How the Loan Analytics SBA Loan Dataset Is Built: Six FOIA Files Reconciled to SBA's Year-End Reports

A feasibility study that quotes an SBA loan figure is making a claim about a public record, and the credit officer who reads it is entitled to ask which record, which release and which definition stand behind the number. Among the conditions SBA may attach to a business loan, its regulations name a feasibility study alongside appraisals and a survey ^[1], so a study written for an SBA file will often lean on the program's own lending history: how many loans went to a kind of business, how large they ran, how many were approved and never funded. This article explains how the Loan Analytics loan dataset turns the files SBA publishes into those figures, and how every step was tested before a single figure reached a page.

The ground covered here is the plumbing. It runs from what SBA posts and what its as-of date means, through what one row of the file represents, how approvals are counted and what the 504 dollar column actually measures, to how the six files were reconciled with each other and with a separate SBA report, and how the whole build was run twice to show that it gives the same answer both times. Three neighbouring questions have articles of their own in the series: the matching of each loan's industry code to one of the asset classes, the reason a charge-off rate measured today can only understate the eventual loss, and the fields that the public file leaves out or withholds. This one stays with the record itself.

Why the Way a Dataset Is Built Matters in a Loan File

A study is read by people who did not build it. When it says that SBA lenders approved so many loans for a type of business in a given state, the credit officer cannot rerun the count from scratch. What the officer can do is check that the figure names its source, its release and its definition, and that the source is one the agency itself stands behind. A loan figure that fails those three checks is, for underwriting purposes, an anecdote, however precise it looks on the page.

Three things can go wrong with a published SBA loan figure even when the arithmetic is flawless. The first is vintage. SBA replaces its loan files every quarter, and a count of cancelled loans or a charge-off rate taken from a spring release will not match the same measure taken from an autumn one, because loans keep moving between statuses. The second is definition. A total can count every approval or only the approvals that funded; it can add the whole 7(a) loan or only the share SBA guarantees; it can add the 504 debenture or the full project the debenture helped pay for. The third is completeness. A download that silently lost a decade of rows, or a file that was truncated on the way in, still produces tidy tables, and nothing in the tables themselves betrays the gap.

The definition problem is sharper than it looks. The statute behind the 7(a) program, 15 U.S.C. 636(a), describes SBA's role as participating with banks and other lenders on an immediate or a deferred, guaranteed basis ^[2]. Under the general rule of the same section, SBA's guaranteed share of a 7(a) loan is 75 percent of the balance outstanding at disbursement when that balance exceeds $150,000, and 85 percent when it does not, subject to exceptions the statute lists ^[3]. A sentence that says SBA lent a certain sum can therefore mean the loans, the guarantees on those loans, or something in between, and the gap between the readings is a sizable part of every loan. The sections below take the three in turn, each with the check the build runs against it.

What SBA Publishes on data.sba.gov

SBA posts the loan-level record of its 7(a) and 504 business loan programs as one dataset on its open data site, data.sba.gov. There are six data files: the 7(a) loans are split into four, one for each decade, while the 504 loans are split into two that cover twenty years apiece ^[4]. A data dictionary that names and describes the fields is provided beside them ^[5], and it is published as an Excel workbook rather than as a CSV like the data ^[6]. Table 1 lists the seven resources in the words the dataset page uses for them.

The same page says how often the record changes: SBA refreshes the 7(a) and 504 FOIA data every quarter, and each release usually lands roughly four weeks after its quarter ends ^[9]. The page lists only the current release, and each new one takes the place of the last rather than sitting beside it. A file downloaded in the spring and a file downloaded in the autumn are two different documents that happen to share a name pattern, and any figure drawn from either one belongs to that release alone. The Loan Analytics dataset is pinned to a single release, the one whose files are dated as of September 30, 2026, and every figure in this series is read from that release and from nothing else.

The word FOIA in the dataset's title points to the Freedom of Information Act. Among other duties, that statute requires each agency to post in electronic form copies of records it has already released when it expects further requests for the same records, or when they have been asked for three or more times ^[10]. Whatever SBA's own reasons for posting the files in this form, the result for a lender is the useful part: the loan record is a public document that anyone can download and check, which is exactly why a study can cite it. The page lists the license as U.S. Government Works and names SBA's Chief Data Officer as the contact ^[11].

