What the SBA Loan File Cannot Tell You: The Limits of FOIA Data in Underwriting

Every quarter SBA posts a loan-level record of its two main business lending programs, and every quarter parts of that record are quoted as if they answered questions the record was never built to answer. A charge-off count is read as a default probability. A list of the banks holding the most loans in an industry is read as a list of the banks that favor that industry. An approval count is read as a measure of demand. Each reading borrows the authority of a federal source while asking that source for something it does not hold.

This article sets out what the file leaves out, why it leaves it out, and what a feasibility study does about each gap. The short version is that the gaps are not accidents of data handling. Put plainly, most limitations of SBA's FOIA loan data are limitations of what the law lets SBA release. They follow from the disclosure rules the agency works under, which sort loan information into what is generally released and what is generally withheld, and the withheld side holds most of what an underwriter needs. The file is a record of approvals and of how they ended. It is not a credit file, and nothing done to it after download turns it into one.

The Loan Analytics loan dataset is built from these files as of September 30, 2026. Where this article prints a loan figure, the figure comes from that dataset through its key, and where it states a rule, the rule is cited to the Code of Federal Regulations or the U.S. Code. The reconciliation companion tests the files against SBA's own year-end totals. This one is about the edges of what they can support.

What SBA publishes, and the rule that decides it

The 7(a) program's records come in four files, one per decade, and the 504 records in two files spanning twenty years each, with a data dictionary alongside ^[1]. The agency says the files are refreshed each quarter and usually appear about a month after the quarter closes ^[2]. The release behind this article is the one dated as of September 30, 2026 ^[3]. One row is one approved loan: who borrowed, which bank or certified development company carries it, how much was approved, when, on what term, in which state and county the project sits, which industry code the borrower reported, and where the loan stands now, described in a handful of coarse status words.

That list is not arbitrary. SBA's FOIA regulations carry an appendix that sorts the agency's records into two groups. On the disclosed side sit who received approved loans, by name and business address ^[4]; what kind of loan was made and for how much, with its terms, rate, maturity and broad purpose, plus statistics on lending and defaults ^[5]; and which banks took part ^[6]. Laid next to the columns of the file, the disclosed side reads almost like a field list. The loan file is roughly what that half of the appendix looks like when it is put into a spreadsheet.

The withheld side reads like the first page of an underwriter's checklist. It lists loan-level information about applications that were declined, withdrawn or cancelled, or that are still pending ^[7]. It lists loan-level information about delinquencies, defaults and losses, and then, in an entry that shapes every rate anyone builds from the file, it lists a loan's status itself, with only two exceptions: a loan that has been charged off and a loan that has been paid in full ^[8]. It lists the applicant's financial statements and credit reports, its business plans and projections, and its pricing and payroll details ^[9]. Put the two halves side by side and the shape of the public record is plain: the program's activity is disclosed, and the borrower's credit is not.

Behind the appendix stand the statute's exemptions. The fourth exemption of the Freedom of Information Act shields trade secrets and confidential commercial or financial information that a person hands to the government ^[10], and the sixth shields personnel, medical and similar files where release would invade personal privacy in a way that is clearly unwarranted ^[11]. SBA's own rules define confidential commercial information by reference to the first of those exemptions ^[12]. The statute also tells an agency to hand over whatever part of a record can be separated out once the exempt parts are cut away ^[13]. That requirement is close to a description of the loan file itself: the segregable remainder of a very large number of loan records after the commercial and personal material has been taken out.

The practical point for a lender is that these gaps are durable. A field that is missing because of a disclosure rule will not turn up in next quarter's release because an analyst would like it, and an analysis that waits for it is waiting on a change in law or agency policy. A study that uses the file has to be designed around its gaps from the start. The figures this article leans on are gathered in Table 1, each with its key, so that a reader can see exactly which numbers come from the dataset and which come from rules.

Loan Data Behind This Article

Every figure in this table comes from the Loan Analytics loan dataset, analytics-loan-data freeze-3 (commit 360c856), and carries its key. SBA rates are approval-cohort floors and are withheld when fewer than 30 loans stand behind them.

