Benford law in audit testing: what it can and cannot prove

The idea

In many naturally occurring financial populations, the leading digit 1 appears in about 30 percent of values, 2 in about 18 percent, and so on down to about 5 percent for 9. Manufactured numbers rarely follow that curve, because people invent digits roughly evenly.

Where the test is valid

• Large populations, ideally several thousand values.
• Values spanning several orders of magnitude.
• No imposed minimum or maximum, and no assigned numbers.

Where it is not

Payroll where salaries cluster in bands, prices ending in 99, invoice numbers, populations under a few hundred records, or any set constrained by a threshold. Running the test there produces a deviation that means nothing.

Reading the result

A deviation is a pointer, never a conclusion. It says: sample this stratum first. Follow it with second-digit tests, duplicate testing and a targeted sample of the over-represented range. Do not report digit analysis as evidence of fraud on its own.

Practical workflow

1. Confirm the population qualifies.
2. Run the first-digit test and record the variance per digit.
3. Stratify the excess: which account, which supplier, which period, which preparer.
4. Sample those entries and inspect the supporting documents.
5. Record what you found, including when the deviation had an innocent cause.

Combine with behavioural checks

Digit analysis pairs well with round-number detection, threshold clustering just below approval limits, and weekend posting analysis. Individually weak, together informative.

NanoFora runs digit analysis alongside duplicate, threshold and period checks, and only raises a finding where the population is large enough to support it. See AI fraud detection in accounting.