New: 6 free SQL practice datasets with 300+ questions — try the SQL Compiler →
Statistics

Distribution Shape: Skewness and Kurtosis

Distribution shape explained: percentiles, quartiles, IQR, skewness and kurtosis, plus two real ways to flag outliers, all worked with real numbers.

Upskly AI Team October 5, 2026 8 min read
Distribution Shape: Skewness and Kurtosis

Two datasets can share the same mean and the same spread and still look nothing alike. What separates them is shape: which side the tail stretches toward, how sharply the middle peaks, and whether a handful of extreme values are quietly distorting the usual summary numbers. This covers percentiles and quartiles, skewness, kurtosis, and two practical ways to flag an outlier, all worked against one real dataset: 55, 62, 58, 70, 65, 59, 61, 100, 63, 67.

In this guide

Percentiles, quartiles and IQR

A percentile describes relative position, not frequency: a value at the 90th percentile simply means roughly 90 percent of the data falls at or below it. A quartile is just a percentile at one of three fixed stops, the 25th, 50th and 75th, which split a sorted dataset into four equal quarters.

Percentile and quartile values (linear interpolation)
MeasureValue for 55, 62, 58, 70, 65, 59, 61, 100, 63, 67
Q1 (25th percentile)59.5
Q3 (75th percentile)66.5
IQR (Q3 minus Q1)7.0
90th percentile73.0

The interquartile range (IQR) is simply Q3 minus Q1, the width of the middle half of the data once the top and bottom quarters are trimmed away. That trimming is exactly why IQR stays close to 7.0 here even though the dataset contains an outlier all the way up at 100, outliers live in the trimmed-off tails, not the middle half IQR measures.

Skewness: which way does the tail point

Skewness measures the asymmetry of a distribution’s shape around its centre. A value near zero suggests a roughly symmetric shape; a value well away from zero shows a lopsided one, and the sign tells you which side.

Reading the sign of skewness
Skew signTail points towardWhere most of the data sits
Positive (right-skewed)Higher valuesBunched at the lower end
Negative (left-skewed)Lower valuesBunched at the higher end

One outlier, two directions

The dataset 55, 62, 58, 70, 65, 59, 61, 100, 63, 67 has a single high outlier at 100, and its skewness comes out to +2.10, strongly positive. Swap that outlier for a single low one instead (90, 88, 85, 87, 20, 86, 89, 91, 84, with 20 sitting far below the rest) and skewness flips to -2.43, strongly negative. One extreme value on either side is enough to drive skewness a long way from zero.

Why skew pulls the mean away from the median

In a strongly right-skewed dataset, the mean typically sits above the median. The reason is mechanical: the mean adds up every value, so a high-value tail drags it upward, while the median only depends on which value sits in the middle of the sorted order and barely moves.

Mean pulled above median by the outlier at 100 (skewness +2.10)
MeasureValue
Mean66.0
Median62.5
Mean minus median+3.5

This gap is a useful sanity check in the other direction too: if a dataset’s mean sits noticeably above its median, that is itself a sign of positive skew, even before computing skewness directly.

Kurtosis: tail weight and peak sharpness

Kurtosis measures how heavy a distribution’s tails are and how sharply its centre peaks, compared to the normal distribution. It is a different property from centre, spread or symmetry: a distribution can be perfectly symmetric and still have unusually heavy or light tails.

Excess kurtosis subtracts the normal distribution’s own kurtosis as a baseline, so the normal distribution’s excess kurtosis is exactly 0.

The three kurtosis regimes
TermExcess kurtosisShape
LeptokurticPositiveHeavier tails, a sharper, taller peak than normal
MesokurticAbout 0Matches the normal distribution's tail weight and peak
PlatykurticNegativeLighter tails, a flatter, more spread-out peak than normal
Same shape measure, opposite stories
DatasetExcess kurtosisRegime
55, 62, 58, 70, 65, 59, 61, 100, 63, 67 (has the 100 outlier)+3.39Leptokurtic
12, 15, 14, 16, 13, 15, 14, 15, 16, 14 (tightly clustered)-0.56Mildly platykurtic

Two ways to flag an outlier

An outlier is a value that lies unusually far from the rest of the data, not simply any value above the mean or the most recently entered one. Two rules of thumb do the actual flagging.

The 1.5 x IQR rule builds a fence on each side of the middle half of the data:

Fences for 55, 62, 58, 70, 65, 59, 61, 100, 63, 67
FenceFormulaValue here
Upper fenceQ3 + 1.5 x IQR66.5 + 1.5 x 7.0 = 77.0
Lower fenceQ1 minus 1.5 x IQR59.5 minus 1.5 x 7.0 = 49.0

Only the value 100 falls outside the 49.0 to 77.0 range, so exactly one value in this dataset counts as an outlier under this rule.