The open loan: a status the rule withholds

In the September 2026 release, a disbursed loan that has neither paid in full nor been charged off carries one status, EXEMPT. The word is meant literally: the status is exempt from disclosure under the appendix quoted above ^[8]. A 7(a) loan in the file can sit in only five places (committed and not yet disbursed, cancelled, paid in full, charged off, or exempt), and the last of those covers every condition an underwriter would want to tell apart. A loan paying on schedule, a loan several payments behind, a loan on a payment deferral, a loan in liquidation and a loan whose guaranty SBA has already honored all look the same.

The number of loans behind that one word is large. In the 7(a) approval cohort of FY2017 to FY2021, the dataset counts 68,103 ^[14] loans that were disbursed and are still open, each with its status withheld, and the 504 cohort of the same years holds 22,353 ^[15] more. These loans are five to ten years old. Some are performing without a missed payment. Some will be charged off next year. The file cannot say which, and nothing in it ranks them by risk.

Two cells make the scale concrete. Among 7(a) loans to gas stations approved in FY2017 to FY2021, 1,033 ^[22] of the disbursed loans are still open, roughly a third of the cohort. Among 504 car wash loans approved in the same years, 199 ^[23] are still open, roughly half. Table 3 sets out the counts from the inputs in Table 2.

Rounded as shown; lines may not foot by one unit in the last place.

What the withheld status does to a rate can be shown with a deliberately simple illustration. The gas station cohort's charge-off rate to date is 1.16 percent ^[22], an approval-cohort floor. Table 4 asks what that rate would become if some share of the still-open loans were to charge off later, at a share that is an assumption and nothing more. If none of them did, the rate stays at 1.16%. If one open loan in twenty did, the rate would be 2.91%, about 2.52x the figure the file shows. If one in five did, it would be 8.19%. The point is not that any of those outcomes is likely. The point is that the file gives no evidence for choosing among them, because the one field that would separate a healthy open loan from a failing one is the field the rule withholds.

Why a rate assembled this way can only understate, and how a within-five-years rate pins the horizon, is the business of the floor piece elsewhere in this series. The narrower point here is about warning. A lender reading the file for signs of stress in an asset class gets none until the stress has run its full course and ended in a charge-off. Delinquency, deferral, liquidation and guaranty purchase all happen inside the exempt label, so the file reports the end of a failure years after its start.

An archived post on this site, listed among the related pages below, tells the story of a single earlier release in which more detailed servicing codes briefly appeared and then went away. Whatever that episode showed, the release dated September 30, 2026 carries none of that detail ^[3], and any delinquency figure taken from the earlier snapshot describes one day that the current files do not repeat. A study that cites such a figure should give its date and say plainly that the current release does not carry the same detail.

What should a lender do with the open-loan gap? First, read any SBA charge-off rate as a floor and ask for its open-loan count; a rate that cannot state how many of its loans are still open is not finished. Second, look for program-level performance figures where SBA publishes them, because statistical data on loans and defaults falls on the disclosed side of the appendix ^[5], even though loan-by-loan servicing status does not. Third, rely on the lender's own servicing history for loans of the same type, which is the only record that sees the middle of a loan's life. A feasibility study adds a fourth answer: it does not import an industry rate as the project's chance of failure at all, but tests the project's own cash flow against the debt under stated assumptions and shows how far those assumptions can move before the project can no longer carry the debt.

Approvals only: the applications the file never sees

The file begins at approval. Loan-level information about declined, withdrawn, cancelled and pending applications is generally kept back ^[7], so there is no row for the borrower who applied and was turned down, no row for the borrower who was offered terms and walked away, and no row for the application that was still being worked when the quarter closed. A reader sees the deals that lenders were willing to approve and nothing of the deals they were not.

For 7(a), the population is selected a second time by the program's own design. SBA lends only where the credit an applicant wants cannot be had on reasonable terms from sources other than federal, state or local government without the guarantee, and the lender must show it weighed conventional factors such as the applicant's industry, how young the business is, and whether its collateral is adequate ^[24]. The file is therefore a record of borrowers who could not get comparable conventional credit, filtered again by the lenders who decided to approve them. That double selection cuts both ways for a reader. It means SBA outcomes are not a fair stand-in for conventional loans in the same industry, and it means they are not a fair stand-in for a marginal applicant either, because the weakest applications in any industry never became rows.

