What is a z-score?
A z-score is a way of asking, "how weird is this number, given the distribution it came from?" It strips away the original units and the scale of the data and replaces them with a single value: the number of standard deviations between your data point and the mean. Positive z means above the mean. Negative z means below it. Zero means right on it.
The point of standardizing is comparison. Test scores from an SAT (out of 1600) and an ACT (out of 36) can't be compared directly — the scales are different, the means are different, the spreads are different. Convert each to a z-score and they're suddenly on the same ruler. A z of 1.5 on the SAT and a z of 1.5 on the ACT mean the same thing: this student scored 1.5 standard deviations above the average test-taker. The original test no longer matters.
Z-scores show up in three places: standardized testing (where percentile rank is the headline number, computed from the z), quality control (where parts more than 3 SDs from the target measurement get flagged), and outlier detection (where any data point with |z| greater than 2 or 3 deserves a second look before it goes into the analysis).
The formula
One line. No exceptions.
z = (x − μ) / σ
Subtract the mean. Divide by the standard deviation. That's the entire calculation.
The symbols come from statistics convention: μ (mu) is the population mean, σ (sigma) is the population standard deviation, and x is your observed value. If you only have a sample rather than the full population, substitute the sample mean (x-bar) and sample standard deviation (s) — the formula doesn't care, though the interpretation of the resulting probabilities shifts slightly when sample sizes are small.
The result is unitless. If x is in pounds, μ is in pounds, and σ is in pounds, the pounds cancel out in both the numerator and the denominator. That's why z-scores from completely different measurement systems are directly comparable.
How to use the Z-Score Calculator
Three inputs, four outputs. The calculator updates as you type; there's no submit button.
- Enter the value you're standardizing into the x field — a test score, a measurement, a single observation.
- Enter the population (or sample) mean μ.
- Enter the population (or sample) standard deviation σ.
- Read the four outputs: the z-score itself, the one-sided P(X ≤ x), the one-sided P(X ≥ x), the two-tailed p-value, and the percentile.
The probabilities assume a normal distribution. If your data is heavily skewed, the z-score is still arithmetically correct, but the percentile and p-value drift from reality. We say more about that further down.
Worked example: an SAT score
A high school senior scores 1400 on the SAT. The College Board's most recent data puts the mean at 1050 and the standard deviation at about 205. How exceptional is that score?
Apply the formula:
z = (1400 − 1050) / 205 = 350 / 205 = 1.71
The student scored 1.71 standard deviations above the mean test-taker. Looking up 1.71 on the normal distribution, P(X ≤ 1400) is about 0.9564 — meaning roughly 95.6% of test-takers scored at or below this student. Put another way, the student is in the top 4.4% of the testing pool. That's the kind of number admissions offices think about, derived from a single division.
If the same student wants to know how they compare to the applicant pool of a specific selective university whose mean SAT is 1450 with σ = 90, the math runs again with new parameters: z = (1400 − 1450) / 90 = −0.56. Against that smaller, more selective distribution, the same 1400 is now slightly below average. The score didn't change. The reference distribution did. That shift is exactly what z-scores are built to make visible.
Z-score, percentile, and the 68-95-99.7 rule
For data that's actually normally distributed, there's a useful set of round numbers to memorize. About 68% of values fall within one standard deviation of the mean. About 95% fall within two. About 99.7% fall within three. Anything beyond three is rare enough that most quality-control systems will flag it as a probable defect or outlier without further questioning.
| z-score | P(X ≤ x) — percentile | What it means |
|---|---|---|
| −3.0 | 0.13% | Bottom 0.13% — very rare on the low side |
| −2.0 | 2.28% | Bottom 2.3% |
| −1.0 | 15.87% | Bottom 16% |
| 0.0 | 50.00% | Exactly the mean — the median in a normal distribution |
| +1.0 | 84.13% | 84th percentile — one SD above the mean |
| +1.71 | 95.64% | The SAT-1400 example — top 4.4% |
| +2.0 | 97.72% | Top 2.3% |
| +3.0 | 99.87% | Top 0.13% — very rare on the high side |
The asymmetry on the high side comes from the cumulative direction: P(X ≤ x) climbs from 0 to 1 as z grows. The percentile is just that probability times 100. A z of +1 puts you above 84% of the distribution. A z of −1 puts you above only 16%. The mean is the 50th percentile.
One-tailed vs two-tailed p-values
If you're running a hypothesis test, the question is whether your observation is too unusual to be explained by chance. The calculator gives both one-sided probabilities and the two-tailed p-value, and which one you use depends on what hypothesis you started with.
One-tailed: you cared about deviations in a specific direction before you saw the data. "Does this new fertilizer increase corn yield?" — you only care about higher yields. A drop wouldn't make the fertilizer work; it would just make it bad. Use the upper tail only. The one-sided p-value for z = 1.71 is about 0.044.
Two-tailed: deviations in either direction would be interesting. "Is this coin fair?" — too many heads OR too many tails both count as evidence against fairness. The two-tailed p-value doubles the more extreme side: about 0.087 for the same z = 1.71. Two-tailed tests are more conservative, which is why most journals and most introductory statistics courses default to them. If your hypothesis didn't specify a direction in advance, you should be using two-tailed.
The honest version of "one-tailed lets me find significance more easily" is a warning, not a feature. Pre-register the direction before collecting data, or stick with two-tailed.
