The Normal Distribution and the 68-95-99.7 Rule

Where this picks up

Part 6 looked at the shape of distributions in plain words: symmetric or skewed, one peak or several, light tails or heavy ones. One shape came up again and again as the reference point for comparison: the symmetric, single-peaked, bell-shaped curve. This article gives that shape a name, a precise definition, and a rule that turns it into one of the most useful tools in a data worker's kit. By the end you should be able to look at a mean and a standard deviation (from part 4) and immediately say roughly what range most of the data falls in, without touching the raw numbers again.

The shape is called the normal distribution, also known as the Gaussian distribution. The rule is called the 68-95-99.7 rule, sometimes called the empirical rule. Together they let you go from two summary numbers to a surprisingly detailed picture of an entire dataset.

A concrete starting point: heights in a population

Suppose you have a very large dataset of adult women's heights, and purely for the sake of a clean, round example, suppose the mean is 165 centimetres and the standard deviation is 7 centimetres. These numbers are illustrative, not a claim about any real population; the point is to fix concrete values we can compute with.

If you plotted a histogram of these heights, you would see something recognisable: most people cluster close to 165 cm, fewer people are found as you move further away in either direction, and the picture looks roughly the same whether you look to the left of the centre or to the right. There is no long tail stretching out on one side only, the kind of skew discussed in part 6. The histogram is close to symmetric and has one peak. That shape, when it holds closely enough, is what we call normal.

Heights are a classic example because a huge number of small, mostly independent factors (genetics, nutrition, health across childhood, and more) add up to produce the final value. When many small independent effects add together, the result tends to look bell-shaped. This is not a coincidence; it is a mathematical fact called the central limit theorem, which later parts of this path will use directly when discussing sampling. For now, the practical point is simpler: many real measurements, from heights to measurement errors to exam scores on well-designed tests, turn out to be approximately normal, which makes the shape worth knowing in detail.

What makes a shape "normal"

Before the formula, build the picture. A normal distribution has three visual features:

• It is symmetric: the left half is a mirror image of the right half around the centre.
• It has a single peak (it is unimodal), and that peak sits exactly at the centre.
• Its tails fall off smoothly and never quite touch zero, but they get very thin very quickly as you move away from the centre.

A direct consequence of symmetry and a single central peak is something covered in part 3: for a perfectly normal distribution, the mean, median and mode are all the same value, sitting at the centre of the curve. If you ever compute these three numbers for a dataset and find them close together, that is a hint, though not proof, that the data might be roughly normal. If they are far apart, as with strongly skewed income data, the distribution is clearly not normal, and the rule in this article should not be applied to it.

It helps to be clear about what the normal distribution is not. It is not the only bell-shaped curve, and not every bell-shaped curve is normal; some have heavier or lighter tails, or are slightly asymmetric while still looking roughly bell-like at a glance. The word "normal" here is a technical label for one specific, precisely defined shape, not a judgement that other shapes are abnormal or wrong. A right-skewed distribution of house prices is not a flawed or broken dataset; it is simply not normal, and treating it as if it were will lead to wrong conclusions, a point this article returns to later.

The normal distribution, precisely

A normal distribution is fully described by exactly two numbers: its mean (usually written with the Greek letter mu, μ) and its standard deviation (usually written sigma, σ). That is the entire description. Once you know those two numbers, you know everything about the shape: where the centre is, and how spread out the values are around it.

The mean controls where the peak sits on the number line; change the mean and the whole curve slides left or right without changing its shape. The standard deviation controls how wide or narrow the curve is: a small standard deviation produces a tall, narrow curve where values cluster tightly around the mean, and a large standard deviation produces a short, wide curve where values are spread over a larger range. This matches exactly how standard deviation was introduced in part 4, as a typical distance of data points from the mean; the normal distribution is simply the shape that makes that description most literal and most useful.

For completeness, here is the formula that generates the curve, though you will rarely need to compute it by hand in practice, since software handles this:

f(x) = (1 / (sigma * sqrt(2 * pi))) * exp(-((x - mu) ** 2) / (2 * sigma ** 2))

Read this piece by piece rather than memorising it. The term (x - mu) squared measures how far a value x is from the mean, and squaring it (as variance does, from part 4) makes the distance always positive and penalises large distances more. Dividing by 2 times sigma squared scales that distance by how spread out the data typically is, so the same raw distance counts as "far" in a tightly clustered dataset but "close" in a widely spread one. The negative sign and the exponential function turn large scaled distances into very small numbers, which is what produces the thin tails. The term out front, involving sigma and pi, simply rescales the whole curve so that the total area underneath it equals exactly 1, since this curve describes a probability distribution and total probability must sum to 1. You do not need to evaluate this formula yourself; every statistics and data science library has it built in. What matters is understanding that the entire family of normal curves, for any mean and any standard deviation, is generated by this one shape, just slid and stretched.