Three misreadings follow from forgetting this. The first treats an approval count as demand. A state with few approvals in a class may have few businesses of that kind, few owners who applied, many applications declined, or plenty of conventional credit that made SBA unnecessary, and the file cannot tell these apart. The second treats an outcome rate as the risk of any project in the class. An approval-cohort rate describes projects that lenders judged sound enough to approve, and a project weaker than the typical approved deal is not represented in its denominator. The third treats an approval as a funded loan. Two to three years on, the FY2024 approvals still waiting for a first disbursement at the release date stood at 3.42 percent ^[19] in 7(a) and 17.30 percent ^[20] in 504. How an approval can end with no dollar lent, and why that ending stays apart from a loss, is traced in the companion on deals that never fund within this series.

A feasibility study answers the demand question from the market itself, not from the program's acceptance record. It counts the population and households in a trade area drawn around the real site, the traffic that passes it, the competitors already serving it and their prices, and it builds a capture rate from that evidence. SBA counts enter a study as evidence of one narrow thing: that lenders in the program have financed this kind of project in this state or county, at about these loan sizes. They are context for the lender's file, not proof that the market wants the project.

No financials, no coverage, no collateral value

The regulation that sets SBA's lending criteria lists what a lender may weigh in reaching a credit decision: the credit history or score of the applicant, its associates and its guarantors; what the business earns; and, where they apply, its equity and its collateral ^[25]. The file holds none of those things. The applicant's financial statements, credit reports, business plans and fiscal projections sit on the withheld side of the appendix ^[9], so the inputs to every credit decision recorded in the file stayed with the lender.

Some columns look closer than they are. The 7(a) files carry a collateral indicator, a yes or no that says whether collateral was pledged but not what it was, what it was worth, or where the lien sat. They carry the approved term in months and the initial interest rate, a jobs figure reported at approval, a code for the age of the business, and a franchise code where one applies. The 504 files carry the certified development company's debenture and the third-party lender's name and dollars, but not the borrower's own contribution or the full project cost. None of these is a coverage ratio, a loan-to-value ratio, an appraised value or a measure of the owner's equity. Any coverage figure attributed to an SBA class or state cannot have come from the file, because the file contains no income, no expense and no debt service.

Conventions. This article follows the conventions of the Loan Analytics worked examples. Coverage is tested against the program floor the site's program page states, where the program has one, and against a stated 1.25x lender assumption, and the higher of the two binds. Equity is measured at the loan the cash flow supports, with fees recomputed on that loan. An acquisition is tested on the seller's historical earnings, not on projections.

A study fills this gap from primary evidence about the project, not from the record of other borrowers. In a study's model, coverage is computed from the project's own operating projection, or for an acquisition from the seller's historical statements, and is checked against both tests the conventions line names. Where the cash flow cannot carry the amount requested, the study sizes the loan down to what it can carry, reworks the fees on that smaller loan, and shows the shortfall as additional equity the sponsor must bring, instead of leaving the gap hidden. The Tennessee gas station worked example, listed among the related pages, runs those tests on one project from start to finish.

Collateral value needs a different professional. The FDIC's appraisal rule defines an appraisal as a written opinion of market value, prepared independently and impartially by a qualified appraiser, for a property that is adequately described, as of a stated date ^[26]. SBA, for its part, may make a loan conditional on professional appraisals of the assets of the applicant and its principals, on a survey, or on a feasibility study ^[27]. Nothing in a feasibility study is an opinion of what the property would sell for. Its job is narrower and different: to show whether demand, project cost and operating income are enough to service the proposed loan, while the opinion of value is left to the appraisal. Reading the file correctly means knowing that neither of those documents can be reconstructed from it.

One shortcut deserves a specific warning. Analysts sometimes divide a median loan size by a guess at project cost and call the result typical leverage. The file has no project cost for 7(a) and no borrower contribution for 504, so the result is the guess restated. A loan size distribution tells a lender how large SBA loans in a class have been. It does not say how much of a project they financed or how hard the debt pressed on the cash flow.

No sponsor record: the people behind the loan stay private

Underwriters judge people as well as numbers, and the file is silent on people by design. An owner's personal financial statements and tax forms, resumes and career history outside government, and communications about the applicant's character all sit on the withheld list ^[28], with the statute's privacy exemption behind them ^[11]. The file can say, through a coarse age code, whether a business was new or established at approval, and through a business type code how it was organized. It cannot say whether the owner had run a similar operation before, whether the managers named in the application are still the managers, or how deep the guarantors' pockets were.

