Clean text, outliers in a column, and changes over time: three layers of anomaly detection in ATX AnalyticsTable, where the rule decides, never the AI.
Your month-end total is right to the cent. One row inside it is not.
That is the anomaly nobody reports, because nothing on screen looks wrong.
In ATX AnalyticsTable, anomaly detection works in three layers. Each answers a different question.
"EMEA" and "EMEA " are two regions to a join. The anomaly scan finds non-breaking and zero-width spaces, control codes, stray tabs and markup left behind by a paste or an Office export. It proposes the fixes; the analyst chooses which to apply. As a pipeline step, the same cleaning runs on every new file, on a schedule.
Eleven checks across numeric, category and date columns: z-score, modified z-score (MAD), IQR fences, Isolation Forest, DBSCAN, rare categories, unexpected values, category drift, impossible dates, time gaps and day-of-week seasonality. Every hit says which test fired and why.
In our training run on a 20-row supplier file, 384,000 was flagged by three tests: an extra zero. −12,600 was flagged by one: a credit note, real, and it stays. Same scan, two different decisions. That is why the result is a report written onto a sheet for sign-off, not an automatic delete.
Ask "anything odd in sales?" and Ask Greg checks each period against the series' own trend and season: a spike, a dip or a lasting level shift. An insight watch turns the question into an alert with low, normal or high sensitivity. The rule decides whether it fires, never an AI model.
All of it runs on your machine or server, under the same row- and column-level security as every query. The same data gives the same result on every machine, so an auditor can reproduce a flagged figure.
One honest limit: an outlier is a question, not a verdict. No test knows your business. It tells you where to look.
Training video, handbook and download: links in the first comment.
Which one cost your team more time last year: hidden characters, wrong numbers, or a trend nobody noticed?
Video: Anomaly detection: three layers, statistics decide.
This site needs JavaScript for the full experience. Every page above is reachable without it.