Article — Empirical Rule Calculator (68-95-99.7)
Empirical Rule Calculator: 68-95-99.7 Percentages for Normal Distributions
The empirical rule (also called the 68-95-99.7 rule) states that for a normal distribution, about 68% of data falls within ±1 standard deviation of the mean, 95% within ±2 standard deviations, and 99.7% within ±3. The empirical rule calculator returns these ranges for any mean and standard deviation, plus a z-score and percentile for an optional specific value.
The rule is a quick reasoning shortcut for normally distributed data. It is mathematically exact only when the data follow a true normal distribution, but it works well for many real-world datasets where the histogram is roughly bell-shaped.
What the empirical rule says
For data drawn from a normal distribution with mean μ and standard deviation σ, the proportions inside symmetric intervals around the mean are:
- 68.27% within μ ± 1σ
- 95.45% within μ ± 2σ
- 99.73% within μ ± 3σ
- 99.994% within μ ± 4σ
- 99.99994% within μ ± 5σ
- 99.9999998% within μ ± 6σ (3.4 per billion outside)
The familiar 68-95-99.7 numbers are rounded to two significant figures. The exact values come from integrating the standard normal probability density over the relevant intervals.
The empirical rule formula
The empirical rule itself is a verbal statement, but it ties to the underlying normal-distribution math:
z = (x − μ) / σ z-scoreP(μ−σ ≤ X ≤ μ+σ) ≈ 0.683 1σ rangeP(μ−2σ ≤ X ≤ μ+2σ) ≈ 0.954 2σ rangeP(μ−3σ ≤ X ≤ μ+3σ) ≈ 0.997 3σ rangeThe z-score expresses any value as a number of standard deviations from the mean. With z computed, the empirical rule maps directly: |z| < 1 covers ~68%, |z| < 2 covers ~95%, |z| < 3 covers ~99.7%.
Applying the empirical rule
Suppose IQ scores are normally distributed with mean μ = 100 and σ = 15. Apply the rule:
68% of people score between 85 and 115 (within ±1σ). 95% score between 70 and 130. 99.7% score between 55 and 145. A score of 130 (which is μ + 2σ) is at the 97.5th percentile — only about 2.5% of people score higher. A score of 145 (μ + 3σ) is at roughly the 99.87th percentile, putting it among the top 1.3 in 1000.
Empirical rule and z-scores
The z-score is the bridge between an arbitrary normal distribution and the standard normal (mean 0, standard deviation 1). Once z is computed, the empirical rule applies regardless of the original units: pounds, inches, dollars, or test points.
| z-score | One-tailed percentile | Two-tailed within range |
|---|---|---|
| 0 | 50.00% | 0% |
| 1 | 84.13% | 68.27% |
| 1.96 | 97.50% | 95.00% |
| 2 | 97.72% | 95.45% |
| 2.58 | 99.50% | 99.00% |
| 3 | 99.87% | 99.73% |
The z-score of 1.96 (not 2) is the conventional threshold for a 95% confidence interval. The empirical rule's friendly "95% at 2σ" is slightly off; the exact value is 1.96.
Empirical rule examples
SAT scores. Total SAT scores have mean ≈ 1050 and σ ≈ 200. Applying the empirical rule: 68% score between 850 and 1250; 95% between 650 and 1450; 99.7% between 450 and 1650 (the actual max is 1600, so the upper tail is bounded). A 1400 score is roughly at the 96th percentile.
Manufacturing tolerance. A factory makes bolts with target diameter 10.00 mm and process σ = 0.10 mm. 99.7% of bolts fall between 9.70 and 10.30 mm. If the specification allows ±0.30 mm, the process is centered exactly on tolerance limits at ±3σ — about 0.3% (3 per 1000) will fall outside spec.
Human height. Adult US male height is roughly normal with μ ≈ 70 inches and σ ≈ 3.5 inches. 68% are between 66.5 and 73.5 inches. 95% are between 63 and 77 inches. 99.7% are between 59.5 and 80.5 inches. Heights below 60 or above 80 inches are extremely rare in a typical sample.
The first published derivation of the normal distribution was by Abraham de Moivre in 1738, who used it as an approximation to the binomial distribution. Carl Friedrich Gauss applied it to astronomical errors around 1809, leading to the alternative name "Gaussian distribution." The bell curve was central to Gauss's method of least squares, the foundation of modern regression analysis.
When the empirical rule fails
The rule is exact only for true normal distributions. Real data may be skewed, multi-modal, or have heavy tails. Common deviations:
Right-skewed data (incomes, response times, lifetimes) has a long tail above the mean. The empirical rule overestimates the proportion within ±1σ on the right side and underestimates on the left. Income distributions are notoriously non-normal.
Bimodal data has two peaks (e.g., heights of a mixed group of children and adults). The mean falls between the two modes where almost no data sits, breaking the rule entirely.
Heavy-tailed data like financial returns produces 4σ and 5σ events much more often than the rule predicts. The 2008 financial crisis featured several "25-sigma" market moves, which under a true normal distribution should be expected once in many billions of years.
Before applying the empirical rule, verify the data is roughly normal. Use a histogram (look for symmetric bell shape), a Q-Q plot (points should follow the diagonal line), or a formal test like Shapiro-Wilk. If the distribution is clearly skewed or has heavy tails, switch to Chebyshev's inequality, which works for any distribution but gives weaker bounds.
Empirical rule versus Chebyshev's inequality
Chebyshev's inequality applies to any distribution — normal, skewed, multimodal, anything. It states that at least (1 − 1/k2) of the data falls within ±kσ of the mean. For k = 2, that is 75% (versus 95% under the empirical rule). For k = 3, Chebyshev gives 89% (versus 99.7%).
The trade-off is clear: Chebyshev makes weaker claims but applies universally. The empirical rule makes stronger claims but only for normal data. For unknown distribution shape, Chebyshev is the safe bet; for verified normal data, the empirical rule provides more useful detail.
Six Sigma and the empirical rule
Six Sigma extends the empirical rule logic to extreme tolerance: aiming for process spread at ±6σ from the mean allows only 3.4 defects per million opportunities (99.99966% defect-free). Motorola introduced the methodology in 1986; companies that implement Six Sigma typically allow for a 1.5σ process drift, so the "3.4 dpm" figure includes that buffer.
At ±3σ tolerance, processes produce about 2,700 defects per million (0.27%). Lifting tolerance to ±6σ (with the 1.5σ drift allowance) drops this to 3.4 per million — a thousandfold improvement in quality. The Six Sigma framework formalizes this jump and provides tools (DMAIC, root-cause analysis, control charts) to achieve it.
The trick: getting to Six Sigma requires understanding the underlying process distribution. If the process is not normal, the empirical-rule percentages do not apply, and the Six Sigma calculation must be adjusted using the actual distribution.