The 68-95-99.7 Rule Explained

Two numbers — a mean and a standard deviation — are enough to sketch an entire bell curve, and the sketch comes with a famous warranty: 68% of the data within one standard deviation, 95% within two, 99.7% within three. That warranty is the empirical rule, and knowing when it applies is one of the most useful things a beginning statistician can learn.

The rule, stated plainly

For roughly bell-shaped data:

μ ± 1σ  holds about 68%   of the values
μ ± 2σ  holds about 95%   of the values
μ ± 3σ  holds about 99.7% of the values

Adult male heights average about 170 cm with a standard deviation near 7 cm. The rule instantly produces three ranges: 163–177 cm holds about two out of three men, 156–184 cm holds about 95%, and 149–191 cm holds all but about 3 in 1,000. One glance at a two-number summary and you can say something concrete about every member of the population.

Why it works: the normal curve's fixed shape

The rule is a property of the normal distribution's exact shape. Every normal curve — of heights, errors, or exam scores — is the same bell resized by two dials: the mean slides it along the axis, the standard deviation stretches it wider or narrower. Because the shape is fixed, the fraction of area within k standard deviations is the same for every normal dataset. Those fractions are 68.27%, 95.45% and 99.73% — the rule's rounded mnemonic values — and they can be computed precisely from the error function for any multiple of σ, not just the friendly integers 1, 2 and 3.

This is also what makes the rule practical rather than academic: with the mean and the standard deviation of a real sample, you can predict how unusual any future observation will be before you ever see it.

Worked example: IQ scores

IQ is designed to be normal with μ = 100 and σ = 15:

68%  → 85 to 115     (the "ordinary" band, 2 in 3 people)
95%  → 70 to 130     (19 in 20 people)
99.7% → 55 to 145    (all but ~3 in 1,000)

Flip the logic and the rule becomes a detector. Roughly 0.15% of a normal population scores above 145 — about 3 in 2,000 — so a classroom of 30 should contain zero such scores. If you keep meeting them, either the group is selected (a gifted programme) or the data was never normal to begin with. The same one-in-a-thousand logic underlies industrial control charts: a 3σ alarm on a stable process is usually a real fault, not bad luck, because 99.7% of everything should already be inside the lines.

When the rule breaks

The rule belongs to symmetric, bell-shaped data. Apply it to incomes, house prices, or website response times — distributions with long right tails — and it fails on the tail: far more than 0.3% of values sit beyond three standard deviations, because the tail is doing the stretching. The quickest sanity check is to compare the mean with the median (the mean vs median guide covers the diagnostic) and to glance at a histogram. A second caveat: small samples might be normal in principle but too sparse to show the promised proportions — 68% of 10 data points is not 6.8 points.

Beyond 1, 2, 3: where 1.96 comes from

The three friendly multiples have an exact-math underbelly worth knowing: μ ± 1.96σ holds exactly 95.0004% of a normal distribution. That is the origin of the ±1.96 margin in 95% confidence intervals and the "z = 1.96" that appears in every intro-statistics exam. Other multiples follow the same pattern: ±2.5σ covers 98.76%, ±0.674σ covers exactly 50% (half the data, in either direction, lies within two-thirds of a standard deviation). Whenever you need a band the mnemonic does not cover, compute the exact normal coverage rather than interpolating the rule.

Try it yourself

Enter any mean and standard deviation in the Empirical Rule Calculator: it draws the three classic bands as concrete low–high ranges, shows the exact normal coverage behind each rounded percentage, and accepts custom multiples like 1.96 or 2.5. For a deeper dive into what the standard deviation itself measures — the engine this whole rule runs on — read How Standard Deviation Works, and to turn band members into explicit probability statements, try the Probability Calculator.