The 68-95-99.7 rule

Here is the central fact this article is building towards. For any normal distribution, no matter what its mean and standard deviation are, the same three proportions of the data always fall within the same number of standard deviations of the mean:

• About 68% of the data falls within 1 standard deviation of the mean (that is, between mean minus 1 SD and mean plus 1 SD).
• About 95% of the data falls within 2 standard deviations of the mean.
• About 99.7% of the data falls within 3 standard deviations of the mean.

This is remarkable precisely because it does not depend on the units, the scale, or the specific values involved. It is true for heights measured in centimetres, exam scores measured in points, or measurement errors in a laboratory instrument, as long as the underlying distribution is genuinely normal. The rule comes directly from the area under the curve described by the formula above; the proportions 68%, 95% and 99.7% are not round numbers chosen for convenience, they are what the area under that specific curve works out to between those specific boundaries. More precise values, if you ever need them, are approximately 68.27%, 95.45% and 99.73%, but 68, 95 and 99.7 are accurate enough for the mental arithmetic this rule is meant to support.

Working it through with the heights example

Return to the heights dataset: mean 165 cm, standard deviation 7 cm. Apply the rule directly by adding and subtracting multiples of the standard deviation from the mean.

• Within 1 SD (165 - 7 = 158 cm to 165 + 7 = 172 cm): about 68% of women fall in this range.
• Within 2 SD (165 - 14 = 151 cm to 165 + 14 = 179 cm): about 95% of women fall in this range.
• Within 3 SD (165 - 21 = 144 cm to 165 + 21 = 186 cm): about 99.7% of women fall in this range.

Notice what just happened. Starting from only two numbers, a mean and a standard deviation, you produced a detailed statement about the entire shape of the data: most people (about two thirds) are between 158 and 172 cm, nearly everyone (95%) is between 151 and 179 cm, and it would be unusual, though not impossible, to find someone outside 144 to 186 cm, since that leaves only 0.3% of the population, roughly 3 people in every 1,000, split between the two extreme tails.

That last point is worth making explicit. The 0.3% outside 3 standard deviations is not all on one side; by symmetry, about 0.15% lies below 144 cm and about 0.15% lies above 186 cm. The same halving applies to the other bands if you only want one side: about 16% of women are shorter than 158 cm (half of the 32% outside 1 SD on either side), and about 16% are taller than 172 cm. This halving trick, using the symmetry of the curve, is often how the rule gets used in practice, since questions are frequently one-sided ("what fraction is above this value" rather than "what fraction is within this band").

Z-scores: putting the rule to work on any value

The 68-95-99.7 rule only speaks in whole standard deviations: 1, 2 or 3. Real questions rarely land exactly on those boundaries. To handle any value, convert it into a z-score, which simply states how many standard deviations a value is from the mean:

z = (x - mu) / sigma

Here x is the value you care about, mu is the mean, and sigma is the standard deviation. The z-score is a plain number with no units, which is part of its usefulness: it strips away the original scale (centimetres, points, seconds, dollars) and replaces it with a universal measure of "how unusual is this value, expressed in standard deviations."

Take a height of 179 cm from the earlier example. Its z-score is (179 - 165) / 7 = 2. That matches the earlier calculation exactly: 179 cm is 2 standard deviations above the mean, which is why it was the upper boundary of the 95% band. A height of 158 cm gives a z-score of (158 - 165) / 7 = -1, confirming it as the lower boundary of the 68% band. The negative sign simply means the value is below the mean rather than above it.

Z-scores also work in reverse. If someone tells you a value has a z-score of 1.5, you can recover the original value with x = mu + z * sigma. For the heights example, that gives 165 + 1.5 * 7 = 175.5 cm. This two-way conversion, from raw value to z-score and back, is the mechanism behind standardisation, a step used constantly in preparing data for machine learning models, which is covered more in the next parts of this path and in future articles on preprocessing.

A worked example: reading an exam score

Suppose a large, well-designed standardized test is built so that scores come out approximately normal, with a mean of 500 and a standard deviation of 100 (these numbers are chosen for round arithmetic, not taken from any specific real test, since the scale of any particular test can change over time). A student scores 650. What can you say about how this student did compared to everyone else who took the test?

