The shape of data explained — and why a bell curve is rarer than you think.
Fundamentals 14 min Beginner May 10, 2026
You already know the centre (mean, median) and the spread (variance, standard deviation) of your data. But imagine two datasets with the same mean of 50 and the same standard deviation of 10 — one looks like a bell, the other has a long tail to the right. The shape changes everything.
The shape of a distribution determines which summaries you can trust, which models will work, and whether rare events are truly rare — or just hidden in the tail. This article introduces three distribution shapes that cover most of what you will encounter in data science.
The Normal Distribution — The Bell Curve
Normal Distribution
AnalogyDefinition
Imagine a darts player throwing many darts at the bulls-eye. Most land near the centre, fewer land far out, and almost none hit the outer ring. Now imagine the darts fall straight down after sticking into transparent tubes — the tube at the centre fills up the most, the tubes at the edges stay nearly empty. The stacks form a bell curve. The centre is μ (the mean), the typical scatter distance is σ (the standard deviation).
Analogy:
Imagine a darts player throwing many darts at the bulls-eye. Most land near the centre, fewer land far out, and almost none hit the outer ring. Now imagine the darts fall straight down after sticking into transparent tubes — the tube at the centre fills up the most, the tubes at the edges stay nearly empty. The stacks form a bell curve. The centre is μ (the mean), the typical scatter distance is σ (the standard deviation).
Definition:
The normal distribution N(μ, σ²) is a continuous, symmetric distribution shaped like a bell. Fully described by two parameters: μ (centre) and σ (width). The 68-95-99.7 rule tells you what fraction of values falls within certain sigma bands. De Moivre first described the bell curve in 1733, Gauss used it for error theory in astronomy, and Galton demonstrated it with his quincunx board.
The dart-board analogy breaks at two points: first, a dart board is two-dimensional while a normal distribution is typically one-dimensional. Second, real players have systematic bias (they don’t perfectly centre). Third — and most importantly: not all data follows a bell curve.
The 68-95-99.7 Rule
1
±1σ = 68% About 68% of all values lie within one standard deviation of the mean.
2
±2σ = 95% About 95% of all values lie within two standard deviations.
3
±3σ = 99.7% Nearly all values (99.7%) lie within three standard deviations.
This rule works only for normally distributed data. Applied to skewed distributions, it massively underestimates extreme events.
Example: Men in Germany average 180 cm tall with σ = 7 cm. 68% are between 173 and 187 cm. 95% are between 166 and 194 cm. A 210 cm man is over 4σ above the mean — extremely rare. If σ were 2 cm instead of 7, 95% would be between 176 and 184 cm — a much narrower range with the same mean.
Deep Dive: Central Limit Theorem
Why does the bell curve appear everywhere? The central limit theorem provides the answer: when you add up many small, independent random quantities, the sum approaches a normal distribution — regardless of how the individual values are distributed. Roll one die: the outcomes are uniformly distributed (1–6, each equally likely). Roll 30 dice and add the results. Repeat 1,000 times. The sums form a nearly perfect bell curve. Height, measurement errors, and exam scores (with many independent contributing factors) follow normal distributions for exactly this reason.
Common Misconception
"All data follows a normal distribution." Wrong. Many real-world datasets — incomes, web traffic, stock returns — are clearly not normal. The 68-95-99.7 rule is invalid for these. Applying it to skewed data massively underestimates extreme events.
The bell curve is symmetric — equally distributed on both sides. But what happens when the bell hangs crooked?
Skewness & Long-Tail — When the Average Lies
Skewness
AnalogyDefinition
Thought experiment: 50 employees earning ~40,000 euros each sit in a room. Mean and median are both ~40,000. Now a CEO walks in earning 10 million euros per year. The mean jumps to over 235,000 euros — but the median barely moves. The mean now describes the CEO, not the typical person. This is exactly what right-skewness does: a few extreme values hijack the mean while the median stays honest.
Analogy:
Thought experiment: 50 employees earning ~40,000 euros each sit in a room. Mean and median are both ~40,000. Now a CEO walks in earning 10 million euros per year. The mean jumps to over 235,000 euros — but the median barely moves. The mean now describes the CEO, not the typical person. This is exactly what right-skewness does: a few extreme values hijack the mean while the median stays honest.
Definition:
Skewness measures how asymmetric a distribution is. In right-skewed (positive skewness) distributions, there is a long tail to the right: many small to medium values and a few extremely high ones. In left-skewed distributions (e.g., exam scores where most students do well but a few fail badly), it is the mirror image. In skewed distributions: mean ≠ median — the mean is pulled toward the tail.
In our thought experiment there are only two extremes (40,000 and 10 million). In the real world, the long tail is continuous — there are also people earning 100,000, 500,000, and 2 million euros. The curve tapers off exponentially to the right rather than jumping abruptly.
Symmetric
Mean = Median. Values are evenly distributed around the centre. The 68-95-99.7 rule works.
Right-Skewed
Mean > Median. The long tail to the right pulls the mean upward. The 68-95-99.7 rule does not work here.
Company A (Symmetric)
Mean: 50,000 € | Median: 50,000 € — Both values agree. Salaries are evenly distributed.
Company B (Right-Skewed)
Mean: 50,000 € | Median: 35,000 € — Most earn 25–40k, a few executives at 500k+ pull the mean up.
Both companies have a mean salary of 50,000 euros. But reporting "average salary 50,000" for Company B is technically correct yet deeply misleading. The median of 35,000 euros describes the typical salary far more honestly. Long-tail distributions are common in finance ("Black Swan" events), social networks (a few superstar accounts, many with few followers), and machine learning (rare but critical edge cases that live in the tail).
