Benford’s law is a logarithmic distribution of significant digits. Its best-known implication is that, for suitable numerical data, the first nonzero decimal digit is 1 about 30.1% of the time and 9 about 4.6% of the time. In financial data analysis, this gives a possible benchmark for screening amounts for unusual digit patterns. The benchmark needs justification: a departure from it does not establish fraud.
The first significant digit of 0.0048, 480 and 480,000 is 4. Leading zeros and the decimal point do not count. This article considers positive amounts; zeros have no nonzero leading digit. Credits, reversals and negative balances need an explicit treatment rather than silently being pooled with payments.
The probability comes from log space
Write a positive value as , where is an integer and . The number is its significand. Under the full Benford law, the fractional part of is uniformly distributed between 0 and 1. Consequently,
Here is the first significant digit. The logarithm turns a ratio into a difference: moving from 1 to 2 doubles the significand, while moving from 9 to 10 multiplies it by only . Those intervals have different lengths on a logarithmic axis. The probabilities sum to one because the log differences telescope to .
This is a derivation conditional on uniform fractional logarithms, not a proof that every ledger has that property. The full law also specifies subsequent significant digits; matching nine first-digit frequencies alone does not establish the full law. Berger and Hill’s mathematical primer, Sections 2–3, states these distinctions and the uniform-log characterization.
Original theoretical diagram. Bar heights are expected counts, not observed payments or a fraud threshold; an actual sample will fluctuate around its model expectations.
For a hypothetical file of 1,000 independent observations with Benford-distributed leading digits, the expected counts are for digit 1 and for digit 9. Fractional expected counts are averages over repeated samples, not instructions to round a ledger into compliance. Independence is needed for standard independent-sample inference; the expectation calculation itself only needs the stated marginal probabilities.
Two honest files can have very different digits
Imagine a fictional accounts-payable review. One file contains invoice amounts arising from different combinations of quantities and prices. Another contains invoice identifiers assigned sequentially from 10,000 to 19,999. The identifiers all begin with 1. That says how the numbering system works, not how honestly the invoices were created.
Now suppose a reimbursement policy admits only whole-dollar claims from 200 through 299. Every admitted amount begins with 2, even if each claim is correct. Comparing that filtered file with the unmodified Benford probabilities would test an inappropriate model.
| What the field contains | What to establish before using Benford |
|---|---|
| Transaction amounts | Whether their generation and selection make Benford a plausible benchmark |
| Assigned account or invoice numbers | The numbering rule; these are labels rather than measured economic sizes |
| Amounts within a policy band | How truncation and eligibility rules alter the digit distribution |
| Rounded prices or repeated fixed charges | How pricing conventions and duplication shape the digits |
Nigrini’s 1999 author article discusses accounting applications and distinguishes amounts from assigned numbers and bounded fields. It is a historical explanation, not a current audit requirement or a universal acceptance rule.
A large range is also insufficient by itself. Multiplicative processes and suitably mixed distributions provide mathematical routes to Benford behaviour, with specific assumptions; “there are many numbers” is not one. Conversely, a model with and uniform on is exactly Benford although its values occupy only one decade. Berger and Hill (2011) explain why broad spread alone is an invalid justification; Hill’s statistical derivation gives precise conditions for mixtures.
Use the anomaly to choose the next question
Define the population and extraction rules before looking for suspicious digits. Keep amount fields separate from identifiers; document zeros, signs, rounding, currency units and filters. With financial statement data, preserve the accounting concept and reporting scope rather than treating every number in a filing as an interchangeable observation.
If the benchmark is defensible, compare observed and expected frequencies. A chi-squared test is one possible goodness-of-fit procedure, subject to its sampling and approximation conditions. Related or duplicated records can invalidate an independent-sample calculation. Searching many vendors, periods and digit combinations also changes the false-alarm problem.
A small p-value concerns the chosen statistical model. It is not the probability that the file is fraudulent. A mismatch can motivate checking the export, business rules and underlying documents. Agreement cannot certify authenticity: fabricated amounts can be designed to reproduce the first-digit pattern, while dishonest transactions can retain authentic amounts.
For the mathematical preparation, the Exponentials and Logarithms lesson explains the log identities. Confidence Intervals and Hypothesis Testing develops model-based evidence and testing errors. Both have public previews; their full lessons require membership.
Sources
- Berger and Hill, The Mathematics of Benford’s Law — A Primer (2019), Sections 2–3 and 5 — significant digits, the uniform-log characterization, exact one-decade construction and qualified convergence results.
- Berger and Hill, Benford’s Law Strikes Back (2011), “Simple Explanations? Are You Sure?” — scale and log-space reasoning; the failure of broad-spread arguments.
- Hill, A Statistical Derivation of the Significant-Digit Law (1996 manuscript), “The Significant-Digit Law” and summary — first-digit versus general digit laws and an explicitly qualified distribution-mixture result.
- Nigrini, I’ve Got Your Number (1999), introductory discussion and dataset characteristics — historical accounting applications and unsuitable assigned or bounded fields.