First compute the z-score: z = (650 - 500) / 100 = 1.5. The student scored 1.5 standard deviations above the mean. That value, 1.5, does not fall on one of the three clean boundaries of the 68-95-99.7 rule, so the rule alone cannot give an exact percentage, but it can still bound the answer usefully. A z-score of 1.5 sits between 1 SD (z = 1) and 2 SD (z = 2). You know that 68% of scores lie within 1 SD (between 400 and 600, so up to z = 1 on the upper side covers 34% above the mean, since 68% is split evenly into 34% on each side by symmetry), and 95% lie within 2 SD (covering 47.5% above the mean on the upper side). So the percentage of students scoring between the mean and a z-score of 1.5 is somewhere between 34% and 47.5%. For an exact figure you would consult a standard normal table or use software (any statistics package can return the precise value, which happens to be close to 43.3% between the mean and z = 1.5, for a total of roughly 93.3% of students scoring below 650 and about 6.7% scoring above it). The point of this example is not to memorise that exact figure, but to see how the rough, three-step rule still gives you an immediately useful bound ("this student is somewhere in the upper few percent, clearly above the 1-SD mark but not quite at the 2-SD mark") without any lookup table at all.

This is the practical value of the rule: it rarely gives you the exact answer, but it gives you an instant, reliable sense of scale, which is often all that is needed to decide whether a value is ordinary, somewhat unusual, or genuinely rare.

Where this shows up in day-to-day data work

• Spotting unusual values quickly. If a sensor reading, a transaction amount, or a user's session length sits more than 2 or 3 standard deviations from the mean, that is an immediate, cheap flag worth a closer look, long before building any complex anomaly-detection model. Part 8 of this path covers outliers in depth and will return to this idea.
• Feature scaling for machine learning. Standardising a feature by subtracting its mean and dividing by its standard deviation, exactly the z-score formula above, is one of the most common preprocessing steps before training models such as logistic regression, k-nearest neighbours or neural networks, because it puts features measured on very different scales (ages in years, incomes in dollars) onto a comparable footing.
• Quality control. Manufacturing processes often use control charts where a measurement is flagged for inspection once it drifts beyond 2 or 3 standard deviations from the process mean, directly borrowing the 95% and 99.7% bands.
• Setting rough expectations. Before fitting any model, simply knowing a variable's mean and standard deviation, and assuming rough normality, lets you sanity-check whether a new observation looks ordinary or surprising, which is often the first and fastest check in exploratory data analysis.
• Communicating uncertainty. Statements like "roughly two thirds of customers spend between $40 and $60" are often just the 68% band of a normal approximation, phrased in plain language for a non-technical audience.

Seeing the rule hold in generated data

It is worth confirming the rule with actual numbers rather than taking it on faith. The code below generates a large sample from a normal distribution and checks what fraction of the sample actually falls within 1, 2 and 3 standard deviations of the sample mean. It assumes numpy (any reasonably recent version, such as 1.2x or later, will work).

import numpy as np

np.random.seed(42)

# Generate 100,000 simulated heights: mean 165 cm, standard deviation 7 cm
heights = np.random.normal(loc=165, scale=7, size=100_000)

mean = heights.mean()
std = heights.std()

for k in (1, 2, 3):
    lower = mean - k * std
    upper = mean + k * std
    fraction_inside = np.mean((heights >= lower) & (heights <= upper))
    print(f"Within {k} SD ({lower:.1f} to {upper:.1f} cm): {fraction_inside * 100:.1f}%")

Running this prints three lines, one for each value of k. The percentages you see should land close to 68%, 95% and 99.7%, though not exactly, because np.random.normal draws random values each time and even 100,000 draws carry some sampling noise (later parts of this path, on populations and samples, explain why this noise shrinks as the sample size grows but never fully disappears). If you rerun the code with a different seed, or a smaller sample size such as 1,000, you will see the percentages wobble a bit more around those three target numbers. That wobble is expected and is itself a useful lesson: the 68-95-99.7 rule describes the exact, true normal distribution; any real or simulated sample only approximates it, more closely as the sample grows larger.