One small detail on the page is worth a lender's attention. The page's own metadata reports that it was last changed on April 28, 2026 ^[11], while every loan file it lists, the 504 pair ^[7] and the four 7(a) files ^[8] alike, is dated as of September 30, 2026. A page's metadata and the files it serves need not agree, and here they do not. The build therefore never takes the release date from the page. It reads the as-of date from the file names and, more importantly, from a column inside every row.

The As-Of Date and the Federal Fiscal Year

Every loan file on the page carries the same as-of date, September 30, 2026, in its name ^{[7] [8]}, and every row in every loan file carries that same date in its AsOfDate column; the dictionary, which holds no loans, carries no as-of date at all. The date is not arbitrary. The fiscal year of the Treasury starts on October 1 and closes on September 30 of the next calendar year ^[12], so the release is a year-end snapshot: it holds every approval through the close of FY2026, with each loan's status as SBA recorded it on that closing day.

Two consequences follow for anyone who quotes the data. The first is that every count in the dataset is a fiscal-year count. FY2025 began on October 1, 2024 and ended on September 30, 2025, and a study that relabels a fiscal-year total as calendar 2025 is three months out of step with its own source. The mistake is easy to make because the numbers look plausible under either label, and it is hard for a reader to catch without the source in hand. The second consequence is that the as-of date describes status, not approval. A loan approved in FY2017 appears in this release with whatever status it held nine to ten years later; a loan approved in FY2025 appears with whatever status it reached in a year or two. Rates built on those statuses depend on how long each approval cohort has had to season, which is why the series treats them as floors and explains the mechanics separately.

FY2026 in this release is a short year. No approval in the file is dated between October 1 and November 12, 2025, while FY2025's first approval falls on October 1, 2024. The dataset records that gap as it found it and does not fill it, so FY2026 holds roughly ten and a half months of approvals: 56,130 ^[13] in 7(a) and 6,075 ^[14] in 504. Set beside FY2025's 78,078 ^[15] 7(a) approvals, the FY2026 count would suggest a sharp fall in lending. One reason the FY2026 count would come in lower is the shorter span it covers, and a study should never present the two years side by side as a trend without saying so.

A handful of dates in the file also fall after the as-of date. Some charge-off dates land in October 2026, and a few first-disbursement dates run as late as June 2028, which cannot describe money that had already moved. The build leaves these values as SBA published them rather than correcting them, and it measured how far they reach: together they can shift a cohort's count of charge-offs or disbursements by no more than a couple of dozen loans.

What One Row of the File Is

One row is one approval. In the 7(a) files it is a loan that SBA agreed to guarantee in the fiscal year the row names; in the 504 files it is a 504 loan approved through a Certified Development Company in that year. A 7(a) row carries 42 fields and a 504 row 39, and the build checked that every row of every file has exactly its program's width, with no short or overlong lines that would signal a broken record.

The fields fall into a few groups. The approval itself: the approval date, the fiscal year, the gross amount approved, the term in months and, in 7(a) only, SBA's guaranteed amount and the initial interest rate. The lender: in 7(a), the bank's name, address and federal identifiers; in 504, the CDC's name and address, plus the name, city, state and dollar amount of the third-party lender that took the first lien. The business: its industry code and description, business type, business age and franchise code. The place: the state and county where the project sits, the SBA district office and the congressional district. The loan's life so far: the first disbursement date, the current status, the paid-in-full date, the charge-off date and the gross amount charged off. The borrower's name, street, city and ZIP code are in the files as well. The build never loads those four fields into its working tables, the loader checks that they are absent, and a scan checks every file the build commits for borrower names and street addresses.

There is no loan number. The 504 regulation says SBA issues a loan number once it agrees to guarantee part of a project's funding ^[16], but no column in either program's public file carries such a number. That absence has a practical cost. A small number of rows repeat another row in every published field, borrower fields included, and without an identifier there is no way to tell a duplicated record from two genuinely separate loans that happen to match in every column. The build keeps every row. The deciding test is the witness described below: in every year where SBA's own published totals tie to the file, they tie with every repeated row kept in. The keep decision is therefore proven for those years and unproven only for a small remainder in the early years where the witness and the file do not tie.

The dataset places each loan by ProjectState, the state where the project occurs, rather than by the borrower's mailing state. The asset being financed sits in the project state, which is the question a feasibility study asks, and the two fields agree for nearly all recent 7(a) loans in any case. What the location fields cannot do, such as place a loan at a street address, is part of the companion article on the file's limits.

