floor belongs to Bank of Hope, which charged off 27.82% of its 35,649 loans in the window but 3.56% of the approved dollars behind them, a gap of 24.3 percentage points.

Source: SBA, 7(a) and 504 FOIA loan data | FDIC, BankFind institution directory A certificate's window population is whatever it originated in the seasoned fiscal years, so lenders that entered the program recently are represented by their earliest cohorts only, and a lender that grew fast is measured on the book it had, not the book it has. Methodology

Where the losses sit, and where they cannot be located

The sector cut is clean. Every populated NaicsCode in these files is exactly six digits (the build checks that the number of distinct code lengths is 1 and aborts otherwise), so a two-digit sector rollup is never mixing grains; 225,772 rows across the two programs carry no code at all and drop out of the sector cut entirely. Over sectors with at least 10,000 loans in the window, 16 of 20 sectors covering 99.6% of the loans, the 5-year rate runs from 3.09% in agriculture, forestry, fishing and hunting to 12.22% in information.

Source: SBA, 7(a) and 504 FOIA loan data | US Census Bureau, NAICS Methodology

State, because county is a name

Geography here stops at the state. The file's ProjectCounty is a NAME with no FIPS code beside it, and county names repeat across states: 1,946 distinct names cover 3,241 distinct state-and-name pairs, and 644 rows carry no county at all. This platform's county tables are keyed on FIPS, so joining them would mean matching names through a gazetteer this build does not have, which is exactly where independent cities, Louisiana parishes, Alaska boroughs and spelling variants turn into silent mismatches. Rather than fake it, there is no county output at all. ProjectState takes 60 distinct values including US territories, and the 12 largest by approved dollars below hold 57.1% of the loans in the window.

Source: SBA, 7(a) and 504 FOIA loan data Methodology

504, which cannot be attributed to a bank at all

The 504 program lends through Certified Development Companies alongside a third-party lender, and its FOIA extract carries neither an FDIC certificate nor an NCUA charter number in any column. What it carries are names: 249 distinct CDC names across 229,049 loans, and 18,147 distinct third-party lender names across 219,444 loans, with 9,605 rows naming no third-party lender at all. A free-text field with 18,147 distinct values is a name, not a key, so nothing in this section is attributed to a bank and no name-matching has been attempted.

What 504 can still be measured on is the program itself. 188,063 of 229,049 approvals were disbursed, $117.4B of approved dollars, and 12,165 have been charged off for $5.6B. It seasons much more slowly than 7(a): by age 5 it has realised 33.6% of the charge-offs it records by age 12, against 62.9% for 7(a), so the censoring trap bites harder here, not less. On the 30 comparable vintages the 5-year rate peaks at 9.89%, against 27.60% for 7(a). The two rates are not like for like: a 504 project is financed by a third-party lender in first position alongside an SBA-guaranteed CDC debenture behind it, so a charge-off rate measured on the debenture says nothing about what the third-party lender lost on the same project.

Source: SBA, 7(a) and 504 FOIA loan data | CRS R41184, SBA 504/CDC Loan Guaranty Program 504 also uses different status tokens from 7(a) (it writes "PIF" where 7(a) writes "P I F", and it has NOT FUNDED and CLOSED, which 7(a) does not), which is why the two enumerations are kept separate throughout. Methodology