Kurtosis Explained
Press Next → or use ← → arrow keys
The Intuition — Why Tails Matter
Kurtosis is the statistic that measures them. It captures how much of a distribution's action lives out in the tails — the propensity for rare, extreme events that mean and spread completely overlook.
The 2008 crisis featured market moves that "normal" risk models rated as once-in-the-history-of-the-universe rare — yet they happened. The models ignored kurtosis. Financial returns have fat tails, so extreme days occur far more often than a bell curve predicts.
Lepto, Meso & Platykurtic
The common myth is that kurtosis measures how "pointy" a distribution is. It doesn't — it measures tail weight. Leptokurtic has heavy tails (more outliers), platykurtic has light tails (fewer), and mesokurtic is the normal-distribution baseline. The tall peak of a leptokurtic curve is a side-effect of mass being pulled from the shoulders into the tails.
The Fourth Standardized Moment
Raising each deviation to the fourth power massively amplifies points far from the mean: a value twice as far contributes sixteen times as much. That's what makes kurtosis a tail-heaviness detector — and also why a single outlier can dominate it. For a normal distribution the maths works out to exactly K = 3, the baseline everything is compared against.
The Three Types, Side By Side
| Type | Kurtosis | Excess | Tails | Examples |
|---|---|---|---|---|
| Leptokurtic | > 3 | > 0 | Heavy / fat | Stock returns, insurance claims, earthquakes |
| Mesokurtic | = 3 | = 0 | Moderate | Heights, IQ, measurement error (normal) |
| Platykurtic | < 3 | < 0 | Light / thin | Uniform data, capped scores, tight tolerances |
Lepto = thin (a thin, tall waist with energy escaping into the tails); platy = broad/flat (think "plateau"); meso = middle. Match the word to the silhouette and you'll never mix them up.
Fat Tails = Extreme Events Are Common
Two distributions can share the same mean and standard deviation, yet the leptokurtic one keeps real probability mass far out in the tail where the normal curve has essentially hit zero. That extra tail area is what turns "impossible" events into merely "rare" ones — the mathematical signature of financial crashes, insurance mega-claims, and structural failures.
One Outlier Dominates The Kurtosis
Nine mild readings plus one 32°C spike. That single outlier's fourth-power contribution is 10,000 out of a total 10,292 — it alone drives raw kurtosis to 6.28 (excess 3.28, firmly leptokurtic). One point, and the whole dataset "looks" heavy-tailed. Kurtosis is exquisitely sensitive to outliers — always visualize before trusting it.
Reading Excess Kurtosis On A Number Line
On the excess scale, 0 is the normal distribution. Negative means lighter tails than normal (platykurtic — like a uniform distribution at −1.2); positive means heavier (leptokurtic — like a Student-t at +5.8). The further from zero, the more the tails depart from the bell.
Kurtosis vs Skewness
| Property | Kurtosis | Skewness |
|---|---|---|
| Measures | Tail heaviness (outlier propensity) | Asymmetry (which way it leans) |
| Moment | 4th central moment | 3rd central moment |
| Normal value | 3 (raw) / 0 (excess) | 0 |
| Outlier sensitivity | Extreme (4th power) | High (3rd power) |
| Answers | "How extreme?" | "Which direction?" |
Mean and standard deviation describe where data sits and how spread it is. Skewness then tells you which way it leans, and kurtosis how heavy its tails are. Together the four give you a complete shape diagnostic — never rely on mean and variance alone.
Where Fat Tails Bite
When a risk model calls a market move a "25-standard-deviation event," it's not that the impossible happened — it's that the model used a thin-tailed normal distribution on fat-tailed data. Measuring kurtosis first tells you when a normal assumption will dangerously understate real-world risk.
Three Traps To Avoid
| ❌ The Trap | ✅ The Reality |
|---|---|
| "Kurtosis measures how pointy the peak is" | It measures tail weight — the peak is a side-effect |
| Assuming your tool returns raw kurtosis | SciPy/Pandas/Excel default to EXCESS (normal = 0) |
| Trusting a high value on messy data | One outlier can fake leptokurtosis — always plot |
A kurtosis of "0" and a kurtosis of "3" can mean the same normal distribution — the first is
excess, the second is raw. Before interpreting any kurtosis number, confirm which convention your tool
uses (in SciPy, that's the fisher parameter), or you'll be off by exactly 3.
Kurtosis In Python
from scipy import stats import numpy as np data = [18, 20, 20, 21, 21, 21, 22, 22, 23, 32] stats.kurtosis(data) # 3.28 → EXCESS (fisher=True, default) stats.kurtosis(data, fisher=False) # 6.28 → RAW (normal = 3) # ── compare three shapes ── np.random.seed(42) stats.kurtosis(np.random.normal(size=10000)) # ≈ 0.02 mesokurtic stats.kurtosis(np.random.standard_t(3, 10000)) # ≈ 5.85 leptokurtic stats.kurtosis(np.random.uniform(-3, 3, 10000)) # ≈ −1.19 platykurtic
Series.kurt() and df.kurt() also return excess kurtosis, but with a
sample-size correction, so the number can differ slightly from SciPy's on small datasets. As always,
check the docs for which formula a tool uses before comparing values across libraries.
Six Rules For Kurtosis
You Now Understand Kurtosis
Kurtosis is the fourth-moment shape statistic that measures tail heaviness — how prone a distribution is to extreme values. Leptokurtic (>3) means fat tails and real risk; platykurtic (<3) means thin tails; mesokurtic (=3) is the normal baseline. It's outlier-sensitive by design, so read it alongside a plot and its partner, skewness.
Pair kurtosis with skewness for full shape diagnostics, connect both to the normal distribution and the empirical rule, and explore fat-tailed distributions (Student-t, power laws) that model real financial and extreme-event data.
📉 End of tutorial · Press ← to review, or click Restart