Skip to content

AI uses for records

What 3,648 CBO revisions teach a variance analysis evidence dataset

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A variance analysis evidence dataset should accept an explanation only if it can be rebuilt from records: a dated baseline, every revision split by cause, decisions after the plan shown separately and a named remainder. CBO's public projection data shows the test working. Of 3,648 revisions, 75.0 percent close from their own rows, 728 more only through components, and 5.1 percent cannot be rebuilt from the file.

License terms

Variance analysis evidence

This page covers what 3,648 CBO revisions teach a variance analysis evidence dataset.Illustrative

Key takeaways

  • Treat a variance explanation as evidence only when its pieces add back to the gap and each piece points to an input record.
  • In CBO's public file, 2,735 of 3,648 outlay revisions close from their own rows and 185 cannot be rebuilt, 84 of them at two 2023 baselines.
  • Removing decisions made after the plan reversed the sign of the outlay gap in 6 of 42 years and the revenue gap in 2 of 44.
  • Closure proves arithmetic, not cause: a catch-all category can close any bridge, so each amount must trace to its own input.
  • Grade baseline, bridge, decisions, inputs, category rule, remainder and scope separately, and record who signed the verdict.

What makes a variance explanation reconstructable?

A variance explanation is reconstructable when someone who did not write it can rebuild it from records: the baseline it compares against, each revision since that baseline with its cause and amount, the decisions made after the plan, and an explicit remainder. If the pieces add back to the gap within a stated tolerance and each piece points to an input record, the explanation is evidence. If they do not, it is commentary.

That is the standard a variance analysis evidence dataset should hold every example to. Month-end commentary usually names a cause, such as volume, price or a delayed project, without the bridge that shows how much of the gap each cause carries or the operating record behind it, such as units shipped, hours billed, headcount or a rate change. A model trained on that commentary learns its tone, not its arithmetic.

Public tests of this standard are rare, because companies do not publish their budget revisions. One public body publishes its revisions in machine-readable form, split by cause. The rest of this page audits that record, then turns the result into a casebook and a reviewer rubric for company data.

Why CBO's own projection revisions make a fair public test

The Congressional Budget Office publishes the data behind its evaluations of its own projections in a public repository, US-CBO/eval-projections, which is in the public domain in the United States. Version 0.12.0 carries baselines through February 2026 and actuals through fiscal 2025. CBO's budget and economic outlook reports split the change between one baseline and the next into three causes, and the repository's change file records them as legislative, economic and technical, by outlay category and fiscal year.

The categories map onto a company budget better than the setting suggests. Legislative changes are decisions made after the projection, much like an approved budget transfer, a price change or a new program. Economic changes come from a revised outside forecast, like a new volume, rate or exchange-rate assumption. Technical changes, in CBO's 2026 to 2036 outlook, cover what is neither legislative nor economic, including new agency data, administrative actions, court rulings and better information on how laws are carried out: the estimating revisions a finance team makes when actuals arrive.

The file also documents its own scope, which turns out to matter. Its README says Fannie Mae and Freddie Mac outlays are removed from the projections and the actuals, and that the effects of the 2022 student loan cancellation plan and the court decision blocking it are excluded. Any reconstruction has to state the same kind of scope rules, or two correct numbers will disagree.

Bridge audit: 3,648 revisions and how many close from their own records

Three quarters of the revisions close from their own rows, most of the rest close once the components are added, and about one in twenty cannot be rebuilt from this file. The audit took total outlays and its three components (mandatory, discretionary and net interest), paired each baseline with the previous baseline for the same fiscal year, and asked whether the recorded legislative, economic and technical changes add up to the difference within $5 million.

Of 3,648 such bridges, 2,735 (75.0 percent) close from the rows recorded at their own level. Another 728, all at the total-outlays level, close only when the missing technical change is rebuilt from the three components; the file carries no total-level technical row at any baseline before January 2019. That leaves 185 bridges (5.1 percent) with a gap: 122 where rows exist but leave a remainder, and 63 with no change rows at all.

The gaps cluster. The February 2023 baseline has only legislative rows in the file, and the May 2023 baseline has none, which together account for 84 of the 185. CBO's February 2023 outlook includes an appendix on changes in its baseline projections since May 2022, so the gap is in this file, not necessarily in CBO's reporting. Most other gaps sit in the earliest years: 46 of the 85 total-level remainders end at baselines before August 1992, the date of the first technical change the file records for any outlay series. That is the point for a dataset builder: an explanation that exists in a report but not in the records attached to the case cannot be reconstructed from the case.