That last gap matters more than it first appears. Under SBA's loan conditions, anyone owning 20 percent or more of the business is generally required to stand as a guarantor ^[29], so the strength of the guarantors is part of what stands behind every SBA loan in the file. Two loans of the same size to the same kind of business can carry very different support, one backed by an owner with substantial outside assets and one by an owner with none, and the file records them identically. Outcome rates for a class average across that difference without revealing it.

Weighing the credit of the guarantors is part of the lender's own analysis under SBA's lending criteria ^[25], and a study does not attempt it. What a study can do is describe the operating plan the sponsor proposes and test it against the market: the staffing the projection assumes, the wages it pays, the hours it keeps, and whether the operator's stated experience matches the kind of business being financed. Where a projection depends on a skill the sponsor has not shown, such as running a fuel margin or filling a care facility, the study says so and shows what happens to the cash flow if the ramp is slower. None of that can be borrowed from the loan file, because the file never held it.

Charge-off dollars, but no recoveries

When a loan is charged off, the file records the date and a gross charge-off amount. It records nothing that follows. There is no field for what SBA paid a lender to honor the guaranty, no field for liquidation proceeds, no field for the price at which a loan or its collateral was sold, and no field for what was collected from guarantors. The regulation leaves no doubt that recoveries happen and are accounted for: payments and recoveries on a loan, sale proceeds included, are split between SBA and the lender in proportion to their interests ^[30]. The amounts simply sit on the withheld side, where non-statistical information on losses is listed ^[8].

That shapes what the dollar rate can mean. Take the 7(a) approval cohort of FY2017 to FY2021. Counted by loans, charge-offs to date reach 6.33 percent ^[16] of disbursed loans, an approval-cohort floor; counted in money, and still a floor, the gross amounts charged off equal 1.95 percent ^[17] of what those disbursed loans were approved for. In 504 the same money measure is 0.29 percent ^[18] of debenture dollars. The dollar rate is useful, but it is easy to overstate what it measures. Its numerator is a balance at the date of charge-off, and its denominator is the amount approved years earlier, so it mixes two moments of a loan's life. It is not a loss given default, because it carries no recovery; it is not a net loss to the lender, because the lender's share and SBA's share are not separated; and it is not a loss rate on outstanding balances, because the balances are not in the file.

A second archived post on this site, on what lenders recover from charged-off gas stations, car washes and hotels, works through this gap for three special-purpose property types and is listed among the related pages. For a methodology reader the lesson is general. A recovery assumption in a study, wherever it appears, is an assumption, and it should be labeled as one and sourced to the appraisal, to the collateral's alternative uses, or to the lender's own workout history. It should never be attributed to the SBA loan file, which cannot support it.

For a lender, the practical questions are about the collateral itself. How specialized is the building, and how many buyers would want it in its current use? What would it cost to convert? Are there environmental or permit conditions that follow the property? A study addresses these as part of the market analysis, by describing the property's competitive position and the depth of the market for its use, while the appraisal states the value. Together they give the lender a reasoned view of the downside that the public file cannot.

The lender column names today's holder

Which banks take part is on the disclosed side of the appendix ^[6], and the 7(a) file names one for every loan. That name is easy to misread. It identifies the bank the loan is currently assigned to, or the last bank to hold it, not necessarily the bank that underwrote it. Loans move. Under SBA's rules any lender may sell a 7(a) loan outright, sell participations in it, or pledge it, subject to the conditions the same subpart sets ^[31]. A lender can transfer its whole interest in a loan to another participating lender ^[32], though the agency must agree in writing beforehand and is free to refuse, and no interest may go to a lender outside the program ^[33]. In the file, the open loans of a merged bank show under the surviving bank, while a loan that closed before the merger keeps the name of the bank that held it then. The same institution can therefore appear under different names in different years, and an originator that sold or merged away its book can vanish from the record entirely.

The 504 file works differently again. Its lender field is the certified development company, keyed by SBA's location identifier, and the senior lender appears only as free text with no identifier. A ranking of 504 lenders is a ranking of development companies, not of the banks that provided the first lien.