Gross Approvals: Counting the Way SBA Counts

The dataset counts approvals gross. Every approval is counted in the fiscal year it was approved, including approvals that were later cancelled before any money moved. This is SBA's own convention: the agency's year-end totals count approvals this way, and the public file ties to those totals only when the cancelled rows are kept in. A count net of cancellation is useful too, and the dataset reports it alongside the gross counts, but it is a different measure and it moves with every release.

Every FIGURES value that the arithmetic in this article uses is printed in Table 2, each with its key, so a reader can check any line of the later tables against its source.

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

Table 3 shows the two counts for the last two full fiscal years of 7(a) lending. Take FY2025: the 7(a) files hold 78,078 ^[15] rows for that year, and by the close of FY2026 SBA had marked 14,124 ^[43] of them cancelled with nothing disbursed, which leaves 63,954 ^[44] once those are set aside. The table subtracts one from the other and finds the stated net figure exactly, for both years.

The two counts behave differently over time, and that is the reason to keep them apart. Once a fiscal year has closed, its gross count should move little from one release to the next, since it records decisions already taken. The cancelled count can only rise as more approvals lapse, and the net count can only fall. A cancelled count quoted from an earlier release is therefore not wrong, only younger: it must sit at or below the later count for the same cohort, and a later release that showed fewer cancellations for a closed year would be a warning sign about one of the two files. The FY2025 cohort was only one to two years old at the as-of date, and 3,191 ^[45] of its 7(a) approvals had neither funded nor been cancelled, so its cancelled count can keep rising for a while yet.

What a cancellation means, and why a cancellation rate is a different animal from a charge-off rate, is the business of the series article on the two ways an approval can end unpaid. The point for this article is narrower: when a study quotes an approval count, it should say gross, and when it quotes a count net of cancellation, it should say as of which release.

The 504 Dollar Figure Is the Debenture

The 504 program splits the financing of a project into separate pieces, and the public file records only one of them. The regulation describes the permanent money for a 504 project as generally arriving in layers: the borrower's own stake, in an amount of at least 10 percent of the Project costs, and a second loan, funded by a debenture the CDC issues, covering up to 40 percent of the Project costs plus certain administrative costs and secured by a second lien ^[46]. The rest comes from a third-party loan that holds the first lien on the project property ^[47]. SBA's own program page describes CDCs as originating a 504 loan in collaboration with a senior lender ^[48].

The debenture is the piece SBA stands behind. The regulation defines it as an obligation that a CDC issues, guaranteed 100 percent by SBA, whose proceeds fund the 504 loan ^[49]. In the 504 files, the gross approval column is that debenture amount. The third-party lender's loan appears in a separate column, and the borrower's own contribution does not appear anywhere in the file. A 504 dollar total in this dataset is therefore a total of debentures. In the regulation's general structure, the debenture funds a loan "for up to 40 percent of the Project costs and certain administrative costs" ^[46], so a debenture total covers well under half of the spending it helped finance. A study that cites 504 volume as if it were project volume understates the projects; one that adds 504 debentures to 7(a) gross loans should say that it is adding a guaranteed piece of one program to the full loans of the other.

The 7(a) files work differently. Their gross approval column is the whole loan, and a separate column holds SBA's guaranteed share of it. That gives three defensible national totals for FY2025, each answering a different question. Gross 7(a) loans plus 504 debentures came to $45.09 billion ^[38]. SBA's guaranteed exposure, the guaranteed share of the 7(a) loans plus the fully guaranteed debentures, came to $35.31 billion ^[40]. The gap between them is the part of the 7(a) loans that SBA does not guarantee. Table 4 recomputes all three from their parts and checks each sum against the total the dataset states.

The guaranteed share of 7(a) gross dollars in FY2025, 73.76 percent, sits below both of the statute's general shares of 75 and 85 percent ^[3], which is consistent with some FY2025 loans carrying a guaranteed share below the general rule. The statute provides for such loans, letting a participating lender ask for a lower guarantee percentage than the general rule sets ^[50], and the figure is a small reminder that a blended share across a whole program is not the share on any one loan. A study that needs the guaranteed share of its own loan should take it from the lender's term sheet, never from a program average.

These are the national figures this article reconciles, collected with their keys and their definitions. Each one is printed from the release named in its source line.

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.

Three Ledgers for Seven Files

The first question about any download is whether everything arrived. The build answers it with three separate ledgers that are compared with one another.