The counts by series are below:

Bridge audit: 3,648 revisions and how many close from their own records
SeriesBridgesClose from own rowsClose only with component rowsRemainder or no record
Total outlays94211772885 remainder, 12 no record
Mandatory outlays902873Not applicable12 remainder, 17 no record
Discretionary outlays902870Not applicable15 remainder, 17 no record
Net interest902875Not applicable10 remainder, 17 no record
All four series3,6482,735728122 remainder, 63 no record

FY2025 outlays traced from the first projection to the actual

Following one fiscal year end to end shows what the audit means in money. CBO's first projection of fiscal 2025 total outlays, in January 2015, was $6,115.1 billion; the actual, in the file, was $7,016.4 billion. Twenty-three revisions sit between the first projection and the last one in January 2025, and the file lets most of the $901.3 billion difference be assigned.

Legislation recorded across those revisions added $735.4 billion, economic changes subtracted $272.9 billion, and technical changes added $195.0 billion, a figure that includes a net $19.6 billion reduction rebuilt from the components for revisions before 2019. That reaches the January 2025 projection of $7,034.7 billion only after a further $262.2 billion that the file does not attribute, nearly all of it from the two 2023 baselines. After January 2025, legislation took $108.4 billion off the projection, and the remaining miss was $90.1 billion: actual outlays came in higher than the projection adjusted for that legislation.

CBO's own accuracy report for fiscal 2025 puts the January 2025 outlay underestimate at $102 billion, or 1 percent. The repository's figure differs, and its README names at least one scope difference, the removal of Fannie Mae and Freddie Mac outlays. Neither number is wrong. They answer slightly different questions, which is why scope belongs in the record of every variance case.

Diverging bars showing how CBO's projection of fiscal 2025 outlays moved from 6,115 billion dollars in January 2015 to the 7,016 billion dollar actual: legislation plus 735, economic minus 273, technical plus 195, not attributed in the file plus 262, later legislation minus 108 and a remaining miss of plus 90.
Fiscal 2025 total outlays, billions of dollars, from US-CBO/eval-projections version 0.12.0; category sums by SourceX's bridge audit.

Decisions made after the plan can reverse a variance's sign

Separating what was decided after the plan from what was mis-estimated can change the sign of a variance, not only its size. Because laws passed after a projection change the outcome it was trying to predict, CBO's evaluations compare each projection with the actual after adding the effects of legislation enacted later; the repository's error files carry that adjusted projection, and the audit's recomputation matched it exactly for every year.

Across 42 first-year outlay projections, for fiscal 1984 through 2025, the adjustment reversed the sign of the gap in 6 years (1989, 2002, 2003, 2006, 2007 and 2025), and later legislation was larger than the raw gap itself in 8. Across 44 first-year revenue projections, for fiscal 1982 through 2025, it reversed the sign in 2 years, 1992 and 2020. The extreme case is fiscal 2020 outlays: the March 2020 projection was $1,848.1 billion below the actual, but $1,778.8 billion of that came from legislation enacted after it, leaving an estimating miss of $69.3 billion.

Fiscal 2025 revenue shows the opposite trap. The raw gap was $71.7 billion too low, which looks modest; after removing legislation that reduced revenue by $131.4 billion, the projection was $203.1 billion too low, matching the $203 billion, or 4 percent, in CBO's accuracy report. CBO attributes about half of that miss to tariff increases the Administration put in place by administrative action, which its framework treats as technical rather than legislative. The category follows the instrument, not the economic effect, and a company dataset needs the same kind of written rule: a price increase approved by the board and one forced by a supplier land in different categories.

Casebook: seven bridge patterns and the records that settle each

Every pattern in the CBO file has a counterpart in company variance work, and each is settled by a specific record. The casebook below pairs them; the verdict column is what a reviewer should record for a case showing that pattern. Amounts refer to total outlays in the CBO file at version 0.12.0.