Scale is easy to underestimate. Gas station 7(a) loans approved from FY2017 to FY2026 are held, as of the release date, by 520 ^[22] distinct lenders. A list of the five largest holders in that class says something real about where the paper sits today. It says much less about which lenders have an appetite for the next gas station loan in a given county, because purchases, sales and mergers have rearranged the names since the loans were made.

A study that relies on a lender list states what the list is: a ranking of current or last holders. Where origination matters, for example to show that lenders active in a market have financed comparable projects, the study traces the bank's FDIC certificate through the FDIC's BankFind Suite, which lists each insured bank's mergers, acquisitions and history ^[34]. It never presents a holder ranking as evidence of a lender's skill, its loss experience, or its willingness to lend on the project at hand, and it never names a borrower from the file.

The industry code belongs to the borrower

The NAICS code on an SBA loan describes the borrower's business, not the property and not the use of proceeds. SBA's size rules assign a concern its primary industry by looking at how its receipts, its workforce and its costs are spread across the lines of business it runs ^[35]. NAICS itself sorts establishments into industries according to how alike their production processes are ^[36], and where a single establishment carries on several activities, the code it receives usually follows whatever it mainly produces or sells ^[37].

Three consequences follow. A mixed operation is coded by whichever activity earns the most, so a fuel station with a car wash and a convenience store appears in one code, and the activity that matters most to a particular lender may be invisible in it. A class built from codes is a class of borrowers, so loans to operating businesses that happen to own their real estate sit beside loans to businesses that lease, and the file does not say which is which. And codes change across NAICS vintages, so a series that crosses the switch from one edition to the next needs a concordance before it can be read as continuous. Shared and split codes, and how the dataset treats each, are the subject of the companion piece on code mapping in this series.

For a study, the code is a starting point for finding comparable loans, not a description of the project. The study classifies the project by what it will actually earn from, with the revenue mix built line by line in the operating projection, and it says so where the project's mix differs from the class it is compared with. A gas station whose margin comes mostly from the store, or a hotel with a large food and beverage operation, should be described in those terms, with the SBA class figures offered as the nearest public benchmark and labeled as such.

Location is a state and a county

The file places each project in a state and a county, using the county name as the file spells it. It does not give a project street address, a parcel number or coordinates. The borrower's own name and business address are generally disclosed ^[4] and do appear in the raw files, but that address belongs to the business, which may differ from the project site, and the Loan Analytics dataset does not load borrower names or street addresses at all. County is therefore the finest place grain the dataset can report.

That is too coarse for underwriting most real estate. A county can hold one interstate interchange or ten, a downtown and a rural fringe, a saturated corridor and an empty one. A count of gas station approvals in a county says nothing about whether those stations sit across the street from the subject site or forty minutes away, and a county total cannot be turned into a trade area by any amount of arithmetic. State figures are coarser still, and a state ranking says as much about a state's size and its lenders' habits as about any market.

A study draws its trade area around the actual site, from drive times and the road network, and identifies competitors by address from fieldwork and public listings. The county and state figures from the dataset then do a modest job: they show whether SBA lenders have been active in this kind of project in the area at all, and at what loan sizes. The site's state-level loan tables and its property-type loan tables report the figures at precisely that grain, never finer.

A dated snapshot, revised every quarter

Every release is a photograph taken on one date. The files are refreshed quarterly ^[2], and between releases statuses change: an exempt loan pays off, a committed loan cancels, a struggling loan is charged off. A figure taken from one release is true as of that release's date and is likely to differ in the next, even when nothing about the method has changed. A rate quoted without its as-of date cannot be checked, and a comparison of two rates taken from different releases may measure the passage of time more than any difference between the things compared.

Recent approvals are the least settled. Gross approvals in the file include loans that will later be cancelled, which is SBA's own counting convention, and the newest cohorts have had the least time to fund, cancel, pay off or fail. Rates on those cohorts move most from one release to the next. The dataset handles this by rating only seasoned cohorts and stating their seasoning in years; a reader of any other analysis should ask what cohort and what seasoning stand behind a figure before relying on it.

Two smaller habits keep a dated figure honest. Cite the release date printed with the files, not the page's own dates: the dataset page shows a last modified date in April 2026 ^[38] while listing files as of September 30, 2026 ^[3], so the page's metadata cannot date a release. And keep the release fixed for the whole of one analysis, so that a count, a rate and a ranking quoted together all describe the same day. The figures in this article all come from the release dated September 30, 2026 ^[3].

