Systems and records
What evidence should back each explanation in a budget variance dataset?
By SourceX Editorial · Updated
9 sections
Short answer
A budget variance explanation belongs in training data only when linked records prove it: the budget line and every revision to it, the posted actuals, the source documents behind the gap, the person who approved the spending, and a later outcome that confirms the cause. Keep the plausible explanations reviewers rejected as counterexamples, so a model learns to cite evidence instead of writing a convincing story.
Value drivers
Budget variance explanation
Key takeaways
- A variance explanation is a claim about cause; without linked documents it is a label nobody can check.
- Keep the budget lineage: original amount, transfers and revisions, each with its approval, so the model compares actuals to the right baseline.
- Rejected explanations are as valuable as accepted ones, because they teach a model which plausible stories the records disprove.
- Grade answers on whether each cause reconciles to documents, authority and the final ledger state, not on how fluent the commentary reads.
- Offsetting causes inside one line item must be split and quantified separately, or the packet teaches a model to stop at the first story that fits.
What evidence should support a budget variance explanation in training data?
Each explanation needs a chain of records a reviewer can follow from the budget line to the cause: the approved budget and its revisions, the posted actuals for the same accounts and period, the transactions or documents that produced the gap, the approval that authorized the spending, and a later record that confirms the cause held up. If any link is missing, the explanation is an opinion, and a model trained on it learns to produce opinions.
Most finance teams already write variance commentary every month. The commentary is rarely the problem; the problem is that it is stored as free text in a slide or a spreadsheet tab, detached from the journal lines and invoices that would prove it. A budget variance explanation dataset reattaches those records, so every sentence of commentary points to something a reviewer can open.
The second half of the answer is less obvious. Keep the explanations that were proposed and then rejected, along with the record that disproved each one. Those counterexamples are what teach a model to check a story against the ledger instead of accepting the first cause that sounds right.
Why is a plausible variance story a bad training label?
Because fluent commentary and correct commentary look identical in text, and only the linked records separate them. Auditors learned this long ago. The PCAOB standard on substantive analytical procedures, AS 2305, tells auditors to evaluate significant unexpected differences and says management responses should ordinarily be corroborated with other evidential matter. An explanation from the person closest to the numbers is the start of the evidence, not the end of it.
Public company reporting makes the same demand in a different setting. In its 2020 amendments to the MD&A rules, the SEC clarified that Item 303(b) requires the underlying reasons for material changes in quantitative and qualitative terms, including where material changes within a line item offset one another. A label that says costs rose because of inflation, with no amount attached to inflation and nothing said about the offsetting savings, fails that standard and should fail yours.
For AI work the cost of a weak label compounds. A model fine-tuned on unchecked commentary learns the house style of variance explanations: the usual suspects, the confident tone, the round attributions. It does not learn to find the cause. An evaluation set built from the same commentary will then reward the model for matching the style, and the problem stays hidden until someone acts on a wrong explanation.
Which records tie a variance to its cause?
Start with the budget lineage, because a variance is only as good as its baseline. ERP budgeting modules usually record more than one budget number per line. In Microsoft Dynamics 365 Finance, for example, budget codes distinguish original budget, transfer and revision entries, and Microsoft describes them as providing an audit trail of approved budget modifications through the budget cycle; a workflow attached to a budget code must be completed before the entry is final. If your export keeps only the latest budget, a model cannot tell a real overrun from a line whose budget was cut mid-quarter.
Then keep the actuals at journal line level with their source documents: purchase orders, receipts, vendor invoices, payroll registers, contracts and the approval records behind them. Keep reversals and reclassifications as their own entries rather than netting them away. Dynamics 365 Finance, like most ledgers, can reverse individual journal lines or an entire journal, and those reversals are often the real story behind a variance that disappears the following month.
Finally, keep whatever recorded the explanation itself: the variance commentary, the reviewer who signed off, comments in the close checklist, tickets in Jira or emails that asked a budget owner why. These records show which causes were proposed, who accepted them and when, which is the raw material for the rejected-explanation set.
The variance task packet: fields every case should carry
Use one packet per variance case: one budget line, one period, one material gap. The schema below is the asset. A team can copy it into a spreadsheet or a JSON schema as is, and each field names the record that fills it, so packets can be assembled from an ERP export rather than written from memory.
- Two further fields complete the packet: a counterexample flag for cases where the first explanation was wrong, and a de-identification note recording which vendor, employee or customer names were replaced with stable pseudonyms.
