Outliers are observations that lie far from the bulk of the data — the extreme values that don't fit the typical pattern. How you treat them shapes understanding and decisions. Ignore them and you may miss the signal (e.g. the one customer who reveals a broken process, or the one event that dominates outcomes). Let them dominate and you may overreact to noise (e.g. one bad quarter that is random). The discipline is to distinguish: is this outlier a data error, a rare draw from the same process, or evidence that the process is different? Each implies a different response.
In normal or near-normal distributions, outliers are rare and the mean is a good summary. In fat-tailed or power-law distributions, "outliers" are structurally part of the process — a few observations account for most of the outcome. There, the mean is misleading and the focus shifts to the tail. Understanding and analysing require knowing which world you are in: thin-tailed (outliers are noise or errors) or fat-tailed (outliers are the story). Use the model to decide whether to exclude, downweight, or feature the extreme values, and to choose the right summary statistic (e.g. median vs mean).
Outliers also drive survivorship bias and narrative fallacy: we notice the extreme success and build stories around it while ignoring the many who had similar inputs and did not succeed. The corrective is to ask what the full distribution looks like and whether the outlier is representative of a repeatable process or a one-off.
Section 2
How to See It
Outliers show up when one or a few points sit far from the rest in a distribution, when someone says "excluding the anomaly..." or "if we take out that one deal...," or when the mean and median diverge sharply. The diagnostic: how do we treat the extremes, and does the distribution have thin or fat tails?
Business
You're seeing Outliers when one customer accounts for 40% of revenue and the rest are long-tail. The mean deal size is misleading; the median and the shape of the tail matter. Strategy (e.g. concentration risk, pricing) should reflect that outliers are structural, not noise.
Investing
You're seeing Outliers when a few positions drive most of the return. The "average" position is not representative. Analysis should separate the tail (the big winners or losers) from the bulk and ask whether the tail is repeatable or luck. Outliers can be the whole story in power-law outcomes.
Operations
You're seeing Outliers when one incident or one complaint is an order of magnitude worse than the rest. Deciding whether it is a fluke (exclude from process design) or a signal (process is fragile to rare events) determines whether you harden the system or ignore it.
Analytics
You're seeing Outliers when the mean and median of a metric diverge — e.g. mean revenue per user is $50 but median is $5. The mean is pulled by a few high-value users. Choosing which statistic to report and to optimise (mean vs median) is an outlier-handling decision.
Section 3
How to Use It
Decision filter
"When you see extreme values, ask: (1) Are they errors? (2) Are they rare draws from the same process? (3) Are they evidence of a different process or a fat-tailed distribution? Use the answer to decide whether to exclude, downweight, or feature them, and whether to use mean or median (or tail metrics) for the decision."
As a founder
Report and plan using the right summary for your distribution. If a few customers or deals dominate, use median and concentration metrics; don't let the mean hide the tail. When one outlier drives a narrative ("our best customer says..."), ask what the full distribution looks like. In product and ops, treat outliers as signals when they reveal a broken process or a segment you're missing; treat them as noise when they're one-off and not actionable.
As an investor
Evaluate returns and risks with the full distribution in mind. If outcomes are power-law, the mean is misleading; focus on the tail and the probability of extreme outcomes. When a fund or company touts "average" performance, check whether that average is driven by a few outliers and whether those outliers are repeatable.
As a decision-maker
When someone presents a metric, ask how outliers are handled. Is the mean or median used? Are extremes excluded, and if so, on what basis? Require clarity on whether the distribution is thin-tailed (outliers are noise) or fat-tailed (outliers are structural). That determines how much to weight the extremes in the decision.
Common misapplication: Automatically excluding outliers without asking why they exist. Exclusion can hide the most important signal — e.g. the one complaint that reveals a systemic flaw. First classify: error, rare draw, or structural. Then decide.
Second misapplication: Assuming the mean is always the right summary. In skewed and fat-tailed distributions, the mean is pulled by outliers and can mislead. Use the median when the distribution is skewed; use tail metrics when the tail is the story.
Buffett's focus on margin of safety and avoiding permanent loss is outlier-aware: he prepares for the tail (the bad outcome that could wipe you out), not just the average. His insistence on understanding the range of outcomes and the worst case is a form of taking outliers seriously in analysis and risk.
Munger's inversion — "invert, always invert" — and his use of edge cases align with outlier thinking: the extreme cases (the one failure mode, the one customer who leaves, the one regulatory change) often matter more than the average. He stresses looking at the full distribution and the tails when understanding a business.
