Empirical Rule Calculator

Two sigma is 95.45%, and the difference travels

The rule rounds three numbers. Two of the roundings are harmless and one is not.

68.2689% → 68% and 99.7300% → 99.7% cost nothing, because nobody carries those figures anywhere.

95.4500% → 95% is different, because 95% is the number that appears in every confidence interval and every hypothesis test. Learning “two standard deviations is 95%” and then reaching for 2 when a 95% interval is wanted is the natural next step, and it is wrong.

The multiplier for exactly 95% is 1.959964. Using 2 instead makes the interval 2.04% wider than it needs to be. That is not catastrophic, and it is in the conservative direction — but it means a 2σ interval is really a 95.45% interval, and reporting it as “95% confident” understates what was actually computed. Two analysts, one using 2 and one using 1.96, produce different intervals from identical data and both call them 95%.

It goes the other way at 99%. Three sigma is 99.73%, not 99% — the exact 99% multiplier is 2.5758. Someone reaching for 3σ because “99%” is wanted overshoots by 16% in width, which is a much bigger error than the 95% case and in the same conservative direction.

The habit to build is to go the other way round: decide the coverage you want, then look up the multiplier. The tool takes either direction, and the critical value calculator is the dedicated tool for the inverse when you need a t or chi-square multiplier rather than a normal one.

What survives when the data is not normal

Every number in the rule is a fact about the normal distribution, not about standard deviations in general. It is worth knowing exactly how much of it is left without that assumption, because the answer is surprisingly little.

Chebyshev’s inequality is the distribution-free version, and it holds for literally any distribution with a finite variance — skewed, bimodal, discrete, heavy-tailed, anything. It guarantees at least 1 − 1/k² within k standard deviations.

At 2σ that is 75%, against the normal’s 95.45%. The normality assumption is worth 20.45 percentage points, and it is an assumption you are making rather than a property you have checked.

At 1σ Chebyshev guarantees nothing at all. The bound is vacuous below k = 1, and not merely as a technicality: a distribution can place every single value more than one standard deviation from its mean — a fair coin scored 0 or 1 does exactly that, with all its mass at exactly 1σ. So the rule’s most-quoted figure has zero support outside the normal case.

One-sided, the gap is wider still. Cantelli’s inequality permits up to 20% of a distribution above μ + 2σ, where a normal puts 2.275% — a factor of 8.8. Since most decisions that matter are one-sided (is this too high? did it exceed the limit?), this is the bound that usually applies, and it is the weakest one.

The practical upshot is not to abandon the rule — it is an excellent quick check when the data really is roughly symmetric and unimodal. It is to stop treating “within two standard deviations” as though it means something without that condition. On right-skewed data the upper tail is heavier than 2.275% and the rule understates how often you will see large values, which is the direction that costs money.

Where three sigma came from, and what six sigma means

The 3σ convention is not arbitrary, and knowing its origin explains why it is 3 and not 2 or 4.

Outside ±3σ is 0.27%, or about 1 in 370. Walter Shewhart chose those limits for control charts in the 1920s on exactly that basis: sample a process hourly and a false alarm arrives roughly every 370 samples, which is a tolerable rate of chasing nothing. At 2σ it would be 1 in 22 — several false alarms a day, and operators would learn to ignore the chart.

Six sigma is a different and more slippery figure. Outside ±6σ on a normal distribution is about 1 in 500 million — two parts per billion. But the quality methodology called Six Sigma quotes 3.4 defects per million, which is far larger. The difference is a deliberate 1.5σ allowance for long-run drift in the process mean, so “six sigma quality” is really the one-sided tail beyond 4.5σ. Both numbers are correct about different things, and comparing them directly is a common error.

The far tail is also where the normal assumption is least defensible. A 6σ calculation is an extrapolation far beyond anything a realistic dataset can verify — you would need hundreds of millions of observations to see one. Real processes have heavier tails than the normal, so those figures are best read as a scale for comparison rather than as literal frequencies. Financial risk models made this mistake at scale, describing daily moves as 25σ events that ought to be impossible and had nonetheless just happened twice that week.

