Normal Distribution Calculator (z-score & probability)
The normal distribution — the famous bell curve — describes a host of phenomena: heights, measurement errors, test scores, quality control. Give the mean μ and the standard deviation σ, then get the probability that a value falls below, above or between two bounds — or the reverse: the threshold value for a given probability.
The z-score expresses a distance from the mean in number of standard deviations: z = (x − μ)/σ. It reduces any normal distribution to the standard N(0, 1), the one in the tables. z = 0 is the mean, z = 2 is two standard deviations above.
Parameters
Result
- P(X < x) (x = 1.960) 0.9750
| Mean μ | 0.0000 |
| Standard deviation σ | 1.0000 |
| Value x | 1.9600 |
| z-score z = (x − μ)/σ | 1.9600 |
| P(X < x) | 0.9750 (97.5002 %) |
| P(X > x) | 0.0250 (2.4998 %) |
z = (x − μ) / σ · P(X < x) = Φ(z)
Making sense of the bell curve
A normal distribution is fully set by two numbers: its mean μ (the peak of the bell, around which values cluster) and its standard deviation σ (its width). Change μ and the whole curve slides. Change σ and it fattens or pinches. Nothing else is free to choose.
The curve is symmetric. A step of +a above the mean is as likely as a step of −a below it. Mean, median and mode sit at the same place — unlike a right-skewed sample, where those three peel apart. Heights in a same-sex adult group, many measurement errors, a well-tuned process on a shop floor: those are the jobs the bell was hired for. On paper it looks like a smooth hill. In a histogram of a few dozen points it looks like a lumpy cousin of that hill, and that is still the right model if the generating process is.
We write X ~ N(μ, σ²). The second argument is the variance, σ squared. People say “N(100, 15)” in conversation and mean σ = 15; the strict notation is N(100, 15²) = N(100, 225). This page asks for σ, not σ². Type 15, not 225, or every probability will be wrong.
The z-score: distance in standard deviations
Raw x is in the units of the problem: IQ points, millimetres, test scores. Probability tables do not speak those units. They speak z: how many standard deviations separate x from the mean.
z = (x − μ) / σ · P(X < x) = Φ(z)
z = 0 is the mean. z = 1 is one σ above. z = −2 is two σ below. The cumulative function Φ(z) — the CDF of the standard normal N(0, 1) — is the area under the bell to the left of z. That area is P(X < x). Because the curve is continuous, P(X = x) is zero; P(X < x) and P(X ≤ x) are the same number. The “above” probability is the complement 1 − Φ(z). An interval is a difference of two Φ values.
Standardising does not change the story. It only changes the ruler. Any N(μ, σ²) reduces to N(0, 1) through that one division. Compute z, read Φ, stop. That is the whole craft of “normal probability” questions, including the ones that look dressed up in words.
- Below x — P(X < x) = Φ(z).
- Above x — P(X > x) = 1 − Φ(z).
- Between a and b — Φ(zb) − Φ(za).
- Quantile — invert: x = μ + σ · Φ⁻¹(p).
This page does the Φ and the inverse for you. A printed z-table does the same job with more page-turning and a rounding rule. We come back to that.
The 68–95–99.7 rule
Three numbers everyone in a stats classroom is expected to know cold. They are areas under the bell, not slogans.
- 68 % of values fall within ± 1 σ of the mean;
- 95 % within ± 2 σ (more precisely ± 1.96 σ for exactly 95 %);
- 99.7 % within ± 3 σ.
A value beyond 3 standard deviations is rare: both tails together hold less than 0.3 %. Six-sigma quality talk lives on that tail. Process capability indices compare a spec width to σ; they assume the bell is a fair picture of the process, which is a separate question.
The “95 % ≈ ± 2 σ” shortcut is good enough to sketch. The number you actually use for a two-sided 95 % interval is 1.96, because Φ(1.96) ≈ 0.975 and the two 2.5 % tails add to 5 %. Mixing 2 with 1.96 on an exam is a style mark, not a conceptual one. Mixing 1.96 with 1.645 is a conceptual one — that is one tail versus two, below.
P(μ − kσ < X < μ + kσ) ≈ 0.68 (k=1), 0.95 (k=2), 0.997 (k=3)
Worked examples: IQ scores on N(100, 15)
Wechsler IQ is scaled to N(100, 15²) by construction. The tests are forced onto that ruler; the population of test-takers is then close to it. Three questions, three modes on this page.
- Share above 130. Mode “P(X > x)”, μ = 100, σ = 15, x = 130. z = (130 − 100)/15 = 2. Φ(2) ≈ 0.9772, so P(X > 130) ≈ 2.3 %. “Gifted” cutoffs sit around here; the exact label varies by programme.
- Central interval 85 to 115. Mode “between”. z = ±1. About 68 % — the first bullet of the empirical rule, recovered from the calculator.
- Score that beats 90 % of people. Mode “quantile”, p = 0.90. z ≈ 1.28, x = 100 + 15 × 1.28 ≈ 119.2. Not a round number. Real cutoffs get rounded; the probability attached to 120 is a shade over 90 %.
Same three questions work for any μ and σ. Heights, assay results, residuals. Type the two parameters, pick the question, read the area. In short: z first, area second, story last.
The probability is computed with an approximation of the error function (accuracy ≈ 10⁻⁷) and Acklam's inverse quantile. Ample for homework, labs and most applied work; the last decimals may differ from a printed table. The normal distribution models a real phenomenon — it is always only an approximation of concrete, ultimately bounded data.