Casebook: seven bridge patterns and the records that settle each
Pattern in the CBO fileCompany counterpartRecords that settle itReviewer verdict
Revision closes from its own rows (2,735 bridges)Reforecast bridge whose drivers sum to the changeBoth budget versions with dates; bridge entries by driver; ledger detailReconstructed
Total-level technical row missing, components carry it (728 bridges)Group commentary with an other line; entities explain itEntity-level bridges that sum to the group figureReconstructed with derivation; record the rule
Small remainder after all rows ($1.3 billion and $1.6 billion in 2021 revisions to fiscal 2025)Unexplained residual under the materiality thresholdThreshold policy dated before the period; residual shown, not pluggedPartly reconstructed; remainder named and sized
Only one cause recorded (February 2023)Commentary names the approved changes and nothing elseThe missing driver entries or the reforecast workpaperNot reconstructable until supplied
Revision with no record (May 2023)Budget silently replaced mid-yearVersion history of the budget file and its approvalNot reconstructable
Legislation after the projection (minus $108.4 billion for fiscal 2025)Budget transfer or price change approved after the planApproval record with date and amountSeparate from the estimating miss before grading
Report and file disagree ($102 billion against $90.1 billion)Board pack and ledger give different variancesWritten scope rules for each figureReconcile scope first, then grade

Reviewer rubric: grading an explanation claim by claim

Grade an explanation one claim at a time, and grade arithmetic, attribution and evidence separately so a fluent answer cannot pass on style. The checklist below works for a model's explanation and for a human's commentary alike; the final verdict uses the same four values as the casebook.

At a glance

Reviewer rubric: grading an explanation claim by claim

  • Baseline: the claim names the budget or forecast version it compares against, with its date, and that version is in the packet.
  • Bridge: every revision between that baseline and the actual is listed, and the amounts by cause sum to the total change within a tolerance written before the period.
  • Decisions after the plan: approved changes made after the baseline are shown separately and removed before the remainder is called an estimating miss.
  • Inputs: each driver links to an operating record (units, hours, headcount, rates, contract terms) and to the ledger lines it moved.
  • Category rule: the reason a change sits in one category rather than another is written down, so a supplier price rise and a board decision are not both called price.
  • Remainder: whatever is unexplained is shown as a named line with its amount, never folded into a driver.
  • Scope: exclusions and entity boundaries are stated, so a second figure for the same variance can be reconciled.
  • Verdict and sign-off: reconstructed, reconstructed with derivation, partly reconstructed or not reconstructable, recorded with the reviewer's accounting qualification and date.

Does a bridge that closes prove the explanation is right?

No. Closure proves the arithmetic, not the explanation. Any bridge can be made to close by putting the difference in a catch-all line, and in CBO's framework the technical category is defined as what is neither legislative nor economic, so it absorbs whatever the other two do not. A bridge that closes because the last category is a remainder is reconstructed only in name.

The test that separates a real attribution from a plug is whether each category's amount traces to its own input. In the CBO file an economic change should follow from a dated revision to the economic forecast, and a legislative change from a cost estimate for an enacted law. In a company, a volume driver should follow from unit counts in the operating system, and a price driver from invoice or contract terms. Where the amount cannot be traced, the category is an opinion.

Two further questions sit outside arithmetic altogether: whether the accounting treatment behind the actual is right, and whether the spending or the budget change was authorized. A bridge can close on a misclassified accrual or an unapproved overspend. Those need a qualified accountant's review and the approval records, and a dataset should label them in fields of their own rather than inside the reconstruction verdict.

Assembling company variance packets that pass the bridge test

A company packet passes the bridge test when it carries the same layers as the CBO file, plus the operating detail CBO's file leaves out. In practice that means the original budget and every reforecast as dated versions, a bridge between each consecutive pair split into decisions, outside drivers, estimating changes and remainder, the operating counts behind each driver, ledger detail for the period, approval records for budget transfers, and the commentary as finally issued.

The audit suggests an acceptance test a buyer can run on delivery: the share of bridges that close from their own rows, the share that close only by derivation, and the share with gaps, each with its rule. A packet set in which gaps cluster in particular periods, like the 2023 baselines here, tells the buyer where records went missing before any model is trained on the commentary.

