Kurtosis Calculator

It measures tails, not peakedness

“Kurtosis measures how peaked a distribution is” appears in most introductory texts, and it is wrong. The correction is not a quibble — the peakedness reading leads to the wrong conclusion about what a high kurtosis is telling you.

The arithmetic settles it. Kurtosis is the average fourth power of the standardised deviation. Values near the mean have standardised deviations below 1, and raising a number below 1 to the fourth power makes it smaller. A value at 0.5 standard deviations contributes 0.0625; one at 4 standard deviations contributes 256. The centre of the distribution is almost entirely written out of the calculation. Whatever kurtosis is measuring, it cannot be the peak.

What it measures is tail weight — how much of the total variance comes from rare, far-out values. High kurtosis means occasional extremes. Low kurtosis means the data is confined.

You can construct a distribution with an infinitely sharp peak and low kurtosis, and one that is flat at the centre with high kurtosis. Westfall documented the history of the error in 2014 under the title “Kurtosis as Peakedness, 1905–2014. R.I.P.” and the reading persists anyway, because a peaked-looking picture usually accompanies heavy tails when the variance is held fixed. The correlation is real; the definition is not.

Try the tool’s uniform preset for the cleanest demonstration. Twelve evenly spaced values — completely flat, no peak at all — give an excess kurtosis of −1.22. Flatness reads as negative kurtosis, because a uniform has no tails, and that is the opposite of what the peakedness story predicts.

The practical reading: a high kurtosis says “this data produces occasional large surprises”. That is a statement about risk, and it is why the measure matters in finance far more than in describing shapes.

Three conventions, and one of them can be impossible

Like skewness, kurtosis names more than one quantity, and here the ambiguity is worse because two of them differ by a constant.

Raw kurtosis is m₄/m₂², and a normal distribution has 3. Excess kurtosis subtracts that 3, so a normal distribution has 0 and the sign tells you immediately whether tails are heavier or lighter than normal.

So “the kurtosis is 3” is genuinely ambiguous — exactly normal under one convention, distinctly heavy-tailed under the other. Those are not near each other, and the word “excess” is frequently dropped.

Excel’s KURT is a third quantity again, a sample-adjusted excess kurtosis. On the tool’s default data the plain excess is 0.094 and KURT returns 1.130 — twelve times larger. Both are defensible; they estimate different things.

And the adjusted version can return a value no distribution can have. Excess kurtosis has an absolute floor of −2. Load the tool’s two-point preset — five zeros and five ones — and the plain excess kurtosis is exactly −2, sitting on the floor. Excel’s KURT returns −2.571.

That is not a spreadsheet bug. The adjusted estimator is constructed to be unbiased for the population value, and an unbiased estimator of a bounded quantity is free to land outside the bound on any particular sample — it only has to be right on average. It does mean the number cannot be read as a description of a distribution, because no distribution has it.

The floor, and what sits on it

Kurtosis has a hard lower bound that is worth knowing, because it turns some reported values into provable errors.

For any distribution, excess kurtosis ≥ skewness² − 2. It follows from the Cauchy–Schwarz inequality applied to the third and fourth moments — algebra, not a tendency. So a symmetric distribution can go as low as −2, and a strongly skewed one cannot get anywhere near it: at a skewness of 2, excess kurtosis cannot be below +2.

The absolute minimum of −2 is reached by exactly one shape — a two-point distribution, with all the mass split between two values and nothing in between or outside. A fair coin scored 0 or 1 is the standard example, and the tool’s preset returns exactly −2.000000.

There is no upper bound. Kurtosis can be arbitrarily large, and for some distributions it is infinite — a t distribution with four or fewer degrees of freedom has no finite fourth moment at all. When kurtosis is infinite, a sample kurtosis is still a number, and it is a number that grows without limit as you collect more data. It is estimating something that does not exist.

Some useful reference values. Normal 0. Uniform −1.2. Exponential +6. Laplace +3. A t distribution with 5 degrees of freedom, +6. Log-returns on most financial assets land between +3 and +10 — comfortably heavier than normal, which is the empirical fact behind most tail-risk modelling.

It is the noisiest of the common statistics

Kurtosis depends on fourth powers, and that makes it far more variable than anything else people compute routinely.

At n = 30 the standard error is about 0.83. So a 95% band under normality runs to roughly ±1.7 — wider than most of the effects people read into a kurtosis. At n = 100 it is still ±0.95. Below a few hundred observations, kurtosis is a weak signal.

A single point can dominate it. Fourth powers grow fast: a value twice as far from the mean contributes sixteen times as much. On heavy-tailed data one observation can supply most of the entire fourth moment, and the tool warns when that happens. In that situation the kurtosis is a statement about that observation, not about the distribution — check it against the outlier calculator before drawing conclusions.

The estimate is also biased downward at small n, for a reason that compounds the problem: a small sample is unlikely to contain the rare extreme values that create high kurtosis in the first place. So the samples most likely to come from a heavy-tailed process are the ones least likely to show it. This is exactly the sample-size trap that the 3σ outlier rule falls into, in a different form.

The honest use: treat kurtosis as a rough flag on samples of a few hundred or more, compare it against its own standard error rather than a fixed threshold, and look at the data before believing it.

Where it earns its keep

Once read as tail weight rather than peakedness, the applications are specific.

Financial risk. Asset returns have consistently positive excess kurtosis, meaning extreme moves happen far more often than a normal model allows. A risk model built on normality underestimates the frequency of large losses, and kurtosis is the single number that says by roughly how much. This is why “25-sigma events” get reported: the sigma is right and the distribution behind it is wrong.

Testing normality alongside skewness. The Jarque–Bera test combines both into a single statistic. It is more informative than either alone, because a distribution can be symmetric and heavy-tailed, or skewed with normal-weight tails, and those need different responses.

Deciding whether the mean is a stable summary. High kurtosis means the sample mean will jump around between samples, because it is occasionally dominated by an extreme value. That is a direct argument for the median, or for a trimmed mean, regardless of any normality question.

Signal processing and quality control, where a rising kurtosis in a vibration or defect signal often indicates impulsive faults appearing before the mean level changes at all.

Where it does not help: describing a distribution’s shape to a general audience. “Excess kurtosis 2.3” conveys nothing. “Values more than three standard deviations out happen four times as often as a normal distribution predicts” conveys the same fact and can be acted on.

Sources and methodology

References for the tail-weight reading and the bound.

• Westfall, “Kurtosis as Peakedness, 1905–2014. R.I.P.”, The American Statistician 68(3), 2014
• NIST/SEMATECH e-Handbook — skewness and kurtosis in the moment-ratio convention
• Microsoft’s KURT reference — the sample-adjusted excess formula, which can leave the valid range
• Skewness Calculator — the third moment, with the same convention problem
• Outlier Calculator — for checking whether one point is producing the kurtosis
• How we calculate — methodology

Method. All three conventions come from the same central moments rather than from rescaling one another. The standard error uses the exact normal-theory expression rather than the √(24/n) approximation, which overstates precision at the sample sizes where kurtosis is actually reported. The tool measures what share of the fourth moment a single observation contributes, and warns above 40% — a check no threshold on the kurtosis itself can perform. The suite asserts that a two-point dataset returns exactly −2, that the g₁² − 2 bound holds across ten thousand random datasets, that a twelve-point uniform returns −1.2168 despite having no peak at all, and that Excel’s adjusted formula returns −2.571 on that two-point data — below a floor no distribution passes. That engine is verified on every change against 62 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.