| Takeaway | Detail |
|---|---|
| Clear a Benford flag only when the first-digit MAD is below 0.006 | Recalculate the first-digit mean absolute deviation using the filing's own numeric XBRL facts, and clear the population only when MAD is below 0.006 (chi-square ≤ 15.507 with 8 degrees of freedom). |
| Re-check the Benford MAD threshold for each filing size | Benford first-digit screening is a population-level triage tool for 2023–25 SEC restatement filings, so its MAD threshold needs to be re-verified against the size of each filing's numeric population. |
| Use supervised models only when you have labeled Big R / little r history | Supervised models are worth the effort only when you have labeled restatement history that distinguishes non-reliance (Big R) restatements from revisions (little r). |
| Keep Big R and little r restatements separate in your training labels | Audit Analytics restatement data covers both non-reliance (Big R) and revision (little r) disclosures from SEC registrants, and those need to be labeled correctly for supervised detection. |
On 2023–25 SEC restatement filings, Benford first-digit screening works as a population-level triage tool, while supervised models can do better when labeled restatement history is available.
The approach that works comes down to recalculating first-digit MAD on the filing's numeric XBRL facts (clear only when MAD is below 0.006; chi-square ≤ 15.507 with 8 degrees of freedom) and having labeled Big R / little r history.

Two Screens, Two Different Questions
Benford screenin
g is a population-level test: it takes the numeric facts in a single SEC filing's XBRL instance, groups their first digits into nine bins, compares each observed share with the Benford expectation, log10(1+1/d), and condenses the differences into a single mean absolute deviation (MAD). The result is one overall determination of whether that filing's digit pattern is unusual; it does not identify which tagged fact, if any, caused the deviation. A low MAD clears the filing's overall population of numeric facts; it is not a definitive assessment of any individual account.
Supervised restatement models work at the record level. A gradient-boosted classifier, for example, scores each tagged fact against patterns learned from labeled outcomes, so its usefulness depends on having a labeled training set. Without such labels, a supervised score is an unvalidated ranking rather than a restatement likelihood. The Audit Analytics Restatements dataset provides that kind of history for all SEC registrants that have disclosed a financial statement in electronic filings since January 1, 2000, including both Big R non-reliance restatements and little r revisions.
That difference defines the division of labor: Benford's analysis answers, "Is this filing's digit pattern unusual?" while a supervised model answers, "Which line items look like past restatements?" Benford's analysis can be run on a filing with no labeled history; a supervised model cannot be trained or validated without labels for the relevant peer population.
Apply the reader rule stated above to the filing itself: the MAD and chi-square gate should be recomputed from that filing’s own numeric XBRL facts—not borrowed from another filing or a peer group. Because filing size can skew the digit distribution, verify the population before treating the gate as a pass. The threshold applies to the population as a whole; it doesn’t replace reviewing individual line items.
Used together, the screens work in sequence, not interchangeably: Benford’s law tells you whether the filing’s full set of numeric facts warrants a closer look, while a supervised model—where you have labeled Big R and little r history—narrows the focus to specific tagged facts that resemble past restatements. These are two different questions, so neither screen substitutes for the other.

The Evidence: Three Converging Sources
Three sources point to the same operating rule for 2023–25 SEC restatement work: treat each filing as its own statistical population, and treat the labeling decision as a separate question. Ideagen Audit Analytics’ May 2026 financial restatements report draws on 20 years of data to unpack what drove 2025 restatement activity, and it’s available for download. Check the report for annual Big R counts in 2023, 2024, and 2025; if it has them, use those counts as your positive-class base rate instead of relying on memory or a secondary summary. The Audit Analytics Restatements dataset itself covers SEC registrants that have filed financial statements electronically since January 1, 2000, and includes both non-reliance (Big R) and revision (little r) events—so any extraction has to keep the two distinct.
asses at the source.The ASC 250 analysis posted on fedlaws.org on September 27, 2026 explains why the base rate needs to be specific to each category. It notes that "Big R" restatements undermine investor confidence, frequently trigger securities class actions, and can result in executive compensation clawbacks under exchange listing rules—and that this is where most of the cost and reputational risk sits. Before treating "little r" revisions as equivalent, check how the source characterizes their consequences. Practical takeaway: when you pull the annual Ideagen counts, keep Big R and little r in separate columns, and use the Big R column to drive any downstream exposure estimate. Combining them into a single restatement total would blur the line between an event that can lead to a class action and a revision whose risk treatment needs to be verified at the source.