Using it on real scores

Most encounters with the rule are through a standardised score, and the arithmetic is the same each time.

IQ is calibrated to a mean of 100 and a standard deviation of 15. So 130 is exactly 2σ, which puts it at the 97.7th percentile — about 1 person in 44, not 1 in 100. 145 is 3σ: 1 in 741. The rule converts a score to a rarity in one step, and the z-score calculator does the conversion for any scale.

Clinical reference ranges are usually the middle 95%, which means one healthy person in twenty falls outside by construction. Run twenty independent tests on a perfectly healthy patient and the chance that all twenty come back inside is 0.95²⁰ = 36% — so an abnormal result on a broad panel is the expected outcome rather than a finding. This is the same multiple-comparisons arithmetic that p-values face.

Grading on a curve applies the rule as policy rather than as a description. Forcing marks into ±1σ and ±2σ bands assumes ability is normally distributed within the class, which for thirty students is an assertion, not an observation.

Check the shape before trusting any of it. The quickest useful test is not a formal one — compare the mean with the median. If they are close and the histogram has one hump, the rule will be about right. If the mean sits well above the median, the data is right-skewed and every tail figure here is optimistic on the high side. The mean, median and mode calculator shows that gap directly, and the lognormal page covers the case where it is large.

One tail or two — the rule only answers one of them

The empirical rule is stated two-sidedly: 95.45% lies within two standard deviations. Most practical questions are one-sided, and using the two-sided figure for them doubles the error.

“How unusual is a value this high?” is a one-tail question. Above ±2σ sits 4.55% of the distribution, but only 2.275% is above +2σ — the rest is below −2σ, which is not what was asked. Quoting 4.55% for a high value overstates its rarity by exactly two.

The multipliers differ too. A two-sided 95% interval uses 1.960; a one-sided 95% bound uses 1.645. That is a 16% difference in width, and picking the wrong one is the same class of error as using 2 instead of 1.96 — only larger.

The direction has to be chosen before seeing the data. A one-sided test is more powerful precisely because it spends all its error budget on one side, and choosing that side after observing which way the result went converts a nominal 5% test into a real 10% one. This is the same arithmetic as the multiple-comparison problem, with two comparisons.

Most quality and safety limits are one-sided. A specification usually cares about too much contamination, too little strength, too long a delay — not deviation in either direction. A control chart with 3σ limits is two-sided by design, but the process capability question behind it is frequently not, which is why the two get quoted with inconsistent tail probabilities.

The tool prints the one-tail figure alongside the two-tail one for exactly this reason, together with the “1 in n” form, which is the phrasing least likely to be halved or doubled by accident.

Sources and methodology

References for the exact figures and the distribution-free bounds.

• NIST/SEMATECH e-Handbook — Shewhart control charts and the reasoning behind 3σ limits
• NIST/SEMATECH e-Handbook — the normal distribution and its exact coverage figures
• ASQ on Six Sigma — where the 3.4 defects per million figure and its 1.5σ shift come from
• Z-Score Calculator — to convert a raw score to the k this page uses
• Critical Value Calculator — the inverse, for a target coverage rather than a target k
• How we calculate — methodology

Method. Coverage is computed as 2Φ(k) − 1 using the higher-precision normal CDF, and the one-tail figure uses the survival function directly rather than 1 − Φ(k), so it stays accurate past k = 8 where the subtraction would lose every significant digit to floating-point cancellation. The exact multiplier for a target coverage comes from the inverse normal, which agrees with the published 1.959964 to fifteen decimal places. The suite asserts the three headline figures to four decimals, that the 2σ interval is 2.0427% wider than the exact 95% one, and that Chebyshev’s bound sits below the normal figure at every k above 1 while being vacuous below it. That engine is verified on every change against 219 assertions. The count and the per-case breakdown are published on the formula verification page.

Related calculators

Where this goes next:

More in Math, or browse all calculators.