The 68-95-99.7 Lens

"Three numbers — 68, 95, 99.7 — are enough to read most bell-shaped claims at first glance."

The Empirical Rule

For a normal distribution with any mean μ and any standard deviation σ:

About 68% of values fall within ±1σ of the mean.
About 95% of values fall within ±2σ.
About 99.7% of values fall within ±3σ.

Three numbers, one rule. The bell's symmetric tails decay exponentially in the squared-distance sense, so these intervals capture progressively more of the distribution. Memorize them and you can read most bell-shaped statistics at first glance without consulting a table.

What the Tails Look Like

Outside ±3σ, less than 0.3% of the distribution remains — split between the two tails (about 0.13% on each side). At ±4σ, the per-tail mass is about 1 in 30,000. At ±5σ, about 1 in 3.5 million. At ±6σ, about 1 in a billion. The tail probabilities shrink very fast — that thin-tail property is exactly what fat-tailed distributions violate, and is exactly the property that Track 07 will exploit to dismantle citizen mistakes.

How to Read a Real Statistic

Hypothetical example: if a runner's performance metric is z = +2 in an approximately normal, clearly defined reference population, it is near the upper 2.3% tail. Time itself reverses direction, so “above” must be defined carefully.

Under an exact normal model, z = +3.5 leaves about 0.023%—roughly 1 in 4,300—in the upper tail. Genuinely uncommon, assuming the score distribution is approximately normal (which most well-designed tests aim for).

'A market move is 6σ from typical.' Translation: about 1 in a billion under normality. Translation under reality: market returns are not normal; this happens far more often than once-per-billion. The arithmetic z-score may be correct, but the one-in-a-billion tail translation is valid only under a calibrated normal model.

The Working Skill

The 68-95-99.7 rule lets you read 'X-sigma' claims as 'about Y% of population' in your head. The skill is reflexive once practiced — and the skill includes pausing to ask whether the bell-curve assumption that underlies the translation actually holds. Under normality, 6σ is 1-in-a-billion. Under reality, often not.

Code

Tail probabilities of the normal at each sigma level·python

import numpy as np
from math import erf, sqrt

# Probability that a standard normal exceeds z, in one tail.
def one_tail(z):
    return 0.5 * (1 - erf(z / sqrt(2)))

# Print the rarity of each sigma level — one tail and both tails.
for z in (1, 2, 3, 4, 5, 6):
    p = one_tail(z)
    one_in = 1 / (2 * p) if p > 0 else float('inf')   # both tails combined
    print(f"z = {z}: per-tail prob = {p:.6e}   total (both tails) ≈ 1 in {one_in:>15,.0f}")

# Notice the speed at which the rarity grows.
# At z=3: ~1 in 370.
# At z=4: ~1 in 16,000.
# At z=5: ~1 in 1.7 million.
# At z=6: ~1 in 500 million.
# Those are the bell curve's tails. Real-world data with fat tails does NOT
# obey these rarities — the same z-value can correspond to far more frequent
# events. Track 07 will show what that costs.

Exercise

Read a recent science or business article that quotes a 'sigma' figure. Translate the σ into the corresponding rarity using the 68-95-99.7 rule. Then ask the second question: is the underlying quantity actually normally distributed? If yes, the rarity is meaningful. If no, the σ statement is a category error and the real rarity is different — usually higher than the bell suggests.

Hint

Domains where the bell tends to be reasonable: physics measurement, biology body measurements, well-designed survey responses. Domains where the bell is usually wrong: financial returns, social-media metrics, file sizes, response times in stressed systems.

The Empirical Rule

What the Tails Look Like

How to Read a Real Statistic

The Working Skill

Code

External links

Exercise

Progress

Comments 0