The first ledger is the publisher's own index: the dataset page and its list of resources, which names seven files, the six CSVs ^[4] and the dictionary ^[6]. The second is the download ledger. Every request the build made was logged with its address, the time, the HTTP status, the Content-Length the server declared and the sha256 hash of the bytes received, and every byte count matched the length the publisher declared. The third is a fresh recount of the disk, made after the downloads finished: the files actually present, each hashed again from scratch.

All three ledgers name the same seven files, and the hash in the download ledger equals the hash recomputed on disk for every one of them. Each ledger guards against a different failure. A file the page lists but the download never fetched shows up as a gap between the first and second ledgers. A download that was cut short shows up as a byte count below the declared length. A file that was altered, replaced or partially overwritten after it arrived shows up as a hash on disk that no longer matches the hash taken at download. The build also proved that the check can fail: with one file hidden from the disk, the comparison reported the missing file instead of passing.

Rows were then counted twice by two unrelated readers, DuckDB's CSV reader and the csv module of Python's standard library, and the two agreed on every file. Two readers matter because a quoted comma, an embedded line break or a stray character in a free-text field can be handled differently by different parsers, and a reader that gets one of those wrong splits a single loan into two rows or merges two into one. Agreement between independent parsers, combined with the full-width check on every row, makes that kind of silent breakage very unlikely.

None of this proves that SBA's files contain every loan SBA approved. The three ledgers prove only that the build holds exactly what SBA posted, byte for byte. Whether what SBA posted is complete is a separate question, and it needs a separate witness.

Reading the Dictionary Against the Files

A dictionary describes what a file is supposed to contain, and the build treats it as a claim to be checked rather than as a description to be trusted. SBA provides the dictionary to help readers navigate the files ^[5], and the build reads it twice: once to name and type the fields, and once as an audit of the files themselves.

The audit compares each file's header with its program's sheet in the dictionary. Every 7(a) file holds exactly the 42 fields the 7(a) sheet lists, and every 504 file exactly the 39 fields the 504 sheet lists, with nothing missing and nothing extra. The check was itself tested with a planted fault: a copy of a header with its first field removed was fed in, and the check reported the missing AsOfDate column, which shows it does what it claims.

Field fill was measured file by file and column by column on the raw text, before any value was converted to a number or a date, so that a conversion error could not hide an empty field. The industry code is a good example of why that matters: it is only partly filled in the files before FY2001 and complete across the FY2017 to FY2026 window the loan pages use, so any industry figure drawn from the 1990s files would rest on only part of the loans.

The status column taught the most useful lesson. The build maps every status value it finds in the files to a closed list, and any value outside that list stops the run, so a new code in a future release cannot slip through as an unknown. Building that list showed that the dictionary and the file disagree: the dictionary documents the paid-in-full code for 7(a) as PIF, while the 7(a) files ship it as P I F, with spaces between the letters. The processing method column ships as spelled-out names rather than the codes the dictionary lists. Neither difference changes a count, but either would have dropped loans from a filter written from the dictionary alone. The file is the record; the dictionary is a description of it, and where they differ, the build follows the file and writes down the difference.

The status vocabulary also changes between releases, which is one more reason to pin a single release. The September 2026 file folds several earlier charge-off codes into one and withholds the status of every loan that is still open. What that withholding costs a reader, and what the earlier releases briefly showed, belongs to the companion article on the file's limits.

The Independent Witness: SBA's Year-End Activity Report

A download can be complete and still describe an incomplete record. To test the record itself, the build needs a second document that SBA produces separately and that counts the same loans another way. SBA publishes one: the Monthly and Yearly Activity Report.

SBA's page for the FY2025 year end gathers six monthly reports as of that year end ^[51]. One of them, the Monthly and Yearly Activity Report, summarizes 7(a) and 504 loan activity by month and by year, from fiscal year 2025 back to 1991 ^[52]. The reports on that page are dated as of September 30, 2025 and are published as Excel workbooks ^[53], and the page itself was issued on November 13, 2025 ^[54]. The build read the yearly sheet of that report, and of the three year-end reports before it, and set each against totals computed from the public file by program and fiscal year with no filter at all: every row, cancelled or not, exactly as the gross convention requires.

The result is the strongest single fact about the dataset. Against the FY2025 year-end report, every fiscal year from FY2010 through FY2025 ties exactly in both programs, to the loan and to the dollar. Gross dollars tie in every year in which the counts tie. Across the whole span the report covers, the 7(a) count ties in 32 of 35 fiscal years and the 504 count in 21 of 35. For 7(a), the guaranteed amount also lines up: it differs by about a hundredth of a percent at most in FY2010 to FY2024, and by under a fifth of a percent in FY2025, the one year in which the later file can still record a revised guaranty.

