One outlier test has an opinion. Eleven checks side by side (IQR, MAD, Isolation Forest, DBSCAN) tell a typo from a rare but real business event.
Your outlier test has an opinion. Run a second one and you find out what it is.
Most teams pick one rule, usually "more than 3 standard deviations" or "outside the IQR fences", and treat whatever it flags as the truth. But every method makes an assumption about what "normal" looks like. When two methods disagree about a value, that disagreement is the most useful thing on the screen.
A real run from our training data, one column of amounts:
One test alone would have given you either too much or too little. With both, you can tell a typo from a rare-but-real business event in seconds.
That is why ATX AnalyticsTable runs eleven outlier checks side by side instead of one:
Each flag comes with the reason it stands out and which checks agreed.
The honest limit: statistics find what is unusual, not what is wrong. A credit note, a year-end booking or a new large client will look exactly like an error. So nothing gets deleted automatically. The flagged rows go into a report for someone who knows the business to sign off.
Swipe through the run, the eleven checks and what happens to a hit.
How do you handle outliers today: one fixed rule, several methods, or a person checking by eye?
Video: Outlier detection: when the tests disagree, that is the information.
This site needs JavaScript for the full experience. Every page above is reachable without it.