outlier
/ OWT-ly-er /
An outlier is a data point that sits far from the crowd — the one value that makes you do a double-take. In a class where everyone scored between 60 and 90, a single score of 5 is an outlier. It might be a genuine extreme case, or it might be a mistake; telling those two apart is much of the work.
Outliers come from two very different sources. Some are errors: a fat-fingered typo (a weight of 1500 kg for a person), a broken sensor, a unit mix-up. Others are real but rare: a legitimately enormous transaction, a genuinely unusual patient. The crucial point is that you cannot tell which by the number alone — you have to investigate. Statistically, outliers are often spotted as values many standard deviations from the mean, or far outside the bulk of the data, but a flag is an invitation to look, not an automatic verdict to delete.
Why it matters: a few extreme points can drag an average, warp a fitted line, and dominate scale-based preprocessing, so a model ends up obsessing over freaks instead of the typical case. But blindly deleting outliers is dangerous — in fraud detection or rare-disease screening, the outliers are precisely the thing you're trying to find. The honest move is to understand each outlier's cause, then choose: fix it, keep it, cap it, or model it specially. There is no one-size-fits-all rule.
Average salary on a team is about $60,000, until the founder's $5,000,000 figure is included — now the "average" jumps to $550,000, a number that describes nobody. The median, $60,000, barely moves. The outlier didn't need deleting; it needed a robust summary that doesn't let one value hijack the story.
One extreme salary hijacks the mean but barely touches the median.
Never delete outliers reflexively. In fraud, fault detection, and rare-disease work the outliers ARE the signal — discarding them throws away the whole point of the project.