Common mistakes

Assuming every dataset is normal

The single most common mistake is applying the 68-95-99.7 rule to data that is not normal. Part 6 covered skewed distributions, such as income or house prices, where the mean, median and mode pull apart and a long tail stretches out on one side. Applying the normal rule to such data gives confidently wrong answers: it will understate how much mass is in the long tail and misplace where the "typical" range actually sits. Always check the shape first, using a histogram or the simple comparison of mean, median and mode from part 3, before reaching for this rule.

Confusing probability density with probability

The curve produced by the formula earlier in this article is a probability density, not a probability. A single point on the x-axis, such as "exactly 165.000000 cm," technically has a probability of zero under a continuous distribution; what the rule and the curve describe are probabilities over ranges (between 158 and 172 cm, for instance), which correspond to areas under the curve. This distinction matters less for everyday use of the rule, but it explains why the rule is always phrased as "within this range" rather than "equal to this value."

Treating small samples as if they were the true distribution

A histogram of 20 or 30 data points can look lumpy or oddly shaped purely due to randomness, even if the underlying process generating the data is genuinely normal. Concluding "this is not normal" from a very small sample, or conversely trusting the 68-95-99.7 proportions precisely from a small sample, is a common error. The rule is a statement about the true, underlying population distribution; small samples are noisy windows onto that distribution and need to be judged with that noise in mind, a theme the sampling article later in this path develops further.

Forgetting that the rule is an approximation even for normal data

The figures 68%, 95% and 99.7% are themselves rounded from the more precise 68.27%, 95.45% and 99.73%. For most practical purposes this rounding is harmless, but if you need an exact probability, for example for a quality threshold where the exact 2% versus 5% distinction matters financially, use software (such as a cumulative normal function in a statistics library) rather than the rounded rule, and reserve the rule for quick, approximate reasoning.

Assuming z-scores fix non-normal data

Standardising a variable (subtracting the mean, dividing by the standard deviation) always produces a new variable with mean 0 and standard deviation 1, but it does not change the shape of the distribution. A right-skewed variable is still right-skewed after standardisation; it has simply been rescaled. Standardisation is about putting variables on comparable scales, not about making them normal. Transforming a skewed variable toward a more symmetric shape (for example with a logarithm, when all values are positive) is a different, separate step, and one that should be considered on its own merits rather than assumed to follow automatically from standardising.

How to tell whether your data is roughly normal

Before applying the rule, a few quick checks, none of them requiring advanced tools, go a long way:

• Plot a histogram. Look for a single central peak and roughly matching tails on both sides, the visual signature discussed in part 6.
• Compare mean, median and mode (part 3). If they are close together, that is consistent with rough symmetry; if they differ a lot, the data is skewed and the rule should be used cautiously, if at all.
• Check for impossible or bounded values. Many real variables cannot go below zero (prices, durations, counts) or have a hard ceiling (percentages capped at 100). A true normal distribution has tails stretching infinitely in both directions, so any variable with a hard floor or ceiling close to its mean, relative to its spread, cannot be exactly normal, even if it looks bell-shaped in the middle.
• Look at the extremes. If the rule predicts that values beyond 3 standard deviations should be very rare (0.3%), but your dataset has many more extreme values than that, the distribution likely has heavier tails than a true normal distribution, and the rule will understate how often extreme values occur.

None of these checks require fitting a formal model or running a statistical test; they are the same descriptive tools built up across parts 3 through 6 of this path, applied with one new question in mind: does this look enough like a normal curve for the 68-95-99.7 rule to be a reasonable approximation, or not. When the answer is genuinely unclear, treat the rule's output as a rough guide rather than a precise claim, and say so when reporting it to others.

Summary and what comes next

The normal distribution is a specific, precisely defined bell-shaped curve, fully described by just two numbers: its mean, which fixes the centre, and its standard deviation, which fixes the spread. When a dataset is approximately normal, the 68-95-99.7 rule lets you read an enormous amount from those two numbers alone: about 68% of values sit within 1 standard deviation of the mean, about 95% within 2, and about 99.7% within 3. Z-scores extend this to any value, not just the three clean boundaries, by expressing how many standard deviations a value sits from the mean, a calculation that also underlies standardisation, one of the most common preprocessing steps in machine learning. The rule is powerful precisely because it is simple, but that simplicity only holds when the underlying assumption, that the data is roughly normal, actually holds; skewed data, bounded data, and small noisy samples all require caution before the rule is applied.

Part 8 turns to a question this article has already brushed against several times: outliers. Having a working sense of the normal distribution and the 68-95-99.7 rule gives you one concrete, numeric way to ask "is this value unusual enough to deserve a closer look," which is exactly where the next part picks up.