- Keep amounts in the ledger currency with the unit stated, and record the materiality threshold the team used, because a cause that explains a small share of the gap is not an explanation of the gap.
| Field | What it holds | Record that fills it | Check before accepting |
|---|---|---|---|
| Starting state | Entity, cost center, account, period and the variance as amount and percentage of budget | Budget vs actual report at period close | Recomputes from the budget and actual fields below |
| Budget lineage | Original budget, each transfer and revision with date and approver | Budget register entries and their workflow history | Baseline used matches the version approved before the period |
| Actuals | Journal lines for the account and period, including reversals and reclassifications | General ledger detail export | Lines sum to the reported actual |
| Source documents | Invoices, receipts, payroll or contract documents behind the lines that drive the gap | AP, procurement, payroll or contract systems | Each driver line has at least one document |
| Applicable policy | Spending, approval or accrual rule in force for the period | Policy document version, delegation of authority matrix | Version dated on or before the transactions |
| Authority | Who approved the spending and the budget change, and their limit | Approval workflow log, PO approvals | Approver held authority at the time |
| Accepted cause | Each cause with its quantified share of the variance and the records that support it | Reviewer sign-off, close commentary | Shares sum to the total variance within a stated tolerance |
| Rejected causes | Plausible explanations that were proposed and disproved, each with the disproving record | Commentary drafts, review comments, emails | Disproving record is linked, not described |
| Outcome check | Later evidence that the cause held: the next period, a credit note, a contract amendment | Following period ledger, vendor credits, renewals | Dated after the period close |
| Label status | Real and approved for use, or illustrative, plus the reviewer and review date | Dataset review log | Illustrative packets are marked and excluded from evaluation scoring |
Which rejected explanations should the task set keep?
Keep every proposed cause that a reviewer disproved, and tag it with the pattern it represents. Six common patterns are listed below. Each has a specific record that disproves it, which is what turns a wrong answer into a training signal rather than noise.
| Rejected pattern | Typical wording | Record that disproves it | What the model should learn |
|---|---|---|---|
| Macro story | Costs rose because of inflation or market prices | Unit prices on invoices unchanged against prior period | Test price and volume separately before blaming prices |
| Wrong baseline | We overspent the budget | Budget revision that cut the line mid-period | Compare to the approved baseline in force, not the latest number |
| Timing mistaken for cause | The vendor raised rates | Invoice for the prior period accrued late, then reversed | Check accrual and reversal entries before naming a cause |
| Misclassification | Marketing spend increased | Journal lines coded to the wrong account and reclassified after close | Look for reclassification entries that move the gap |
| Single cause for a net figure | Volume was higher | Higher volume offset by a supplier credit, both material | Split offsetting drivers and quantify each one |
| Unauthorized framing | The project needed it | No approval in the workflow log, or an approver above their limit | An explanation that skips authority is incomplete |
Illustrative worked case: a freight overrun with three candidate causes
The case below is illustrative, not drawn from a real company, and uses counts and percentages only. It shows how one packet moves from starting state to verified outcome, and where the rejected explanations come from.
- Starting state: a distributor's outbound freight account closes the second quarter 18% over its approved budget. The original budget was not revised during the quarter, so the baseline is clean.
- Candidate cause 1, fuel surcharges: the operations manager's first commentary blames fuel. The packet links the carrier invoices, which show the fuel surcharge rate unchanged from the first quarter. Rejected, pattern: macro story.
- Candidate cause 2, volume: shipment counts rose 6% against the prior quarter. Accepted as partial: it explains roughly a third of the gap once the unit cost per shipment is held constant, and the packet records that share.
- Candidate cause 3, expedited shipping after a carrier change: the ledger shows 212 expedited shipments in the quarter against 40 in the first quarter, starting the week a contract with the previous carrier ended. The packet links the contract end notice, the new carrier's rate sheet and the approval chain for expedite requests, which shows a regional manager approved them under a temporary exception. Accepted, with its quantified share.
- Residual: a small remainder traces to one invoice for the first quarter accrued late. The packet keeps the accrual and its reversal and labels the remainder as timing, not cause.
- Outcome check: third-quarter actuals return within 3% of budget after the exception expires, and the expedited count drops back to the first-quarter level. The cause held.
- Counterexample flag: set, because the first explanation was wrong. This packet appears in the evaluation set as a case where a model must reject the fuel story using the invoices.
How should a model's variance explanation be graded?