The Z-score rule instead asks how many standard deviations a value sits from the mean, and a value beyond roughly plus or minus 3 is often treated as a possible outlier, since about 99.7 percent of normally distributed data falls within 3 standard deviations of the mean. Here, 100 sits at a Z-score of 2.67, close to that boundary but not quite past it, a reminder that the two rules will not always agree exactly on the same cutoff.

Median absolute deviation (MAD)

The median absolute deviation (MAD) measures spread using distances from the median rather than the mean, then takes the median of those distances too. Because it leans on the median at every step, a single extreme outlier barely shifts it, unlike the standard deviation.

The same outlier, two very different reactions
MeasureValue for 55, 62, 58, 70, 65, 59, 61, 100, 63, 67
Standard deviation12.73 (inflated by the outlier at 100)
MAD4.0 (barely moved by it)

Common mistakes

  • Mixing up the direction of skew. Positive (right) skew means the tail stretches toward higher values, not that most of the data sits on the high side; most values actually bunch at the lower end.
  • Treating skewness and kurtosis as the same thing. Skewness is about asymmetry; kurtosis is about tail weight and peak sharpness. A distribution can be symmetric and still be strongly leptokurtic.
  • Assuming a skewness near zero means no outliers. It only means the data is not lopsided to either side; it says nothing by itself about outliers or peak shape.
  • Using the standard deviation to judge spread when an outlier is present. The outlier inflates it; MAD or the IQR stay far more representative of the typical spread.
  • Treating the 1.5 x IQR and Z-score rules as interchangeable. They can disagree on the same dataset, as they do here, so a value worth a second look under one rule is not automatically flagged by the other.

Cheat sheet

Quick answers
QuestionAnswer
What splits sorted data into four equal quarters?The quartiles (25th, 50th, 75th percentiles)
What is Q3 minus Q1 called?The interquartile range (IQR)
Which way does a positively skewed tail point?Toward higher values
Which way does a negatively skewed tail point?Toward lower values
In a right-skewed dataset, is the mean above or below the median?Above
What does excess kurtosis of 0 indicate?Tail weight and peak match the normal distribution
What is a leptokurtic distribution?Heavier tails and a sharper peak than normal (positive excess kurtosis)
What is a platykurtic distribution?Lighter tails and a flatter peak than normal (negative excess kurtosis)
What two values define the 1.5 x IQR fences?Q3 + 1.5 x IQR and Q1 minus 1.5 x IQR
What Z-score is commonly used to flag a possible outlier?Roughly plus or minus 3
Why is MAD more outlier-resistant than standard deviation?It is built entirely from distances to the median, not the mean

Try it yourself

Quick check

Eight quiz scores: 68, 72, 75, 74, 70, 73, 71, 30. Without computing anything, what sign would you expect the skewness to have, and why?

Show answer

Negative. One score, 30, sits far below the rest, which pulls the tail toward lower values. Checking the real numbers confirms it: the mean is 66.62 but the median is 71.5, the mean sitting below the median, exactly the pattern left skew predicts, and the computed skewness comes out to -2.17.

Frequently asked questions

What is the difference between a percentile and a quartile?

Percentiles come in a continuum from the 1st to the 99th. Quartiles are just three specific percentiles, the 25th, 50th and 75th, which split data into four equal quarters.

What does skewness measure?

Skewness measures the asymmetry of a distribution’s shape around its centre. A value near zero suggests a roughly symmetric shape; larger positive or negative values show a lopsided one, and the sign shows which side the tail stretches toward.

What does kurtosis measure, and how is it different from skewness?

Kurtosis measures tail weight and peak sharpness relative to the normal distribution, not asymmetry. A distribution can be perfectly symmetric (skewness near zero) and still have unusually heavy tails (positive excess kurtosis).

How does the 1.5 x IQR rule flag an outlier?

It builds an upper fence at Q3 plus 1.5 times the IQR and a lower fence at Q1 minus 1.5 times the IQR. Any value outside those two fences is flagged as a possible outlier.

Why does a right-skewed distribution usually have a mean above its median?

The mean sums every value, so a high-value tail drags it upward. The median only depends on which value sits in the middle of the sorted order, so it barely moves even with an extreme tail present.

Why is the median absolute deviation (MAD) more resistant to outliers than the standard deviation?

MAD measures distances from the median, then takes the median of those distances, so every step leans on a position-based statistic rather than one that sums every value. A single extreme outlier barely shifts it.

Test yourself

Timed questions on Distribution Shape, with an explanation for every answer.

Take the Quiz
Upskly AI Team
Learning made simple
Scroll to Top