SAT scores and grading on a curve
College Board's graduating class of 2024 had a mean SAT total of 1,024 out of 1,600. Published summaries put the standard deviation a little over 200 points (around 229). The old classroom model, from the 200–800 section scale, was even cleaner: each section treated as N(500, 100). Either way, a reported score is a point on a bell, and a percentile is an area.
A 1,400 total against μ = 1,024 and σ = 229 is z ≈ 1.64. Φ(1.64) ≈ 0.95, so roughly the 95th percentile if the bell were exact. It is not exact: the SAT is bounded at 400 and 1,600, and the actual histogram has a bit of shape the Gaussian does not. Close enough to place a student. Not close enough to pretend a 1,600 is “z = +2.5 of an unbounded normal”.
Grading “on a curve” in a classroom is a different animal. Some instructors force the class onto a bell — a fixed share of As, Bs, Cs — which is a ranking policy, not a claim that knowledge is normal. Other instructors keep a raw cut (90 / 80 / 70) and the histogram can pile up on the right. Ask which rule is in force before you convert a z-score into a letter. Frankly, a lot of “the curve” talk is just “we ranked you”, dressed in σ.
For a section modelled as N(500, 100), P(X > 700) is P(z > 2) ≈ 2.3 %, the same tail as IQ 130. That is why 700-per-section used to sound rare. Recent totals sit lower because the testing population changed, not because the algebra of Φ changed.
The central limit theorem: why means look normal
The bell shows up in places that are not themselves normal. Average 40 independent waiting times, 40 exam scores, 40 measurement errors, and the distribution of the mean is close to normal even when a single observation is skewed. That is the central limit theorem. The sample mean X̄ of n i.i.d. draws with mean μ and finite variance σ² is approximately N(μ, σ²/n). The standard deviation of the mean — the standard error — shrinks as 1/√n.
X̄ ≈ N(μ, σ / √n) for large n
The folklore threshold is n = 30. It is a rule of thumb, not a law. A mildly skewed population is already fine at n = 20. A brutal skew (income, claims, city sizes) wants more. Binary data (a proportion) is a special case: np and n(1 − p) both need to be decently large before the normal approximation to the binomial is safe.
This page does not need n. It already assumes X is normal. The CLT is why that assumption is often reasonable for a mean or a total even when it is a poor model for a single raw value. At bottom, “use the bell for the average, not for the individual” is the sentence that saves a lot of bad charts.
Confidence intervals for a mean, z-tests, control charts for x-bar: they all ride this result. If you only remember one reason the normal distribution is famous, remember this one. The IQ scaling is a convention. The CLT is a theorem.
When the bell lies: skew, floors and ceilings
The catch: a lot of real variables refuse the job. Household income is right-skewed; the Census headline is a median for that reason. Reaction times, insurance claims, city populations, file sizes: same shape, often closer to log-normal. Taking logs and then using this page on ln X is a standard move. Using this page on raw income is how you invent negative dollars in the left tail.
The mathematical normal has support (−∞, +∞). Heights cannot be negative. SAT totals cannot be 200 or 1,800. Test scores with a floor of 0 and a ceiling of 100 pile up at the ends. For a process well inside its bounds, with σ small compared with the distance to the wall, the overflow is negligible. For a distribution that can see the wall — a hard 0, a hard 100, a physical constraint — the tail probabilities this page prints are the model's, not the world's.
- Use the bell — symmetric unimodal data, measurement error, a mean of many things, a process in control.
- Do not — income, counts of rare events (Poisson), waiting times, anything with a hard bound sitting a couple of σ away from μ.
- Check — histogram, mean vs median, a normal probability plot. If mean and median disagree by much, stop.
For what it's worth, “approximately normal” is the honest phrase. Quality texts that treat every dimension as N(μ, σ²) are choosing a model that is good enough for spec limits, not discovering a law of nature.
One-tailed vs two-tailed, and what a z-table shows
A printed z-table almost always gives Φ(z) = P(Z ≤ z) for z ≥ 0, in steps of 0.01. Negative z is recovered by symmetry: Φ(−z) = 1 − Φ(z). This page computes Φ directly, sign and all, so you do not have to flip. Inverse mode is the table read backwards: given p, find z, then x = μ + σz. Acklam's inverse is what sits behind the quantile option; a table would give you z = 1.96 for p = 0.975, same idea, two-decimal z.
Tails are where exams (and papers) go wrong. A two-sided 5 % test splits the 5 % across both tails: 2.5 % each, critical z = ±1.96. A one-sided 5 % test puts the whole 5 % on one side: z = 1.645 (or −1.645). Same α, different geometry. Asking “is the mean different from 100?” is two-tailed. Asking “is it greater than 100?” is one-tailed. You do not get to pick the prettier p-value after seeing the data.
two-sided 95 %: z = ±1.96 · one-sided 95 %: z = 1.645
P(X > x) on this page is already a right-tail probability. Doubling it to “make it two-tailed” is only valid when the alternative is two-sided and the statistic is symmetric, which the normal is. Doubling an interval probability is nonsense. Read the question: one cut, two cuts, or a quantile. Then pick the matching mode.
Tables stop around z = 3.49. Beyond that they print 1.0000 and you have lost the tail. This page still returns a small probability out in the far wing, within the 10⁻⁷ accuracy of the erf approximation. For six-sigma work you want a better tail expansion. For a stats class, Φ(4) ≈ 0.99997 is already “essentially everyone”.