The comparison was tested in the same way as the other checks. With a single FY2024 loan removed from the file side, the reconciliation reported a gap of exactly one loan for that program and year, which shows the comparison is sharp enough to see one missing row among tens of thousands.

It is fair to ask how independent the witness is. It is a different product, an aggregate workbook rather than a loan-level file; it was produced a year earlier, against the FY2025 year end rather than the FY2026 one; and it is published on a separate page. It is not independent in the strongest sense, because both documents come out of SBA's own systems, and a loan that never entered those systems would be missing from both. What the witness proves is that the public loan file is the same record SBA itself reports to the public, row for row and dollar for dollar, for every year from FY2010 to FY2025. It does not prove that SBA's records are free of error, and the dataset never claims that.

The witness also has a blind spot that defines the limit of what it can prove. It counts rows and adds dollars; it cannot see the values inside a row. A loan coded to the wrong state, the wrong industry or the wrong status still ties in the yearly totals. The tie therefore proves rows and dollars, not coding, and the coding checks in this article and its companions are a separate line of defence.

Where the Witness Disagrees, and What That Means

The years before FY2010 do not tie as neatly. In the 504 program, the public file runs short of the year-end report in all but one year from FY1991 to FY2004, by as much as a few dozen loans a year, and holds one loan more than the report for FY2009. In 7(a), three years of the 1990s show the file one or two loans over the report. These gaps are small against the size of the programs, and none touches the FY2017 to FY2026 window that the loan pages use, but they are reported rather than smoothed over.

The more instructive finding is that the witness revises itself. The build read four year-end reports, not one, and compared them with each other. Of the 38 program-years before FY2010 that appear in two or more of those reports, 25 change from one report to the next. Two later program-years change as well: 504 FY2011 and 7(a) FY2020. A historical count for a year long closed is therefore not a fixed quantity even inside SBA's own reporting, and a difference of a few loans between two published sources for the 1990s is not evidence that either one is wrong. For a study, the practical rule is to rest every historical comparison on years from FY2010 onward, where the file and the latest report agree exactly, and to describe any older figure as approximate.

FY2026 has no witness yet. SBA also maintains a current-month version of the same reports: its page links to six reports that are updated monthly ^[55], among them a Monthly and Yearly Activity Report that runs from the current fiscal year back to 1991 ^[56]. The workbook links on that page lead to a separate SBA lender-portal domain rather than to data.sba.gov, and the build reads only from federal primary hosts, so the current-month workbooks were not fetched. Until SBA posts a year-end page for FY2026, as it did for FY2025, the FY2026 counts in the dataset rest on the public file alone. A study that cites FY2026 should say so, in addition to noting that the year is short.

Reproducibility: Hashes, Two Rebuilds and a Frozen Release

A dataset that gives one answer today and a slightly different answer tomorrow cannot anchor a loan file, so the build was designed to be rerun and was rerun.

The first layer is hashing. The sha256 hash of every raw file was recorded at download, and every national figure in the dataset carries, beside its value, its definition, the exact query that produced it, its data vintage and the hashes of the six source files it was computed from. A reader who downloads the same release from SBA can confirm that the bytes match before checking a single number.

The second layer is the rebuild. The whole pipeline, from the raw files to the final tables, was run twice from the raw bytes, and the sha256 hash of every output was compared between the two runs: the cell table, the threshold table, the national figures, every manifest, and every cell file taken together. All 19 hashes were identical. The same test was run again after the last changes to the asset map, with the same result.

Getting there exposed four sources of instability, and each was fixed rather than tolerated. A sum of dollar amounts computed in parallel with floating-point numbers landed a cent apart between runs, because the order of additions changed; dollar sums now use exact decimal arithmetic. Two lenders sharing a display name, ties in ranked lists, and ties in the listed composition of one asset class could each come out in a different order between runs; every one of those now breaks ties by a fixed rule. Each was small, a cent here or the order of tied items there, but each would have made a rerun look different from the original, which is the one thing a reproducible build cannot allow.