How a study fills each gap

Table 5 pulls the gaps together: the question a lender asks, what the file shows, the rule that explains it, and what a feasibility study relies on instead.

The table also works as a reading checklist for any figure drawn from the file, whether it appears in a study, a lender's credit memo or a published article. A careful reader asks the same few questions every time.

• What release is it from? A figure without an as-of date cannot be checked, because the files are replaced every quarter ^[2].
• If it is a rate, is it labeled a floor, and how many open loans sit in its denominator? The open-loan count is the measure of how much of the cohort the rate cannot see ^[8].
• If it is a count, does it include approvals later cancelled? SBA's convention counts them, which suits measuring program activity and does not suit measuring funded loans.
• If it names lenders, does it say they are holders? A holder list is not an origination record ^[31].
• If it describes an asset class, does it say the class is built from the borrower's industry code? The class is a set of borrowers, not of buildings ^[35].
• If it is a rate for a small group, how many loans stand behind it? Loan Analytics prints no rate resting on under thirty loans, and the small-sample companion gives the reasoning.

Each question takes a minute to ask and saves a lender from building a credit view on a number that means something other than it seems to. None of them requires distrusting the file. They require reading it as what it is.

What the file does well

None of this makes the file weak. Within its limits it is unusually solid. Its approval counts and dollars match SBA's own year-end reporting for every fiscal year from FY2010 to FY2025, it records every approval in both programs rather than a sample, and the dataset built from it covers 647,329 ^[21] approvals across 7(a) and 504 in FY2017 to FY2026. For questions about program activity it is the best public evidence there is: how many loans of a kind lenders have made, in which states and counties, at what sizes, in which years, through which program, and how often the seasoned ones have ended in a charge-off at least.

A feasibility study uses the file for exactly those questions and states the limits each time it does. Each loan-outcome page on the site gathers counts, sizes, holders, counties, trends and the two floors in one place, with the keys behind them, and a separate article in this series walks through reading one block by block. Two national examples are the gas station page for 7(a) and the car wash page for 504. What a study does not do is ask the file about the project's market, its cash flow, its collateral, its sponsor or its odds, because the disclosure rule kept those answers out of the public record before the first row was written ^[9].

The honest summary for a lender is short. The SBA loan file tells you what the program did, and it tells you that reliably. It does not tell you what any one borrower could afford, what any one building was worth, how any open loan is performing, or how much was recovered when a loan failed. Those answers come from the borrower's records, the appraisal, the lender's own history and the feasibility study, and a credit decision that skips them in favor of the public file is resting on the part of the record that was never meant to carry it.

Related Pages on Loan Analytics

• The Loan Analytics methodology series (all articles)
• How a feasibility study is done: the Loan Analytics methodology
• SBA feasibility studies (7(a) and 504)
• Loan data by state
• Loan data by asset class
• Worked examples (case studies)
• Worked Example: SBA 7(a) Assisted Living Feasibility Study in Utah
• Worked Example: SBA 504 Express Car Wash Feasibility Study in Maryland
• Worked Example: SBA 7(a) Gas Station Feasibility Study in Tennessee
• Worked Example: SBA 7(a) Gym Feasibility Study in Hawaii
• Worked Example: SBA 504 RV and Boat Storage Feasibility Study in New Jersey
• Worked Example: SBA 504 Owner-Occupied Retail Feasibility Study in Michigan
• For One Quarter, SBA Showed Which Loans Were Late. Then the Window Closed.
• What Do Lenders Recover on a Charged-Off Gas Station, Car Wash or Hotel? Nobody Publishes the Answer
• How the Loan Analytics SBA Loan Dataset Is Built: Six FOIA Files Reconciled to SBA's Year-End Reports
• USDA Business and Industry Loans in Public Data: What USAspending Shows and What It Hides

Commission a Feasibility Study

Loan Analytics prepares lender-grade feasibility studies for SBA 7(a), SBA 504, USDA Business and Industry, conventional and EB-5 financing, built on the methods this series sets out. A study is not an appraisal: it tests whether a project's market, costs and cash flow can carry the debt a lender is asked to approve. To discuss a study for your project, contact Loan Analytics.

Daniel Smith, MAI, Loan Analytics