Flag an outlier relative to a defined comparison group, then investigate its source and meaning. Retain genuine exceptional events; correct verified errors through a documented rule. Explain sensitivity where uncertainty remains, because statistical unusualness alone does not establish that a record is wrong.
Choose a meaningful reference group
A large order is unusual only in relation to something. Compare it with orders using the same currency, commercial scope and time basis. An annual enterprise contract will look extreme beside individual monthly purchases even when both are correct. Inspect the distribution and relevant business segments before selecting a threshold. A rule based on the entire file can hide unusual values inside one segment while flagging every legitimate record in another.
Separate screening from diagnosis. A box-plot fence, a percentile or a business limit can identify candidates for review, but each has assumptions and a purpose. Record the threshold and population version so the same candidate can be reproduced later. Do not describe every flagged value as an error in a report title. Readers may otherwise interpret an investigation queue as proof of faulty operations.
Trace the observation through the data pipeline
Check the original transaction, unit, decimal scale, sign and transformation history. A tenfold difference might result from a real contract, a pack-size conversion or a misplaced decimal. Look for duplicates and joined rows as well as field-level mistakes. Compare the reported quantity with supporting evidence available under the authorised process, and record which evidence was actually reviewed rather than implying a complete audit.
Choose among retaining the value, correcting a demonstrated error, excluding an ineligible record or leaving an unresolved qualification. Keep the original observation and the reason for any changed value. A cap or winsorisation rule changes the analytical estimand even when it is useful for a particular model. It should not silently overwrite the transaction ledger or make an operational chart appear more comfortable.
Show how exceptional events affect the decision
Report a relevant total together with a measure that explains concentration. If one contract dominates a month, show its contribution and a clearly labelled comparison excluding it where that comparison helps planning. The full total remains necessary for questions about actual recorded activity. The supplementary view answers a different question about the rest of the business, so avoid presenting it as the corrected truth.
NIST's outlier guidance distinguishes unusual observations from established bad data; it is background reading for the screening principle. The example below is independently invented. A reviewer should be able to see the chosen comparison group, inspected evidence, action and metric effect in a short investigation note. Recheck the rule when products or commercial models change, because yesterday's unusual annual contract can become tomorrow's normal business.
Also record the investigator and review date so unresolved candidates do not silently become permanent exclusions.
Fictional orders: an annual contract changes the picture
A fictional scenario to explain the method, not a customer result or a market benchmark.
A fictional supplier records nine orders of 100 units and one order of 1,000 units. The total is 1,900 and the mean order value is 190. The median is 100. The largest order accounts for approximately 52.6% of recorded value, so it deserves attention even though it is only one of ten orders.
The supporting order confirms that 1,000 is an annual service contract with the correct currency and no duplicated line. The analyst retains it. A supplementary ordinary-order view reports nine orders totalling 900, explicitly excluding the annual contract for workload planning. It does not replace the complete recorded-sales total.
If the source instead confirmed a decimal error, a correction would need a separate change record and revised total. In this fictional case the evidence supports retention, not trimming. The difference between mean and median helps explain concentration but does not itself authenticate the large contract or establish whether the month represents a lasting trend.
TAKE THIS INTO YOUR NEXT CONVERSATION
A practical next-step checklist.
- Define the comparison population and screening threshold.
- Inspect source evidence and transformation history.
- Preserve original values and resolution reasons.
- Show material concentration and sensitivity clearly.
Common questions.
Does an extreme percentile justify deleting a value?
No. It identifies a review candidate under that population. Establish ineligibility or error before changing the recorded event.
Can a capped value be useful?
Yes for a defined analytical purpose, provided the cap is disclosed and the uncapped source remains available.
Further reading & source context.
These primary references provide background. They are not evidence for the fictional example above and do not endorse Accuralytix.
Found an error or a missing qualification? Send an editorial correction.
Check my data readiness.
Use a practical tool to turn this idea into a clearer starting point.
Check my data readiness