Benford's Law: When Numbers Tell on Themselves

⬅️ Back to Articles

If I asked you to pick a random real-world number, what are the chances it starts with 1? One in nine, about 11 percent, right? Nine digits, equal odds.

That intuition is spectacularly wrong. In almost any large natural dataset (river lengths, company revenues, city populations, physical constants), the digit 1 leads about 30 percent of the time. Digit 9 shows up less than 5 percent. And the same curve holds whether you measure rivers in miles or kilometres, whether you count populations in 1938 or 2022, whether the data comes from Earth or the stars.

This is Benford’s Law. A simple statistical pattern that is both beautifully clean and practically useful for catching people who are lying with numbers. Vatsal Bakshi’s writeup is the best single explanation I have found, and an ACM paper by Ferreira and Levy (2023) showed me an application I had never considered.

  1. The formula fits on one line: P(d) = log₁₀(1 + 1/d). For digit 1, that gives 30.1 percent. For 2, it’s 17.6 percent. For 9, it’s 4.6 percent. The numbers sum to exactly 100, and they describe the first-digit distribution of almost any dataset that spans multiple orders of magnitude. Human heights (1.5m to 2.1m) are too narrow a range to follow the law. Mountain heights (100m to 8,849m) span nearly two orders. It works.

  2. The law was discovered twice. Simon Newcomb, an astronomer, noticed in 1881 that the first pages of his logarithm tables were far more worn than the later pages. People kept looking up numbers starting with 1. He published a paper that was essentially ignored. Fifty-seven years later, Frank Benford independently spotted the same pattern at General Electric and ran a systematic study across 20,229 data points. The law got his name. Theodore Hill finally proved it rigorously in 1995, showing that if you mix enough different distributions together, the aggregate converges to Benford regardless of the components.

  3. The key is scale invariance. If every number in a Benford-compliant dataset is multiplied by any positive constant, the distribution doesn’t change. Measure river lengths in miles then switch to kilometres. The first-digit curve stays identical. Benford’s Law is the unique first-digit distribution that is invariant under multiplication, which is why it applies across currencies, units, and time periods.

  4. Fraud detection is the best-known application, and it works because humans are bad at faking randomness. When people fabricate financial numbers, they avoid 1 as a leading digit (it feels too small, too obvious) and gravitate toward 3, 5, and 7. The resulting distribution is flatter and more uniform than Benford predicts. The gap is measurable and often dramatic. Real accounting data hugs the curve. Fabricated data visibly diverges. Forensic accountants use this as a screening tool. Deviation is not proof of fraud, but it is a strong signal worth investigating.

  5. The 2009 Iranian presidential election is a famous case. Political scientist Walter Mebane applied Benford’s Law to the reported vote totals and found significant second-digit deviation from the expected distribution. Separate analysis by Roukema in 2014 confirmed first-digit anomalies in the same data. Benford analysis did not prove the election was rigged, but it added statistical weight to concerns raised by opposition groups. The same technique flagged Greek macroeconomic data before the 2009 deficit revelation.

  6. I did not expect Benford’s Law to also apply to supercomputer failures. A 2023 paper by Ferreira and Levy at SC Workshops used the law to analyse failure data from the Astra supercomputer. The idea is simple. When a system enters a period of unusually frequent or patterned failures, the error counts deviate from the Benford distribution. The deviation becomes a signal that the current failure-mitigation strategy might be suboptimal. If the numbers look wrong, something unusual is happening.

  7. The law has limits. It only applies to data spanning multiple orders of magnitude. Phone numbers, ZIP codes, and identity numbers are assigned arbitrarily and won’t follow it. Even for natural data, a Benford deviation is a signal to investigate, not a conviction. The Enron data showed deviation, but the actual accounting fraud there was far more direct. Benford was a supporting clue, not the smoking gun.

The takeaway: Benford’s Law is the closest thing statistics has to a universal cheat detector. If a dataset is supposed to reflect real-world measurement and doesn’t follow the curve, someone has probably been in the kitchen. Run the test yourself next time you look at financial data. It takes thirty seconds and a logarithm table, and it catches people who think they are being clever.

Related TMFNK Content

Crepi il lupo! 🐺