Tools › Finance and Business › Auditing › Benford's Law Test › Example

Benford's law, first two digits: 1,430 invoice amounts bunched just under a 5,000 approval limit

ToolBenford's Law Test
Sourceoriginal
LevelProfessional

Problem

A payables ledger holds 1,430 invoice amounts between 10.39 and 194,228.26. Purchases of 5,000 or more need a second signature. Test the first two digits of the amounts against Benford's law at the 5 % level, and count the amounts in the 5 % band just under the limit against the equal band at and above it.

Inputs to Enter

  • Amounts: the 1,430 seeded illustrative amounts of the tool's default case (one per line); amount column automatic; negative amounts by absolute value; smallest amount 10
  • Test in full: first two digits; significance level 5 %; approval limit 5,000 (band 5 %)
Open the Benford's Law Test tool (loaded with this example)

Answers

  • Benford proportion of the pair 48 = log10(1 + 1/48) = 0.008955; expected count = 1,430 × 0.008955 = 12.81; observed 64, or 64/1,430 = 0.04476
  • z = (|0.04476 − 0.008955| − 1/(2 × 1,430)) / √(0.008955 × (1 − 0.008955)/1,430) = 14.23, beyond the critical value 1.960; pairs 49 and 47 follow with z = 11.61 and 11.51
  • First two digits: chi-square 578.75 on 89 degrees of freedom, p < 0.0001; mean absolute deviation 0.00337; 7 of 90 pairs beyond ±1.96 (48, 49, 47, 63, 14, 10, 80) where about 4.5 are expected by chance alone
  • First digit: chi-square 103.71 on 8 degrees of freedom (digit 4: 250 amounts where 138.6 are expected); second digit: chi-square 41.46 on 9
  • 118 amounts lie from 4,750.00 up to 5,000.00 against 6 from 5,000.00 up to 5,250.00; the middle 90 % of the amounts cover 2.16 orders of magnitude

Notes

The tool's default inputs are this case; press Submit, or open /analysis-tools/business/audit/benfords-law-test/?try=under-the-limit. Seeded illustrative data, no real ledger; every figure is checked against an independent numpy and scipy computation. A deviation from Benford's law is a reason to look at those entries, not evidence of fraud or error.

More Examples for This Tool