What is Outlier? — one analogy

1The analogy

A sheet of paper taped next to a piggy bank keeps track of how much gets dropped in each day. Scrolling through a month of it, most days land somewhere between one and four dollars, except for one single day marked fifty dollars. That one day is an outlier.

The question is what to make of it. If an extra zero got added, or a note meant for a different sheet landed here by mistake, it needs fixing. But if that day really was a windfall dropped in all at once, that fifty dollars is the single most important line describing this piggy bank. Deleting it just because it stands out leaves behind the wrong conclusion that only a dollar a day ever got saved.

So the real work starts after a stray value turns up. What happened that day gets thought through, the days around it get compared, and any sign of a slip of the hand gets checked for.

2In detail

What counts as "standing out"

Deciding something sits alone needs a yardstick. The most common method lines the values up by size, measures the width the middle half of them takes up, and stretches a fence about one and a half times that width above and below. Anything landing outside the fence becomes an outlier candidate.

When the values cluster around a center, distance from the average gets used instead. Either way, a person decides where "far enough" starts, and shifting that line even slightly changes the list of candidates — which is why it helps to write the rule down before looking at the values, and stick to that order.

Some outliers only show up when several columns get looked at together. A height of six feet is common enough on its own, and so is a weight of ninety pounds, but the two written down for the same person catches the eye.

How much can one value drag

What makes an outlier dangerous isn't how many there are — it's how much pull it has. Averaging a month of savings with one fifty-dollar day pulls the average up a long way. No matter how carefully the other twenty-nine days were recorded, this one day decides the result.

The same happens during training. A common way to score a miss squares the distance, so a single far-off point can outweigh the combined penalty from a hundred points sitting close together. The model ends up pulled toward getting that one point right at everything else's expense.

Deciding whether to keep it or drop it

The first thing to check is how the record got made. A field typed in by hand is a likely candidate for a mistyped digit; a value logged automatically by a machine calls for checking what the equipment was doing at that moment. Once a cause is found, the choice to fix or delete gets much easier.

When no cause turns up, flagging the value is safer than deleting it. A deleted value can't be brought back, but a flagged one can still be revisited later. Pulling the value in only as far as the fence, rather than removing it entirely, is another option that's used.

Sometimes the outlier is the whole point

Everything so far has treated an outlier as an obstacle, but plenty of fields flip that around. Catching card fraud, spotting equipment about to fail, and finding an intrusion on a network are all jobs whose entire purpose is finding "the thing that stands alone." Here, the ordinary values are the backdrop and the stray one is the main event.

The same value gets treated differently depending on the goal. A spike in sales on a promotion day is a nuisance when computing an average, but it's the only clue there is for figuring out which promotion actually worked.

The handling flips too. Instead of deleting the stray value, it gets collected separately, and the question becomes what conditions tend to produce it. The ordinary values now serve only as background, and where to draw the line becomes the whole job.

3More precisely

An outlier isn't a property of the value itself — it's a label that comes from its relationship to the values around it. The same number is ordinary in one dataset and stands out in another. Writing down "this value is an outlier" on its own isn't useful later; what needs recording alongside it is which dataset and which yardstick led to that call.

The analogy has a limit worth noting too. The piggy bank record was written by hand, so that one day can be recalled and checked. Most real data was recorded by someone else, long ago, with no way to trace the cause — so the judgment usually leans on probability and context instead.

It's worth telling apart from noise, too. Noise is a small wobble mixed into every value; an outlier is the one that strayed far more than the rest. Cutting down noise means changing how something gets measured; handling an outlier means deciding what to do with that one value. The two can show up in the same dataset without being the same problem, and mixing them up leads to fixing the wrong thing.

4Try it yourself

TensorFlow Playground ailearn.space Turn up the noise to send a few points to a strange spot, then watch how far the boundary line gets dragged

5Common misconceptions

It's easy to think an outlier is a wrong value, but actually it often captures something rare that genuinely happened.
It's easy to think an outlier should just be deleted whenever it's found, but actually deleting it makes the remaining data look neater than it really is, so the resulting score looks better than it should.
It's easy to think a large dataset buries any outlier, but actually when misses get scored by squaring the distance, a single far-off point keeps its pull no matter how many other points there are.

7One-line summary

In shortAn outlier is a single value sitting far apart from the rest, and whether it gets kept or dropped should wait until there's an answer for why it happened.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02