Quality control and outlier detection
Manufacturing uses z-scores constantly. A bolt is supposed to be 50.00 mm long, with a process standard deviation of 0.05 mm. A part comes off the line at 50.18 mm. The z is (50.18 − 50.00) / 0.05 = 3.6. The 68-95-99.7 rule says only about 0.03% of normal output should be that far from the target. One out of 3,000 parts. If you see two in a row, something has drifted — a tool is worn, a temperature has shifted, a setting was bumped. Six Sigma quality programs are built around the same idea, just with stricter thresholds.
For data analysis, the same logic flags outliers. Any observation with |z| greater than 3 (or sometimes 2.5, depending on how aggressive you want to be) gets a second look before going into a regression or a t-test. It might be a typo. It might be a real but uninformative observation that's going to distort everything. It might be the most important data point in your file. The z-score doesn't tell you which — it just tells you to look.
When the calculator's probabilities are wrong
The probabilities all assume the underlying data is normally distributed. The arithmetic of the z-score itself doesn't depend on that assumption — subtracting a mean and dividing by an SD works on any data. But the cumulative probabilities and p-values come from the normal CDF, and they're only correct when the data is actually normal.
The calculator uses the Abramowitz-Stegun 7.1.26 rational approximation for the normal CDF, accurate to about 1.5 × 10⁻⁷ — better than six decimal places. That's the same approximation used inside R, SciPy, Excel, and most statistical software. For homework, scientific reports, business analysis, and quality control, the precision is more than your input data justifies.
But for skewed data (income, wait times, anything with a long right tail), heavy-tailed data (financial returns, network latency), or small samples (use a t-distribution instead), the normal-distribution probabilities will be misleading. The same z = 2.0 corresponds to different p-values under different distributions. Always check whether the normality assumption holds before reading too much into the percentiles. A quick histogram, a Q-Q plot, or a Shapiro-Wilk test is usually enough.
Related calculators
The z-score is rarely the end of an analysis. A few tools that pair with it:
- Standard Deviation Calculator — if you have raw data instead of a pre-computed σ, start here to get the standard deviation, then feed it into the Z-Score Calculator.
- Confidence Interval Calculator — z-scores at 1.96 and 2.576 give you the 95% and 99% confidence interval multipliers. The interval calculator handles the full computation including sample size.
- Probability Calculator — for working with combined or conditional probabilities once you've translated z-scores into single-event probabilities.
- Sample Size Calculator — works out how many observations you need before z-based tests have enough statistical power.
Frequently asked questions
What does a negative z-score mean?
The value is below the mean. The sign tells you direction; the magnitude tells you how far. A z of −1.5 is one and a half standard deviations below the mean, which (under normality) is around the 7th percentile — you're below about 93% of the distribution. The interpretation is symmetric with the positive side: −1.5 and +1.5 are equally unusual, just in opposite directions.
Can z-scores be greater than 3?
Absolutely, but they're rare in normally distributed data — less than 0.3% of observations should fall outside ±3. When you see |z| greater than 4 or 5, three things might be happening: the data isn't actually normal (heavy tails are real, especially in finance), the value is a true rare event, or there's an error in the data. Check the source before drawing conclusions from extreme z-scores.
Why use population σ vs sample s?
If you know the true population standard deviation (rare in practice), use μ and σ — the z-score is exact and follows the standard normal distribution. If you're estimating from a sample (common), substitute the sample mean and sample SD. For small samples (n less than 30 or so), the resulting "z" actually follows a t-distribution with n − 1 degrees of freedom, and you should look up p-values on a t-table rather than the normal one. For large samples, the difference is negligible.
How is this different from a t-score?
A z-score uses the population standard deviation. A t-score uses the sample standard deviation and accounts for the extra uncertainty from estimating σ from data. For samples larger than about 30, the t-distribution converges to the normal distribution and the two are interchangeable. For small samples, the t-distribution has slightly heavier tails — meaning the same numerical "z" corresponds to a more cautious (higher) p-value. The Z-Score Calculator assumes you've got the population σ or a large enough sample that the distinction doesn't matter.
What's the connection between z-scores and IQ?
IQ tests are scaled so that the mean is 100 and the standard deviation is 15. So an IQ of 130 has z = (130 − 100) / 15 = 2.0, which corresponds to the 97.7th percentile — the cutoff most "gifted" programs use. An IQ of 145 (z = 3.0) is in the top 0.13%. The Mensa cutoff is the 98th percentile, roughly z = 2.05 or IQ 131. The whole scoring system is built around standardizing a normal distribution.
How is the calculator accurate to seven decimal places?
The normal CDF doesn't have a closed-form expression — there's no shortcut formula like "sin(x)" that gives an exact answer. Computing it requires either numerical integration or a rational approximation. The calculator uses Abramowitz and Stegun's formula 7.1.26, a polynomial-fraction approximation with a guaranteed error bound of 1.5 × 10⁻⁷. Excel uses a similar approximation. R and SciPy use higher-precision routines, but for everything short of drug-trial-grade work, the difference doesn't matter.
Can I use a z-score with non-numerical data?
No. Z-scores require quantitative data with a meaningful mean and standard deviation. Categorical data (red/blue/green), ordinal data without spacing (ratings of "love it / like it / hate it"), and binary outcomes (yes/no) don't have z-scores in the standard sense. For those, you'd use different statistics: chi-squared tests for categorical data, the Mann-Whitney U for ordinal, or proportion-based z-tests for binary outcomes (which are derived but distinct).