Data Quality

Outlier Detector

Runs three outlier tests on every numeric column and reports which of them flagged each value.

Loading the tool…

Processing happens locally in your browser. What you paste or load is processed by this page and is not uploaded to a server. Nothing is stored unless you use a control that says it stores something, and you can clear anything this site has kept from the privacy page.

How to use this tool

  1. Paste a CSV with at least one mostly numeric column.
  2. Leave the thresholds alone the first time; 1.5 and 3 are the conventional ones.
  3. Select Find outliers.
  4. Start with the values all three methods flagged, then raise the agreement setting if there is too much to look at.

What outlier detector does

Every outlier method disagrees with the others somewhere, and which one a tool picked is usually invisible in the answer. The IQR rule makes no assumption about the shape of the data. The z-score assumes something roughly normal and is dragged around by the very extreme values it is meant to find. The modified z-score uses the median absolute deviation instead of the standard deviation, so it is not.

All three run here and every flagged value is reported with the methods that flagged it. That turns the output into something you can reason about: a value all three dislike is worth investigating, and a value only the z-score dislikes usually just means the column is skewed. A column is only tested when most of it is numeric, so an identifier column with a few numbers in it does not produce nonsense bounds.

Frequently asked questions

The agreement between them. IQR assumes nothing about the distribution, the z-score assumes something roughly normal, and the modified z-score is resistant to the extremes it is looking for. A value all three flag is worth investigating; one only the z-score flags usually just means the column is skewed.

No. It means the value is far from the others, which is sometimes an error, sometimes a genuine extreme and occasionally the most important row in the file. The tool reports candidates and shows the bounds it used; the judgement is yours.

A column is only tested when it has at least four numbers and most of its values parse as numbers. An identifier column with a few numeric entries would otherwise produce bounds computed from a handful of values, which is worse than no answer.

It sets how far beyond the middle half of the data a value has to sit before it is flagged. The conventional 1.5 marks the whiskers of a box plot; 3 is the usual setting for "extreme" outliers only. Raising it reports fewer values, not better ones.