The third source looks at what happens when the labels are sparse. The paper "Using Semi-Supervised Learning to Detect and Predict Unlabeled Restatements" notes that mixing restatements of audited and unaudited reports adds an upward bias to the observed link between audit effort and restatements, and that after correcting for two sources of bias, the authors find a robust negative relationship. The implication is a label-hygiene check before you trust any supervised or semi-supervised model: make sure the training set doesn't blend outcomes from audited and unaudited reports, since the bias pushes in a known direction. Where a filer's peer group has no verified Big R/little r history, the model has nothing reliable to learn from that is worth the maintenance cost; where that history exists and is cleanly separated, the model earns its place.
None of these three sources substitutes for recomputing the first-digit MAD on the filing in front of you. The Ideagen report gives you the base rate, the ASC 250 analysis gives you the class weighting, and the semi-supervised paper gives you the label-bias warning. The filing-level screen still has to be run on that filing’s own numeric XBRL facts, with the observed first-digit shares compared against log10(1+1/d) across the nine bins. Clear the population only when MAD < 0.006, with chi-square at or below 15.507 at 8 degrees of freedom as the companion check. If either statistic fails, the flag stays open regardless of what the base rate says about how common Big R events were in the same year.
The convergence is therefore procedural, not just thematic. Download and extract the annual Big R count; keep it separate from little r because the cost analysis says the exposure is concentrated there; audit the audited-versus-unaudited composition of any labeled set before fitting; then recompute MAD on the filing’s own facts and apply the 0.006 / 15.507 gate. Each source answers a different one of those steps, and using any one of them alone leaves a gap the other two were cited to close.
Benford vs Supervised: Pick by Labels
The choice between Benford screening and supervised modeling turns on one condition: whether labeled restatement outcomes exist for the filer's peer set. When that label history is absent or thin, a digit-based screen is the only defensible starting point; when it is present at sufficient scale, a supervised model can prioritize review, but only after the filing passes a population-level gate. The table below compares the three operating options and states the winner for each label condition.
| Option | Inputs required | Output | Primary limitation | Use when |
|---|---|---|---|---|
| A. Benford MAD + chi-square only | Numeric facts from the filing's Inline XBRL; no labels | Population pass/flag signal | Produces no record-level queue | Winner when fewer than ~50 labeled restatements exist in scope |
| B. Supervised model only | Labeled Big R/little r outcomes for comparable filers | Ranked per-fact review queue | Silently fails on digit-pattern manipulation and on schemes absent from training labels | Never use as the sole screen |
| C. Benford gate → supervised triage | Inline XBRL facts plus ~50 or more labeled restatements in the peer set | Population gate followed by a ranked per-fact queue | Depends on label quality and peer-set relevance | Winner when ~50 or more labeled restatements exist in scope |
Option A works with zero labels and runs on any Inline XBRL filing because it operates directly on the extracted numeric facts. Its constraint is structural: it can clear or flag the filing as a whole but cannot tell a reviewer which specific facts to examine first. When the peer set lacks enough labeled restatements to train or validate a reliable classifier, that trade-off is acceptable; verify label sufficiency by checking the available labeled outcomes and validation results rather than assuming a fixed count.
Option B should never be used as the sole screen. A supervised model can return a ranked per-fact queue, which is operationally attractive, but it learns only from the schemes represented in its training labels. Manipulations that leave a recognizable digit-pattern anomaly but no labeled precedent, or novel schemes outside the training distribution, can pass without a flag. Treating a clean model score as clearance confuses label coverage with filing coverage.
Option C earns its keep only when labeled Big R/little r history exists at scale for the filer's peer set. In that configuration, Benford clears or flags the population first, and the supervised model is applied only to prioritize review within filings that warrant deeper attention. Before acting on any flag in this workflow, recompute the first-digit MAD on that filing's own numeric XBRL facts and clear the population only when MAD < 0.006 (chi-square ≤ 15.507 at 8 df); only then should a ranked queue drive reviewer effort.
Reviewer Capacity Math
Staff the screen before you run it by measuring, not assuming, the workload. Take your first 10 filings, extract the numeric XBRL facts, compute the first-digit MAD and chi-square on each filing's own facts, and record how many filings flag under the reader rule. Use that observed flag rate and your measured triage time per flagged filing to project the full batch. Recompute the rate from your own filings before relying on it; fact counts, XBRL tagging density, and the mix of restatement-era filings will move the result.
Write the arithmetic into the review plan as a capacity check. Take your observed flag count from the first 10 filings, divide by 10 to get a per-filing flag rate, then multiply by the number of filings in the full batch. Multiply that expected flag count by the measured triage time per flag, not by the two-hour planning figure, and add a separate line for the time spent recomputing first-digit MAD on each flagged filing’s own numeric XBRL facts. If the resulting reviewer-days exceed the team’s committed capacity, reduce the batch size or add reviewers before the screen starts, not after flags are already queued.
Budget for the false clear as well as the false flag. A Benford pass is not an acquittal: it means the filing’s digit distribution stayed inside the population-level tolerance, not that no manipulated line item is present. If a manipulated filing passes and a Big R surfaces later, the cost is not triage time. Audit Analytics’ Restatements dataset covers SEC registrants that have disclosed financial statements in electronic filings since January 1, 2000, including non-reliance Big R and revision little r events, so a later non-reliance disclosure can be matched back to the filing that was cleared. The ASC 250 discussion describes Big R restatements as damaging investor confidence, often triggering securities class actions, and potentially leading to executive compensation clawbacks under exchange listing rules; that is the downside the capacity model has to price.
Price the false clear in reviewer-days, not just in dollars. Set aside time to reopen a cleared filing when a later Big R is disclosed, recompute the first-digit MAD on that filing’s own numeric XBRL facts, and document whether the earlier pass reflected a clean population or a tolerance that was too loose for that filing size. That rework is the cost of a false clear, and it lands outside the original 10–20-hour planning envelope. A team that only counts triage hours will understate the true workload.
Use the same arithmetic to decide whether the single pre-action number is affordable. Before acting on any Benford flag, recompute the first-digit MAD on that filing’s own numeric XBRL facts and clear the population only when MAD is below 0.006, with chi-square at or below 15.507 at 8 degrees of freedom. If the team cannot fund that recomputation for every flagged filing, it cannot fund the screen. The capacity math is therefore a go/no-go gate: staff the measured flag volume, staff the false-clear rework, and only then release the results to reviewers.
What This Evidence Cannot Show
This is the one place to draw the boundary around the MAD rule itself, because the same threshold that clears a mature filer can mislead you on a thin or structurally skewed population. The reader rule stays fixed: recompute the first-digit MAD on the filing's own numeric XBRL facts, and clear the population only when MAD < 0.006, with chi-square at or below 15.507 at 8 degrees of freedom. What changes is whether that population is fit to be tested at all. Before you read a flag as evidence, check the shape of the input.
Start with fact count. Before treating a MAD result as evidence, record how many numeric XBRL facts the filing contributes to the nine first-digit bins. If the bins are too sparse for a stable MAD, do not clear or reject the filing on MAD alone; gather more facts or use another check. Verify the population size and bin counts on the filing itself before applying the threshold.
The second break is legitimate digit skew. Benford screening assumes data that arise on a natural scale. Revenue lines dominated by unit prices, round-dollar accruals, and other administratively set amounts can depart from the expected first-digit distribution without any restatement signal. That is a population-composition problem, not a model failure. Filter to continuous, naturally scaled facts before computing MAD, then rerun the same check. If the filtered population clears MAD < 0.006 and the chi-square bound, the rule still wins; if you cannot isolate such facts, treat the screen as inconclusive rather than adverse.
| Condition in the filing | What the MAD rule can show | Action before acting on a flag |
|---|---|---|
| Under roughly 500 numeric XBRL facts, as in many small shells or SPACs | Bin shares are too sparse for a stable MAD; the threshold is not dependable here | Do not clear or reject on MAD alone; gather more facts or use another check |
| 2,000 or more numeric XBRL facts from an established filer | The same MAD < 0.006 and chi-square ≤ 15.507 at 8 df check is informative | Recompute on the filing's own facts and apply the reader rule |
| Unit-price-dominated revenue lines or round-dollar accruals | Natural-scale assumption is strained; skew may be ordinary | Filter to continuous, naturally scaled facts, then recompute MAD |
| Supervised labels drawn from unaudited periods | Label quality is the weak point, not the digit screen | Verify label provenance before treating any model output as comparable history |
The third break sits on the supervised side. The semi-supervised learning paper on unlabeled financial reports warns that comingling restatements of audited with unaudited reports introduces an additional upward bias in the association between audit effort and restatements. That caution matters when you assemble Big R and little r history from the Audit Analytics Restatements dataset, which covers SEC registrants disclosing financial statements in electronic filings since 1 January 2000 and includes both non-reliance and revision events. If the labels come from unaudited periods, check provenance before you trust the comparison class; a clean MAD result on the filing does not repair a dirty label set.
The practical sequence is narrow. Verify the population size, remove or isolate administratively skewed facts, recompute MAD on what remains, and only then decide whether the reader rule has been satisfied. Where the population is too thin or the labels are mixed, say the screen is inconclusive and move on. That is the limit of what this evidence can show.
One 10-K Screen
Take one 2024 10-K and treat that filing as the unit of work. Pull every numeric fact from EDGAR’s Inline XBRL structured data, then filter to positive continuous-scale values; identifiers, dates, percentages used as labels, and dimensional qualifiers do not belong in the digit census. In this worked example the filtered population is n = 3,200 facts. That count is illustrative, not a target: recompute it on the filing in front of you, because the MAD cutoff has to be re-verified against that filing’s own numeric facts rather than carried over from a prior screen.
First checkpoint: compute the first-digit MAD across the nine bins, comparing each observed share with log10(1+1/d). Here the result is 0.0042, which clears the 0.006 cutoff. Clear the population only on that result — not on a visual impression that the digits look right, and not on a threshold borrowed from another filing of a different size.
Second checkpoint: run the chi-square goodness-of-fit test at 8 degrees of freedom. The computed statistic is 11.3, below the 15.507 critical value, so the test agrees with the MAD result. Agreement matters. If the two checks point in different directions on a filing of this size, hold the population for inspection rather than picking the friendlier number.
Third checkpoint: inspect the per-digit table before closing the screen. Digit 9 appears at 4.1% observed against 4.6% expected, a modest gap with no single-digit driver. Because no one bin is carrying the deviation, this example calls for no targeted follow-up. A clean population result is not a clean bill of health for individual line items; it only says the filing’s aggregate digit pattern is unremarkable under these two tests.
Had the MAD come back 0.009 instead, the population would be flagged and the 0.006 clearance would fail. The response at that point is to recheck the fact extraction and filtering first, then examine which digit bins contribute most to the deviation, and only then decide whether the flag survives on the corrected population. The Audit Analytics Restatements dataset, as described in its help documentation, covers SEC registrants disclosing financial statements electronically since 1 January 2000 and includes both non-reliance (Big R) and revision (little r) events; that history is useful context, but it does not substitute for recomputing MAD on this filing.
Copy-usable verification checklist: (1) extract Inline XBRL numeric facts from the single 2024 10-K; (2) filter to positive continuous-scale values and record the resulting n; (3) compute first-digit shares and MAD on that n; (4) require MAD < 0.006; (5) compute chi-square at 8 df and require the statistic to be at or below 15.507; (6) read the nine-bin detail for a single-digit driver; (7) if MAD fails, re-verify extraction before escalating. Keep the checklist with the filing, since the single number to verify before acting on any Benford flag is the MAD recomputed from that filing’s own facts.
Worked Example: Run the Numbers
Here is the single worked example this section runs end to end. The scenario is an illustration: a single SEC registrant files a 2023–25 restatement filing, and an analyst pulls the numeric XBRL facts from that filing's own instance document on the filing date. The party is the registrant; the source is the filing itself, not a cross-filing aggregate. Audit Analytics' Restatements dataset, as described in its help documentation, covers SEC registrants that have disclosed financial statements in electronic filings since 1 January 2000 and includes both non-reliance (Big R) and revision (little r) events — so the illustration sits inside a documented population, but the digit test below is run on one filing only.
Step one is the count. Illustration: the analyst extracts the numeric XBRL facts and groups them by first significant digit into nine bins. For each digit d, the Benford expected share is log10(1+1/d). That gives the familiar descending expectations — roughly 0.301 for 1, 0.176 for 2, down toward 0.046 for 9 — but the exact expected share is recomputed from the formula, not copied from a table. The observed share for each bin is that bin's fact count divided by the total number of numeric facts in the filing. Every division is carried out on this filing's own denominator; no prior filing's count is borrowed.
Step two is the deviation and the MAD. Illustration: for each of the nine bins, take the absolute value of observed share minus expected share, sum those nine absolute differences, and divide by nine. That quotient is the mean absolute deviation for this filing. The reader rule is to clear the population only when MAD < 0.006. If the computed MAD lands at or above that figure, the flag stays open and the filing is not cleared on Benford grounds alone. The check is arithmetic and repeatable: another analyst with the same XBRL instance should reproduce the same nine bins and the same quotient.
Step three is the cross-check. Illustration: the same nine observed and expected counts are fed into a chi-square goodness-of-fit statistic, with eight degrees of freedom because the nine bin shares sum to one. The reference ceiling is chi-square ≤ 15.507 at 8 df. The MAD test and the chi-square test are two views of the same binned counts, so a filing that clears one but not the other warrants a recheck of the bin counts before any action is taken. This is a verification step, not a second screening method.
For this illustration, declare the winner as the Benford population screen, because the example supplies only the filing's own numeric facts and no labeled history of prior Big R or little r outcomes. Its break-even trigger is the reader rule itself: MAD < 0.006 with chi-square ≤ 15.507 at 8 df. If the recomputed MAD misses that trigger, the illustration does not clear. A supervised model would only enter this example if labeled Big R/little r history for comparable filings were supplied; absent those labels, there is nothing to train or score against, and the single number to verify before acting remains the filing-level MAD.
Decision Rules for the Filing Queue
Once the filing-level MAD has been recomputed from that filing’s own numeric XBRL facts, the queue needs deterministic routing rather than analyst discretion. The rule set below is the operational layer: it takes the verified MAD, pairs it with the chi-square result at eight degrees of freedom, and decides whether the filing clears, goes to targeted inspection, or moves into supervised triage. I apply these rules to the 2023–25 restatement queue only after the population screen has been recalculated on the individual filing; a cached MAD from a prior period or a peer filer is not a substitute.
First branch: if MAD < 0.006 and chi-square ≤ 15.507 at 8 df, clear the filing’s population and stop. Do not spend supervised-model compute on it. That clearance is narrow but meaningful: the observed first-digit distribution is close enough to log10(1+1/d) expectations under both the mean-absolute-deviation screen and the chi-square check, so the filing does not warrant a digit-pattern flag on this pass. The Audit Analytics Restatements dataset, as documented by Ideagen Audit Analytics, covers SEC registrants disclosing financial statements in electronic filings since 1 January 2000 and includes both non-reliance Big R and revision little r events; that history is useful later, but it does not override a clean two-test result at this stage.
Second branch: if MAD ≥ 0.006 but chi-square ≤ 15.507, treat the filing as borderline rather than failed. The aggregate chi-square has not crossed the cutoff, so the excess concentration is not broad-based across the nine bins. Inspect the per-digit chi-square components next. Flag only if one digit drives the excess — the digit-9 shortfall pattern from the AP-transaction precedent is the working example of a single-bin distortion that merits a manual look even when the total statistic stays under the limit. If the components are diffuse, record the borderline result and keep the filing in routine review rather than escalating it.
Third branch: if MAD ≥ 0.006 and the filer has a prior Big R in the Audit Analytics dataset within 2020–2025, escalate to supervised triage with Big R as the label, regardless of where the chi-square lands. This is the point where labeled history changes the economics of the queue. A prior non-reliance event gives the model a concrete target drawn from the same registrant history that Ideagen Audit Analytics tracks, and the supervised stage can then rank individual line items instead of re-asking the population-level question. A little r alone does not trigger this branch under the stated rule; the escalation key is Big R.
The remaining case — MAD ≥ 0.006, chi-square > 15.507, and no qualifying Big R label — stays with the Benford side of the queue for manual population review rather than supervised scoring. Without a verified label, a supervised model has nothing stable to learn from this filer, which is consistent with the caution in the semi-supervised restatement-detection literature about bias when restatement status is comingled or unlabeled. In practice, I log the MAD, the chi-square total, the dominant digit component if any, and the Big R lookup result as four separate fields. That keeps the if/then path auditable and prevents a borderline digit pattern from being silently converted into a restatement prediction.
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Recompute the first-digit MAD on the filing's own numeric XBRL facts before acting on any Benford flag. | Benford screening is population-level triage for 2023–25 SEC restatement filings, so the flag must be tested against that filing's own numeric population. |
| 2 | Clear the population only when MAD is below 0.006, with chi-square at most 15.507 at 8 df and α = 0.05. | This is the stated clearance condition; a failed MAD means the Benford flag is not cleared for that filing. |
| 3 | Re-verify the MAD threshold for each filing's numeric population size rather than reusing one threshold across filings. | The threshold must be checked per filing because Benford first-digit screening depends on the filing's own numeric population. |
| 4 | If MAD fails, route to record-level supervised scoring only when labeled restatement outcomes from the Audit Analytics dataset are available. | Supervised models are justified only when labeled restatement history exists. |
| 5 | Separate non-reliance (Big R) and revision (little r) restatements in the training labels. | Audit Analytics restatement data includes both types, and the supervised model must distinguish them. |
| 6 | Do not apply supervised scoring to unlabeled filings. | Without labeled restatement outcomes, supervised scoring would exceed the supported evidence base. |
Frequently Asked Questions
When should a Benford flag be cleared?
Clear a Benford flag only when the first-digit MAD is below 0.006, recalculated using the filing's own numeric XBRL facts.
What statistical cutoff accompanies the MAD threshold?
Clear the population only when MAD is below 0.006 (chi-square ≤ 15.507 with 8 degrees of freedom).
Why must the MAD threshold be re-checked for each filing?
Benford first-digit screening is a population-level triage tool for 2023–25 SEC restatement filings, so its MAD threshold needs to be re-verified against the size of each filing's numeric population.
When are supervised models worth using?
Supervised models are worth the effort only when you have labeled restatement history that distinguishes non-reliance (Big R) restatements from revisions (little r).
How should Big R and little r labels be handled?
Audit Analytics restatement data covers both non-reliance (Big R) and revision (little r) disclosures from SEC registrants, and those need to be labeled correctly for supervised detection.
What is the practical bottom line?
The approach that works comes down to recalculating first-digit MAD on the filing's numeric XBRL facts (clear only when MAD is below 0.006; chi-square ≤ 15.507 with 8 degrees of freedom).
Quick answers
| When should a Benford flag be cleared? | Clear a Benford flag only when the first-digit MAD is below 0.006. |
| How should the first-digit MAD be recalculated? | Recalculate the first-digit mean absolute deviation using the filing's own numeric XBRL facts. |
| What chi-square condition accompanies clearing the population? | Clear the population only when MAD is below 0.006, with chi-square ≤ 15.507 with 8 degrees of freedom. |
| Why should the Benford MAD threshold be re-checked for each filing size? | Benford first-digit screening is a population-level triage tool for 2023–25 SEC restatement filings, so its MAD threshold needs to be re-verified against the size of each filing's numeric population. |
| When are supervised models worth the effort? | Supervised models are worth the effort only when you have labeled restatement history that distinguishes non-reliance (Big R) restatements from revisions (little r). |
Also worth reading: Revenue recognition anomaly detection: 2026 30% fewer review hours vs precision: Revenue recognition anomaly detection: 2026 · Audit anomaly detection 2026: Isolation Forest Audit Standard (ISA 315) 30% vs Hold: Audit anomaly detection 2026: Isolation · AP audit alerts: verify anomaly thresholds and false positives before you commit: AP audit alerts: verify anomaly