The third layer is the frozen release. The dataset is tagged, and every figure in this series is read from one tag, analytics-loan-data freeze-3, at one commit. The source line for each figure names that release, so a figure in a study can be traced to the tag, the definition and the query that made it. Two reviewers were also set to argue that the dataset was wrong, each recomputing the gate counts, the witness, the completeness hashes and the reproducibility hashes on their own. Several of their findings were upheld and led to changes, among them the rule that every charge-off rate is labelled a floor, and every finding and its disposition is recorded with the release; none of them broke a count, a tie or a hash.

The simplest arithmetic check is open to any reader. Table 6 adds the twenty yearly approval counts, ten fiscal years in each program, and compares the sum with the total the dataset uses for its state and asset class cells. The two agree exactly, so nothing in the window was dropped or counted twice between the yearly figures and the cells.

What a Feasibility Study Can Rely On, and What It Cannot

The checks above draw a fairly sharp line around what the dataset can bear.

A study can rely on gross approval counts and gross dollars by program and fiscal year for FY2010 through FY2025. Those figures are the same figures SBA reports in its own year-end activity report, they rest on files that were proven complete against the publisher's index, and they reproduce exactly on a rebuild. A study can also rely on the definitions behind them: approvals counted gross with cancellations included, 7(a) dollars as whole loans with the guaranteed share held separately, and 504 dollars as debentures. And it can rely on the as-of date, which is written into every row and fixes the status of every loan at the close of FY2026.

A study should treat other figures with more care. FY2026 is short and has no witness. Years before FY2010 differ slightly from SBA's own reports, which themselves change from one edition to the next. Field values that the witness cannot see, such as the state, the industry code and the status of a loan, were checked for width, fill and vocabulary but not against an outside count, and the companion articles explain how far each can be trusted. Rates of any kind are built on approval cohorts and statuses that are still changing, and the series treats every one of them as a floor. Why each one can only understate is argued in the series piece on approval cohorts and seasoning, and a separate piece explains why the dataset leaves a rate blank whenever its denominator is too thin to bear it.

A study must not infer anything about a single project from these totals. A national count of approvals in a class says how much SBA lending a kind of business has drawn, and where; it says nothing about whether one proposed project's market, costs and cash flow will carry its debt. That remains the work of the study itself, built from evidence about the site, the trade area and the operation. The loan record supplies context and a benchmark, and a careful study keeps the two roles distinct in the way it writes them up.

The same discipline applies to the release. Every quarterly release replaces the last ^[9], and the next one will move cancelled counts, statuses and rates for recent cohorts even as the gross counts for closed years hold still. A study written today should cite the September 30, 2026 release by name. A study revised next year should either keep that citation or move every figure to the new release at once, never mixing the two.

What a Lender or Borrower Should Check

A short list of questions will expose most problems with a loan figure in a study, whether the figure comes from this dataset or from somewhere else.

Does the figure name its release? SBA refreshes its files each quarter ^[9], so a count or a rate that gives no release date leaves the reader nothing to hold it against. The source line should name the release and, for this dataset, the key.

Is the year fiscal or calendar? SBA's figures run on the federal fiscal year, from October 1 to September 30 ^[12]. A calendar label on a fiscal count is an error even when the number is right.

Is the count gross or net? Gross approvals include later cancellations; net counts do not, and they fall with each release. Either is legitimate, but the study must say which, and must not compare a gross count from one source with a net count from another.

What do the dollars measure? For 7(a), the whole loan or SBA's guaranteed share of it, which the general rule sets at 75 or 85 percent of the balance at disbursement depending on loan size ^[3]. For 504, the debenture alone, a piece the regulation generally limits to 40 percent of the project's cost before certain administrative costs ^[46], and never the whole project.

Is the year witnessed? FY2010 through FY2025 tie exactly to SBA's year-end report. Older years are approximate, and FY2026 is short and not yet witnessed.

Is a rate labelled as a floor, with its cohort and its loan count? If not, ask for them, and read the series' articles on cohorts, small samples and cancellation before relying on the rate.

Can the figure be recomputed? Every figure from this dataset carries a FIGURES id or a cell key, and every key resolves to a definition and a query in the frozen release. A figure that cannot be traced that way should be treated as unverified, however authoritative the document that prints it.

None of these questions requires the reader to rebuild the data. Each of them requires only that the study say what it is quoting, from which release, under which definition, which is the standard the Loan Analytics loan dataset is built to meet for every figure it supplies.

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
• From NAICS Code to Asset Class: How SBA Loans Are Assigned, and Where Two Industries Share One Code
• What the SBA Loan File Cannot Tell You: The Limits of FOIA Data in Underwriting

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