Common Misconception
"The 68-95-99.7 rule works for any distribution." Wrong — it works only for normal (symmetric, bell-shaped) data. For right-skewed income data, far more than 5% of values can lie beyond 2 standard deviations on the right side.
So far we have measured continuous data — salaries, heights, temperatures. But what if we are counting events?
Discrete vs. Continuous — Counting vs. Measuring
Discrete vs. Continuous
AnalogyDefinition
"How many calls did I get today?" — you count: 0, 1, 2, 3, ... This is a discrete question, and the Poisson distribution is often a good model. "What is today’s temperature?" — you measure: 23.4°C, any real number in a range. This is continuous, and the normal distribution might fit. A fair die shows 1–6 with equal probability — that is discrete uniform.
Analogy:
"How many calls did I get today?" — you count: 0, 1, 2, 3, ... This is a discrete question, and the Poisson distribution is often a good model. "What is today’s temperature?" — you measure: 23.4°C, any real number in a range. This is continuous, and the normal distribution might fit. A fair die shows 1–6 with equal probability — that is discrete uniform.
Definition:
Discrete distributions assign probabilities to countable outcomes (0, 1, 2, ...). Continuous distributions assign densities to ranges of real numbers. The Poisson distribution models the count of independent events at a fixed rate λ per time period. Exactly one parameter: λ equals both the mean and the variance. Example: emails per hour with λ = 5. The uniform distribution: every outcome is equally likely — fair die (each face 1/6) or random number between 0 and 1.
The analogy breaks at two points: first, real phone calls are not perfectly independent — a call can trigger follow-up calls, violating Poisson assumptions. Second, temperature is measured with finite precision (one decimal place), technically making it discrete in practice — but the continuous model is the better abstraction.
Poisson Counts events (emails/hour, calls/day). One parameter: λ. Shape: right-skewed, peaking at λ.
Normal Measures values (height, weight). Two parameters: μ and σ. Shape: symmetric bell.
Uniform Every outcome equally likely (die, random number). Shape: flat.
Deep Dive: Distributions in AI
Distributions are everywhere in AI. Neural network weights are often initialised from a normal distribution N(0, σ²) — too large and training explodes; too small and gradients vanish. The softmax function converts a vector of raw values into a discrete probability distribution over tokens or classes. And in anomaly detection: a data point more than 3σ from the mean gets flagged as an outlier — but only if the data is actually normally distributed.
Common Misconception
"All distributions are bell-shaped." Wrong. The bell curve is just one shape. Poisson distributions are right-skewed, uniform distributions are flat, and many real-world distributions have heavy tails or multiple peaks. Always visualise your data (histogram) before assuming a shape.
Interactive: How Statistical Operations Scale
When working with distributions, the amount of data matters enormously. Move the slider to see how fast different statistical operations grow with increasing sample size: looking up a stored parameter costs O(1), computing the mean costs O(n), but pairwise comparisons explode at O(n²).
1010000
O(1)1
O(log n)6.6
O(n)100
O(n²)10.000
Moderate Input
At n=100, the difference becomes visible: O(n²) requires 10.000 operations, while O(n) needs only 100. O(log n) needs just 6.6 — that's 15x less than O(n).
Ratio to O(n)
Complexity
Operations
Factor vs. O(n)
O(1)
1
100x faster
O(log n)
6.6
15x faster
O(n)
100
1x (Reference)
O(n²)
10.000
100x slower
Takeaways
Three Key Insights
The normal distribution is symmetric and fully defined by μ and σ. The 68-95-99.7 rule gives you an instant sense of where most values lie — but only when the data is actually bell-shaped.
Skewed distributions pull the mean away from the median. When these two values diverge significantly, the distribution has a long tail — and the median tells the truer story.
Discrete distributions count events (0, 1, 2, ...), continuous distributions measure values (any real number in a range). Choosing the right distribution type is the first modelling decision in any data analysis.
Quiz: Distributions
Question 1 / 4
Not completed
What does the 68-95-99.7 rule describe?
1. What does the 68-95-99.7 rule describe?
☐ A) How many data points exist in a dataset
☐ B) What fraction of values lies within 1, 2, or 3 standard deviations — for normally distributed data
☐ C) How fast an algorithm runs
☐ D) The difference between mean and median
2. A dataset has mean = 80 and median = 45. What does this tell you about the distribution?
☐ A) It is symmetric
☐ B) It is left-skewed
☐ C) It is right-skewed — a few high values pull the mean upward
☐ D) Nothing can be concluded
3. A helpdesk receives an average of 12 tickets per hour. Which distribution models the number of tickets per hour?
☐ A) Normal distribution
☐ B) Uniform distribution
☐ C) Poisson distribution
☐ D) None — tickets do not follow a distribution
4. Why is the mean misleading for right-skewed salary data?
☐ A) The mean is always calculated incorrectly
☐ B) A few extreme salaries pull the mean up, so it does not reflect the typical salary
☐ C) The median is always larger than the mean
☐ D) Salary data cannot be analysed
Answer Key: 1) B · 2) C · 3) C · 4) B
Checkpoint
Learning Goals
A machine produces parts with mean weight 500g and σ = 5g (normal distribution). Quality control rejects parts outside 490–510g. What percentage of parts pass?
A dataset has mean = 60 and median = 35. Is this distribution likely symmetric, right-skewed, or left-skewed? Which summary statistic better represents the "typical" value?
A call centre receives an average of 8 calls per minute. What distribution would you use to model the number of calls per minute? What about the duration of each call?