Grade the explanation against the packet, not against the original commentary text. Agent benchmarks have moved this way for a reason: tau-bench, a published agent benchmark, scores an agent by comparing the database state at the end of an episode with an annotated goal state, rather than matching the wording of a reference answer. For variance work the equivalent goal state is the set of causes, their shares and the records that support each.
At a glance
How should a model's variance explanation be graded
Citation checking is the other half. Research on attributed question answering treats human judgments of whether a cited source supports an answer as the gold standard, and found the automatic AutoAIS metric useful for development but only moderately correlated with human ratings on individual answers. Plan for a finance reviewer to grade a sample of answers by hand, and use automated support checks as a screen, not the verdict.
- Cause correctness: every accepted cause in the packet is named, and no rejected cause is offered as an explanation.
- Quantification: each cause carries a share of the variance that reconciles to the total within the packet's tolerance.
- Evidence: each cause cites at least one linked record, and the cited record actually supports it.
- Authority: the answer notes whether the spending and any budget change were approved by someone with authority at the time.
- Abstention: when the packet lacks the record needed to decide, the correct answer says which record is missing instead of guessing.
What breaks a variance dataset before a model ever sees it?
The first failure to check for is an export that flattens history. Keeping only the current budget, netting reversals out of the actuals, or exporting commentary without the period and reviewer it belongs to removes exactly the information a packet needs. Ask for line-level ledger detail with entry dates, budget entries with their codes and approval history, and commentary with timestamps.
The second failure is a chart of accounts that changed mid-history. If cost centers were merged or accounts renumbered, record the mapping as its own versioned table and attach the version to every packet, or a model will learn variances that are artifacts of the remapping.
The third is de-identification that breaks the links. Replace vendor, employee and customer names with stable pseudonyms that stay the same across invoices, journal lines and approvals, so the chain of evidence survives. De-identification makes records safer to share; it does not by itself give a company the right to license them, which depends on contracts and permissions reviewed separately.
What a supplier of variance packets should prepare first
A company considering licensing its finance records need not build packets before learning whether it qualifies; SourceX's first step is a short company assessment that requires no uploaded data. What helps that conversation is knowing which systems hold budgets, ledger detail, approvals and commentary, and for how many years they are consistent.
As a general guide, AI buyers look for years of consistent history, typically two or more, and tens of thousands of records or more in one system, though smaller sets can qualify when the work is rare, and each deal is assessed on its own. Rights review, contracting and data preparation are SourceX's part, and records are delivered only after the company has approved the scope and terms and signed.
Frequently asked questions
What is a budget variance explanation dataset?
It is a set of variance cases in which each explanation is linked to the records that prove it: the budget lineage, ledger lines, source documents, approvals and a later outcome check. It usually includes rejected explanations as counterexamples. It is used to train or evaluate AI systems that analyze budget vs actual results.
Is month-end variance commentary enough on its own?
No. Commentary is a claim about cause written by someone close to the numbers, and auditing standards treat such explanations as something to corroborate. Without the linked ledger lines and documents, a model learns the style of commentary rather than how to find the cause.
Why keep explanations that turned out to be wrong?
Rejected explanations, each tied to the record that disproved it, teach a model to test a plausible story against the ledger. They also make an evaluation set harder to pass by sounding confident. Tag each one with its pattern, such as wrong baseline or timing mistaken for cause.
How do you handle variances with several offsetting causes?
Split them and quantify each driver separately, including drivers that reduce the variance. The SEC's MD&A rules call out offsetting changes within a line item for the same reason: a net figure explained by one cause hides material movements in both directions.
Can illustrative cases be mixed with real ones?
They can be used for training examples or documentation, but label them clearly and keep them out of evaluation scoring. An evaluation set should measure performance on records that actually happened and were reviewed.
Sources
- PCAOB AS 2305 says auditors should evaluate significant unexpected differences and that management responses should ordinarily be corroborated with other evidential matter. Source
- The SEC's 2020 MD&A amendments clarify that Item 303(b) requires the underlying reasons for material changes in quantitative and qualitative terms, including where changes within a line item offset one another. Source
- Dynamics 365 Finance budget codes distinguish original budget, transfer and revision entries and provide an audit trail of approved budget modifications; workflows can be attached to budget codes. Source
- Dynamics 365 Finance can reverse individual journal lines or an entire financial journal. Source
- tau-bench scores agents by comparing the final database state with an annotated goal state rather than matching reference wording. Source
- Attributed QA research uses human attribution judgments as the gold standard and finds AutoAIS only moderately correlated with human ratings at the instance level. Source
See if your company qualifies
A short company assessment. No data uploads are needed.