Section 6
Visual Explanation
Outliers — Points far from the bulk. In thin-tailed distributions they are noise; in fat-tailed they are structural. Choose mean vs median and exclude vs feature accordingly.
Section 7
Connected Models
Outliers sit with distribution, summary statistics, and tail risk. The models below reinforce how to handle them or create tension with average-based thinking.
Reinforces
Mean Median Mode
Mean, median, and mode are summary statistics. The mean is pulled by outliers; the median is robust. When outliers are present or the distribution is skewed, choosing median over mean is the direct application of outlier-aware thinking.
Reinforces
Standard Deviation & Normal Distribution
Standard deviation measures spread; in a normal distribution, points beyond ~2σ are rare. Outliers are often defined relative to the normal (e.g. beyond 3σ). Knowing whether the distribution is normal or fat-tailed tells you whether "outlier" means noise or structure.
Tension
Black Swan Theory
Black swans are extreme, rare, high-impact events. In Taleb's framing, they are outliers that dominate history but are often ignored by models that assume thin tails. The tension: treating outliers as noise (exclude, downweight) can hide black swans; treating all extremes as structural can overreact to noise.
Tension
Power Law Distribution
In power-law distributions, a few observations account for most of the total — "outliers" are the rule, not the exception. The mean is misleading. The tension with "outlier as anomaly" is that in power-law worlds, the tail is the story; you don't exclude it, you analyse it.
Section 8
One Key Quote
"The median is not the message."
— Stephen Jay Gould, Full House
Gould used the phrase when discussing his cancer prognosis: the mean survival was misleading because the distribution was skewed by a few long-term survivors. The median was more representative of the typical outcome. The lesson: when outliers or skew are present, the mean can be the wrong message. Choose the summary statistic that matches the question and the distribution.
Section 9
Analyst's Take
Faster Than Normal — Editorial View
Always ask how outliers are handled. When someone reports a metric, ask: mean or median? Any exclusions? In skewed and fat-tailed data, the answer changes the story. Require transparency so you can judge whether the summary is appropriate.
Thin vs fat tail is the key distinction. In thin-tailed domains (e.g. heights, many manufacturing tolerances), outliers are often errors or rare noise; robust stats or exclusion can help. In fat-tailed domains (returns, deal size, viral growth), outliers are structural; focus on the tail and use median or tail metrics.
Outliers can be the most important signal. The one complaint that reveals a broken process, the one deal that shows a new segment — don't auto-exclude. Classify first: error, rare draw, or structural. Then decide whether to exclude, downweight, or feature.
Section 10
Test Yourself
Is this mental model at work here?
Scenario 1
A team reports average deal size of $80k. One deal was $2M; the rest were $20k–$50k.
Scenario 2
A founder says 'our best customer loves us' and uses that to justify product direction.
Scenario 3
Revenue is power-law: 5% of customers drive 60% of revenue. The analyst reports mean revenue per customer.
Scenario 4
One data point is a clear typo (revenue entered as 100 instead of 10,000). The analyst excludes it and recomputes.
Section 11
Summary & Further Reading
Summary: Outliers are extreme values in a distribution. How you treat them — exclude, downweight, or feature — depends on whether they are errors, rare draws, or structural (fat-tailed). Use the right summary statistic: median when skewed or fat-tailed, mean when thin-tailed. In understanding & analysing, distinguish thin-tailed (outliers as noise) from fat-tailed (outliers as the story). Pair with mean/median/mode, standard deviation & normal distribution, regression to the mean, and survivorship bias; be aware of black swans and power-law distributions.
Gould on the median vs the mean when distributions are skewed. "The median is not the message" in the context of cancer survival — a classic outlier and summary-statistic lesson.
On extreme outliers that dominate outcomes and the failure of models that assume thin tails. Reframes "outliers" as the main event in fat-tailed domains.
Accessible guide to common statistical errors, including mishandling outliers and choosing the wrong summary statistic.
Leads-to
Regression to the Mean
Extreme first observations (outliers) often regress toward the mean on the next measurement when there is noise or partial skill. So one interpretation of an outlier is "expect regression" — but only when the process is stable and the outlier is a draw from the same distribution, not a structural shift.
Leads-to
Survivorship Bias
We see the extreme successes (outliers) and build stories around them; we don't see the many who tried and failed. Survivorship bias is outlier bias in narrative form. The corrective is to ask for the full distribution, not just the tail we notice.