Packets need authorization and de-identification first; stable pseudonyms for vendors, customers and employees keep ledger totals reconciling after names are removed. Teams that need such packets can describe them to SourceX, which looks for datasets at U.S. businesses after a request arrives instead of holding an inventory; the request covers scope, volume, language, format, timing and licensing terms, and can state the bridge-closure test as the acceptance condition. As a rough guide, buyers look for two or more years of consistent history and tens of thousands of records in one system, though rarer work can qualify with less, judged deal by deal.

Counting rules, scope and limits of the audit

The audit reads three files from the repository at commit 682559c (version 0.12.0) and runs no repository code. A bridge is a pair of consecutive baselines for the same series and fiscal year; it closes when the change rows dated after the earlier baseline and on or before the later one sum to the difference within $0.005 billion, using exact decimal arithmetic on the values as written. That window picks up 65 change dates that have no row of their own in the baselines file, summed into the 40 bridges that span them; 14 more predate a fiscal year's first baseline and fall in no bridge. For total outlays, a bridge without a technical row closes by derivation when the remainder equals the sum of the three components' technical rows in the same window, within the same tolerance. One change row with an empty value, for net interest, was skipped and is printed by the script. Change dates after a fiscal year's last baseline, 142 series, year and date combinations, are not bridges.

The decision adjustment uses CBO's spring baselines for outlays and winter baselines for revenues, first projection year only, adding every legislative change dated after the baseline. A sign reversal means the raw and adjusted gaps have opposite signs. All figures are in billions of dollars as the file records them; percentages are rounded once, from exact counts.

The limits are real. The file is an evaluation input, not the whole of CBO's published analysis, so its gaps say nothing about CBO's reports. Government projections are not company budgets: CBO publishes its revisions by cause, while companies change plans inside the year and rarely record why. The audit shows how to test reconstructability, not how often company commentary passes; that needs authorized company packets labeled by qualified accountants. The scripts, cbo_bridge_audit.py and its saved output, contain no randomness.

Frequently asked questions

What is a variance bridge, and why must it close?

A variance bridge lists every change between a baseline and an actual result, split by cause, so the pieces sum to the total difference. If the pieces do not sum within a stated tolerance, part of the variance is unexplained, and an explanation built on that bridge is incomplete however fluent it sounds.

Why separate decisions made after the budget from forecast errors?

Because they can point in opposite directions. In CBO's data, removing legislation enacted after the projection reversed the sign of the outlay gap in 6 of 42 years. In a company, an approved budget transfer can make a real estimating miss look like an overrun or hide it entirely.

What should each case in a variance analysis evidence dataset contain?

The dated baseline, every revision with amounts by cause, the operating records behind each driver, ledger detail, approvals for changes made after the plan, the issued commentary and a named remainder. A reviewer verdict and the scope rules complete the case.

How large an unexplained remainder is acceptable?

Set the threshold before the period and record it with the case, then show any remainder as its own line. In the CBO audit, bridges that closed did so within $5 million, while some 2021 revisions left remainders above $1 billion at the total level. What matters is that the remainder is named and sized rather than folded into a driver.

Can public government data stand in for company variance records?

It can test a method, not replace company data. CBO publishes its revisions by cause, which makes it a clean public example of a reconstructable explanation. Company budgets change inside the year, use different drivers and rarely record why, so real training and evaluation sets need authorized company packets reviewed by accountants.

Sources

  • CBO's repository documents its input files, the legislative, economic and technical split of baseline changes, and scope rules removing Fannie Mae and Freddie Mac outlays and excluding the 2022 student loan cancellation effects. Source
  • The baselines, baseline changes and actuals files audited here, read at commit 682559c. Source
  • Version 0.12.0 (2026-02-12) carries baselines and baseline changes through the February 2026 baseline and actuals through 2025. Source
  • CBO's outlay error file carries the legislative change and the adjusted projection used to compare each projection with the actual. Source
  • The repository's data, code and documentation were created by CBO employees and are in the public domain in the United States. Source
  • CBO's 2026 to 2036 outlook describes technical changes as revisions that are neither legislative nor economic, citing new agency data, administrative actions, judicial rulings and implementation information, and treats tariff increases made by administrative action as technical. Source
  • CBO's fiscal 2025 accuracy report finds the January 2025 projection too low by $203 billion (4 percent) for revenues and $102 billion (1 percent) for outlays, with tariff increases made by administrative action about half of the revenue miss. Source
  • CBO's February 2023 outlook includes an appendix on changes in its baseline projections since May 2022. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies