Fake Invoice Detection in 2026: Benford's Law vs Isolation Forest on 10,000 Records

With 10,000 records, the Benford calculation compares nine observed first-digit bins against nine expected proportions and condenses them into a single MAD value. Isolation Forest, by contrast, uses the same 10,000 records to learn the structure of the feature space and then scores each record individually. As a result, Benford answers whether the population's digit pattern is unusual, while Isolation Forest identifies which records fall in sparse regions of the feature space.

The two methods are not interchangeable: a population can pass Benford while still containing individually unusual invoices, and a population can fail Benford because of a broad shift in digit distribution rather than a handful of suspicious invoices. Benford screens the ledger's overall distribution; Isolation Forest ranks the review queue based on record-level anomalies.

Neither output is proof of fraud on its own. Benford's MAD is sensitive to rounding, minimum invoice thresholds, and mixed populations, while Isolation Forest scores depend on the choice of features and the assumed contamination rate. A high Benford MAD or a high isolation score should prompt document-level verification before any invoice is labeled as fake.

What 10,000 records can and cannot prove

For a 10,000-invoice ledger, Benford's first-digit law yields method-calculated expected counts of 3,010 invoices beginning with 1 and 458 beginning with 9, because P(1)=0.301 and P(9)=0.046. A uniform benchmark of 1,111 invoices per digit is the wrong comparison; Benford's distribution is not uniform, so a leading-1 share near 30% is expected rather than suspicious. Comparing observed counts to 1,111 would mislabel a normal Benford population as anomalous.

Nigrini's first-digit MAD is (1/9)Σ|observed share−Benford share|. Treat the test protocol's MAD cutoff as a screening trigger: verify the cutoff before use, and if the population's first-digit distribution departs enough from Benford, decompose by vendor, month, or amount band. It is not a fraud conclusion, and it should not be ported to first-two-digit MAD because the expected vector changes from nine leading digits to ninety.

Liu, Ting, and Zhou's Isolation Forest score, s(x,n)=2^{−E(h(x))/c(n)}, turns an observation's expected isolation path length into a normalized score; shorter expected paths push the score toward 1. That score is useful for ranking records into a review queue, but it remains a relative anomaly measure rather than proof about any individual invoice. A high score says "review this first," not "this invoice is fake."

The table separates what the 10,000-record test can establish from what it cannot.

The practical takeaway is that 10,000 records are enough to make a population-level Benford screen useful, but not enough to turn a statistical flag into an accusation. Any invoice that still looks suspicious after ranking needs document-level corroboration before it can be called fake.

Options Compared: Screening vs. Ranking

The most defensible approach for the 2026 10,000-invoice test is a hybrid: use Benford’s first-digit MAD to screen the population, use Isolation Forest to rank the review queue, and do not label any invoice as fake without document-level corroboration. Benford alone can flag a population but cannot rank individual invoices; Isolation Forest alone can rank records but cannot explain why the population distribution is off. The hybrid approach keeps each tool in its proper role.

Benford first-digit MAD is the gate because it outputs a single population flag based on gross invoice amounts. It is useful for screening only: it asks whether the ledger's leading-digit pattern departs from the expected population pattern. If MAD exceeds 0.0018, the population should be broken down before any invoice-level conclusions are drawn. That cutoff is a screening trigger, not a fraud verdict.

Isolation Forest is the triage layer because it outputs an individual rank based on numeric features. With contamination set to 0.02, it produces a quota-driven queue rather than a population diagnosis. It is useful for triage only: it tells reviewers where to start, not whether an invoice is false. Its weakness is that it cannot explain the anomaly in audit terms.

The hybrid combines the two in sequence: MAD gate first, then contamination quota. The gate keeps a ranking model from being asked to explain a population-level distribution problem; the ranking model keeps the gate from becoming an unmanageable list of suspects. The hybrid still needs document corroboration, because neither method sees invoices, approvals, or supporting evidence.

Therefore, the winner is not the screen alone or the ranker alone. It is the hybrid gate plus triage, with the final determination reserved for document-level review.

Costs and review capacity at 10,000

The Benford pass is cheap in computational terms: one scan of 10,000 gross amounts to fill nine first-digit bins and compute nine absolute deviations. The cost is data preparation—cleaning amounts, handling credits, reversals, and currency normalization—not model training or tuning.

Isolation Forest scoring is the heavier step; if the test protocol specifies an ensemble size, verify the resulting path-evaluation count before relying on it. With contamination set to 0.02, the 10,000-record population maps to a 200-invoice review queue. That queue is a triage output, not a fraud finding; each item still needs document-level corroboration before any invoice is called fake, so the queue's cost is review labor rather than an automatic fraud count.

Review capacity is where the budget bites. Measure reviewer throughput before sizing the queue: reviewer-days equal the queue size divided by observed invoices per reviewer-day, and hours equal the queue size times observed hours per invoice. Use those measured rates to compare a 200-invoice queue with a larger queue instead of assuming fixed productivity numbers.

Those two quotas frame the operating choice: the 200-invoice queue fits a short review sprint, while the 500-invoice queue consumes a full month of one reviewer's half-time capacity. The arithmetic assumes reviewer time only; it excludes document retrieval, vendor follow-up, and any re-performance of the screen after decomposition.

If the Benford screen does not clear its threshold, the population should be decomposed before scoring; otherwise the review queue inherits population-level distortion and the 200-invoice budget is spent on noise. The cost model therefore treats decomposition as part of the screen, not as an optional cleanup step.

Costs are therefore front-loaded into data preparation and back-loaded into review, with scoring in between as a fixed, repeatable pass. The hybrid design keeps the first pass as a population screen and the second as a ranked queue so review hours are spent on records the model has already isolated, not on the whole ledger.

Worked synthetic 10,000-invoice example

For a synthetic 10,000-invoice ledger, choose an illustrative fake-label rate and observed first-digit counts that sum to 10,000. Keep the fake labels separate for evaluation only; in a live review they would be unknown. Verify every count against the ledger before computing MAD.

Compute expected counts at n=10,000 from Benford's first-digit probabilities, sum the absolute differences between observed and expected counts, divide by 10,000 and by 9 to get MAD, and compare that result with the documented screening threshold. A result above the threshold triggers decomposition before ranking; it is not, by itself, a fraud conclusion.

Isolation Forest with contamination = 0.02 queues 200 invoices. The queue is a review order, not a verdict: no queued invoice is called fake until document-level corroboration is checked, such as vendor master match, purchase order, receiving record, or approval trail. The 9,800 unqueued invoices remain tied to the population screen result rather than being treated as cleared.

Use the worksheet as a gate: compare the computed MAD with the documented screening threshold; if it exceeds the threshold, decompose the population, rank the capped queue, and require corroboration for every queued invoice. The synthetic arithmetic illustrates the method and should not be read as proof that a live ledger has the same fake rate.

Decision rules for the 2026 audit test

If the Benford first-digit MAD screen fires above 0.0018, treat it as a population-level trigger rather than an invoice-level accusation: decompose the ledger by vendor, period, and amount band before any invoice is opened, and do not open an invoice solely because the population MAD is high. The screen identifies where to look; it does not identify which invoice is wrong.

If the Isolation Forest queue is capped at the top 200, require a named feature before assignment to review: duplicate bank account, missing purchase order, vendor age under 30 days, or a bank-detail change. A high anomaly score without one of those features is not enough to consume review capacity, because the score ranks records but does not by itself establish a control failure.

If a record is in the top-ranked queue but has no document or vendor-master mismatch, mark it "monitor" rather than "fraud review"; no invoice is called fake without document-level corroboration. Monitoring preserves the record for trend review while keeping the fraud queue limited to records with an evidence path.

If a decomposed stratum still shows MAD above 0.0018 after removing known one-time payments or credit adjustments, rerun Isolation Forest inside that stratum before expanding the review queue beyond the capped size. This keeps the second pass tied to the stratum that actually failed the screen instead of reopening the whole population.

If two or more named features appear on the same record, escalate it ahead of single-feature records, but still require document-level corroboration before any fraud conclusion. If no named feature appears, the record remains in monitoring even when its anomaly score is high.

If a named feature cannot be verified from source documents or the vendor master, downgrade the record to "monitor" until the underlying data is corrected; do not let an unverified flag drive a fraud review.

Frequently Asked Questions

Can a 10,000-invoice ledger pass Benford's Law and still contain suspicious invoices?

A population can pass Benford while still containing individually unusual invoices that fall in sparse regions of the feature space.

What non-fraud factors can cause Benford's MAD value to be high?

Benford's MAD is sensitive to rounding, minimum invoice thresholds, and mixed populations.

Why might a ledger fail Benford's Law even if there are no individually suspicious invoices?

A population can fail Benford because of a broad shift in digit distribution rather than a handful of suspicious invoices.

What inputs influence the anomaly scores produced by Isolation Forest?

Isolation Forest scores depend on the choice of features and the assumed contamination rate.

Should a high Benford MAD or a high Isolation Forest score be treated as proof of a fake invoice?

A high Benford MAD or a high isolation score should prompt document-level verification before any invoice is labeled as fake.

How many invoices beginning with the digit 1 are expected in a 10,000-invoice ledger under Benford's first-digit law?

For a 10,000-invoice ledger, Benford's first-digit law yields method-calculated expected counts of 3,010 invoices beginning with 1.

Quick answers

Also worth reading: Benford's Law in Audit 2026: 12,000 Accounts Payable Transactions, Flag or Clear?: Benford's Law in Audit 2026: · Benford's Law 2026: MAD Zero and Sequential Tests from 2025 Filings: Benford's Law 2026: MAD Zero · Benford's Law: 8% False-Positive Rate in Q1 2026 10-Q Revenue: Benford's Law: 8% False-Positive Rate